ITFS Volume 17 Issue 2 April 2009

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 17, NO.
2, APRIL 2009 253
Piecewise Fuzzy Anti-Windup Dynamic Output

Feedback Control of Nonlinear Processes With
Amplitude and Rate Actuator Saturations
Tiejun Zhang, Member, IEEE, Gang Feng, Fellow, IEEE, Huaping Liu, Member, IEEE,
and Jianhong Lu
Abstract—In this paper, a novel anti-windup dynamic output state feedback and output feedback controllers have been de-
compensator is developed to deal with the robust H∞ output veloped for linear systems [10]–[16]. In many cases, to avoid
feedback control problem of nonlinear processes with amplitude the input saturation, low-gain control laws are designed for a
and rate actuator saturations and external disturbances. Via fuzzy
modeling of nonlinear systems, the proposed piecewise fuzzy anti- prescribed control constraint and known initial system states.
windup dynamic output feedback controller is designed based on More specifically, the saturation nonlinearity for linear sys-
piecewise quadratic Lyapunov functions. It is shown that with sec- tems is characterized by the convex hull of several linear feed-
tor conditions, robust output feedback stabilization of an input- backs [11], [12]. Alternatively, the saturation term has been
constrained nonlinear process can be formulated as a convex opti- transformed into a deadzone nonlinearity in [14], [15], and ac-
mization problem subject to linear matrix inequalities. Simulation
study on a strongly nonlinear continuously stirred tank reactor cordingly, a sector condition and S-procedure were used to
(CSTR) benchmark plant is given to show the performance of the derive some stability conditions. These explicit conditions can
proposed anti-windup dynamic compensator. be formulated as linear matrix inequalities (LMIs), and thus,
Index Terms—Actuator saturation, amplitude and rate con- it is very easy to obtain the controller gains by solving these
straints, anti-windup, dynamic output feedback, fuzzy systems, LMIs. In fact, this LMI-based anti-windup design can also deal
linear matrix inequality (LMI), piecewise quadratic Lyapunov with the tougher actuator rate constraint problem conveniently
function, process control. [16].
However, to our best knowledge, few works have been re-
I. INTRODUCTION ported in open literature to consider the robust output feedback
control problem for strongly nonlinear industrial processes in
UTPUT feedback control plays a major role in process
O control for industrial plant operation and optimization
[1]–[4] since state variables are not always fully measurable for
the presence of actuator saturations, external disturbances, and
plant uncertainties. Takagi–Sugeno (T–S) fuzzy dynamic mod-
els have been widely accepted in the control community due
most of the industrial plants in practice. There have been many to its excellent interpretability and capability of representing
results on dynamic output feedback stabilization for continu- the dynamics of nonlinear systems [17]–[22]. There have been
ous linear systems and discrete-time systems (for example, see many controller design methods for T–S fuzzy systems via both
[6]–[8]). More recently, the robust H2 and H∞ control problem state feedback and observer-based output feedback [20]–[26]. It
of discrete-time linear systems with polytopic uncertainties via has been demonstrated that piecewise quadratic Lyapunov func-
dynamic output feedback was investigated in [9]. tions [27], [28] can be used to reduce the conservatism in the
On the other hand, in process industry, actuators are always fuzzy stabilizing controller design [25], [26]. The controller de-
subject to physical limitations [1], [2], [5], which would lead to sign for fuzzy systems with actuator saturation has also attracted
performance deterioration and even instability for large pertur- much attention during the last several years, and representative
bations [10], [11]. To deal with these problems, some bounded works include [29]–[32]. It is noted that only low-gain design
approach was used in [29]. A less conservative approach that is
Manuscript received February 12, 2007; revised June 9, 2007; accepted based on convex hull representations was proposed in [30]. Very
September 3, 2007. First published April 30, 2008; current version published recently, the saturation function was formulated into a specific
April 1, 2009. The work was supported in part by the Hong Kong Research
Grant Council under Grant CityU 112806, in part by the National Natural Sci- nonlinear saturation sector in [31] and a new design approach
ence Foundation of China under Grant 50640460116, Grant 60504003, and was given that requires less number of LMIs. All of [29]–[31]
Grant 60625304, and in part by the National High-Tech R&D Program of China dealt with the case of continuous-time systems. A fuzzy Popov
under Grant 2006AA05A107.
T. J. Zhang and G. Feng are with the Department of Manufacturing Engineer- criterion design approach for both continuous- and discrete-time
ing and Engineering Management, City University of Hong Kong, Kowloon, fuzzy systems was given in [32]. However, all of the aforemen-
Hong Kong (e-mail: tjzhang@ieee.org; megfeng@cityu.edu.hk). tioned results are only restricted to state-feedback controllers.
H. P. Liu is with the Department of Computer Science and Technology,
Tsinghua University, Beijing 100084, China (e-mail: hpliu@tsinghua.edu.cn). To our best knowledge, there is no result available in the open lit-
J. H. Lu is with the School of Energy and Environment, Southeast University, erature on dynamic output feedback controller for fuzzy systems
Nanjing 210096, China (e-mail: jhlu_seu@yahoo.com.cn). with actuator saturation. Furthermore, the rate-limited controller
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. design, due to its difficulty, has never been considered for fuzzy
Digital Object Identifier 10.1109/TFUZZ.2008.924238 systems.
1063-6706/$25.00 © 2009 IEEE
Authorized licensed use limited to: University of Michigan Library. Downloaded on May 13, 2009 at 17:01 from IEEE Xplore. Restrictions apply.
254 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 17, NO. 2, APRIL 2009
To deal with the aforementioned robust constrained dynamic proximation error between the original nonlinear system and its
output feedback control problem, an uncertain fuzzy model is approximating model with the assumption that
used to approximate the uncertain nonlinear process and also to 1 1 2 3
deal with plant modeling errors that are inevitable for approxi- [∆Al , ∆Bl , ∆Dl ] = M l ∆[N l , N l , N l ], ∆ T ∆ ≤ In .
mation. Based on this fuzzy model, we then develop a piecewise (2)
Let µl (ξ(k)) be the normalized membership function of the
robust fuzzy anti-windup dynamic output compensator for the
uncertain nonlinear processes with amplitude and rate satura- inferred fuzzy set M l where M l = rj =1 Mjl and Li=1 µl = 1.
tions on manipulated inputs. Motivated by the recent result [16], By using a center-average defuzzifier, product inference, and
the design of this anti-windup dynamic compensator is based singleton fuzzifier [17]–[22], the uncertain fuzzy model (1) can
on piecewise quadratic Lyapunov functions and can be accom- be expressed by the following global model:

plished by a convex optimization procedure subject to several 

x(k + 1) = A(µ)x(k)
+ B(µ)u(k)
+ D(µ)w(k)
LMI conditions. Both controller parameters and the anti-windup 



compensation gain can be obtained simultaneously. This makes 
 y(k) =  C(µ)x(k)
1/2 
joint optimization possible and may lead to better performance. Qy y(k) (3)

  1/2 
Finally, the proposed controller design method is applied to the 

 z(k) =  Ra u(k) 

control of the strongly nonlinear chemical continuously stirred 
 1/2
tank reactor (CSTR) plant. Rr ∆u(k)
The rest of the paper is organized as follows. Section II where
introduces some preliminaries about uncertain fuzzy models,

L
L
piecewise Lyapunov stability, and sector conditions for input
A(µ) = l ,
µl A
B(µ) = l
µl B
amplitude and rate saturations. The novel piecewise fuzzy anti- l=1 l=1
windup dynamic output compensator is proposed in Section III.
In Section IV, the case study on a nonlinear CSTR bench-
L
L
C(µ) = µl Cl ,
D(µ) = l
µl D
mark is present to demonstrate the effectiveness of the proposed
l=1 l=1
approach. Finally, some conclusions are drawn in Section V.
Notations: For any two vectors x, y ∈ m , x y means that µ := µ(ξ) := (µ1 , . . . , µL )
x(i) − y(i) ≥ 0 ∀i = 1, . . . , m, where x(i) and y(i) denotes the l = Al + ∆Al , l = Bl + ∆Bl , l = Dl + ∆Dl .
A B D
ith component of x and y, respectively. For any two symmetric
matrices A and B, A > B means that A − B is positive definite. It should be pointed out that Cl in (1) can be re-
Further, stands for symmetric blocks in matrices and ∗ stands garded as one constant matrix C with 0 or 1 ele-
for blocks in matrices without subsequent use. ments since y(k) stand for some measurable system states.
Without loss of generality, the disturbance w(k) is as-
II. PRELIMINARIES
sumed to be with finite energy ∞ k =0 w(k) w(k) ≤ wm ax <
T
A. Uncertain Fuzzy Dynamic Model ∞. Note that z(k)T z(k) = y(k)T Qy y(k) + u(k)T Ra u(k) +
∆u(k)T Rr ∆u(k) implies that the H∞ control of (1) also en-
T–S fuzzy dynamic models are widely used to represent com-
forces the transient control performance.
plex systems [17]–[22] with both fuzzy inference rules and local
analytic linear models as follows.
B. Piecewise Quadratic Lyapunov Stability
Rule Rl : IF ξ1 (k) is M1l AND . . . ξr (k) is Mrl , THEN
 Similar to [27] and [28], the output space can be parti-

 x(k + 1) = (Al + ∆Al )x(k) + (Bl + ∆Bl )u(k)

 tioned into a number of closed polyhedral regions, denoted as

 + (Dl + ∆Dl )w(k)

 {Ss }s∈I ⊆ p , where I is the index set of regions. For subse-

 y(k) = Cl x(k) quent use, a set Ω that represents all possible transitions from
 1/2  (1) one region to any possible region including itself is defined as

 Qy y(k)

  

 z(k) =  Ra1/2 u(k) , l = 1, . . . , L Ω := {s, t|y(k) ∈ Ss , y(k + 1) ∈ St }. (4)



 1/2
Rr ∆u(k) Define a piecewise Lyapunov function candidate for the region
s by
where Rl denotes the lth fuzzy inference rule, L the num-
ber of fuzzy rules, Mjl are fuzzy sets, x(k) ∈ n the system V (x(k)) = x(k)T Ps x(k), s∈I (5)
state variables, u(k) ∈ m the control input variables, ∆u(k)
then, we have
the one-step control difference or also called the control rate,
y(k) ∈ p the outputs available for measurement, z(k) ∈ p ∆V (x(k)) = V (x(k + 1)) − V (x(k))
the controlled variable that needs to be regulated, w(k) ∈ q
= x(k + 1)T Pt x(k + 1) − x(k)T Ps x(k). (6)
the external disturbance that belongs to L2 [0, ∞), ξ(k) ∈ r
some measurable system variables, and (Al , Bl , Cl , Dl ) is the If the Lyapunov function (5) satisfies that ∆V (k) < 0 for all
lth nominal local model of the fuzzy system (1). The uncertain trajectories of the system, the system can be guaranteed to be
terms (∆Al , ∆Bl , ∆Dl ) represent the plant uncertainty and ap- stable in the sense of Lyapunov.
ZHANG et al.: PIECEWISE FUZZY ANTI-WINDUP DYNAMIC OUTPUT FEEDBACK CONTROL OF NONLINEAR PROCESSES 255
Lemma (Piecewise Lyapunov Stability [28]): The au- where sata (•) and satr (•) denote the amplitude and rate satu-
tonomous piecewise linear system x(k + 1) = As x(k) is rations, respectively, and ω the input of this integrator system.
globally exponentially stable if the following LMIs admit a If |σ(k)(i) | ≤ ρr (i) , i = 1, . . . , m, it follows that
set of positive definite matrix solutions Ps :
|∆u(i) (k)| = |u(i) (k + 1) − u(i) (k)| ≤ ρr (i) . (14)
ATs Pt As − Ps < 0, s, t ∈ Ω. (7)
Proof: Provided that the Lipschitz constant of the saturation
Notice that the piecewise quadratic Lyapunov functions [28] are function sat(•) equals 1, it follows that
used here to reduce the conservatism in the stabilization problem |∆u(i) (k)| = |sata(i) (v(k) + σ(k))(i) − sata(i) (v(k))(i) |
of a piecewise linear system, which is quite different from the
case for a linear time-invariant system. ≤ |σ(k)(i) | = |satr (ω(i) (k))| ≤ ρr (i) . (15)

C. Input Saturation and Anti-Windup Condition
Assume now that the manipulated input u is subject to am- III. MAIN RESULTS
plitude limitations defined as follows: Considering the system with some input saturation, the ob-
−ρa
u(k)
ρa (8) jective here is to design a suitable anti-windup output feedback
controller to guarantee the stability of the resulting closed-loop
where ρa(i) > 0, i = 1, . . . , m, denote the component-wise con- control system with an optimal level of disturbance attenuation,
trol amplitude bounds. Thus, the actual control input for im- that is, to find a controller such that the upper bound γ > 0 of
plementation is a saturated one, u(k) = sat(v(k)), with each induced L2 -norm from the disturbance w(k) to the controlled
component defined by output z(k)

 +ρa(i) , if ρa(i)
v(i)
 z(k)
≤γ (16)
sat(v)(i) = v(i) , if − ρa(i)
v(i)
ρa(i) (9) w(k)

 −ρ , if v
−ρ
a(i) (i) a(i). is minimized for all nonzero w(k) ∈ L2 under zero initial
conditions.
Let the function ψa (v) := v − sat(v), then with (9), one has
 To achieve this goal, we choose the following piecewise fuzzy
 v(i) − ρa(i) , if ρa(i)
v(i)
 anti-windup dynamic output feedback compensator as a candi-
ψa (v(i) ) = 0, if − ρa(i)
v(i)
ρa(i) (10) date controller.

 v + ρ , if v
−ρ Rule Rl : IF ξ1 (k) is M1l AND . . . ξr (k) is Mrl , THEN
(i) a(i) (i) a(i)

 v(k + 1) = Im v(k) + σ(k)
which behaves like a deadzone nonlinearity. 


 x (k + 1) = Acls xc (k) + Bsc [y(k)T v(k)T ]T
Moreover, as in [5], the control input in industry might also 
 c


be subject to rate constraint 

c
+ Els (sata (v(k)) − v(k))
+ Fs (satr (yc (k)) − yc (k))
c (17)
−ρr
∆u(k)
ρr (11) 

 yc (k) = C c xc (k) + Dc [y(k)T v(k)T ]T


 s s
where ρr (i) > 0, i = 1, . . . , m, denote the component-wise rate 


 u(k) = sat a (v(k)), σ(k) = satr (yc (k)),
constraint bounds of control inputs. Then, the control variation 
∆u(k) has the similar saturated components to (9), and thus, s ∈ I, l = 1, . . . , L
we have the corresponding deadzone nonlinearity ψr . where xc (k) ∈ n , yc (k) ∈ m , and v(k) ∈ m . Let
For one polyhedral set Ψ := {(α, β)| − ρ(i) ≤ (α(i) −
β(i) ) ≤ ρ(i) , i = 1, . . . , m}, the so-called sector condition has ψr (k) := ψr (yc (k)) = yc (k) − satr (yc (k)) (18)
been described as follows. ψa (k) := ψa (v(k)) = v(k) − sata (v(k)) (19)
Lemma 2 ([14], [15]): If (α, β) ∈ Ψ, then the nonlinearity
ψ(α) satisfies the inequality it follows that
u(k) = v(k) − ψa (k) (20)
ψ(α)T · W · [ψ(α) − β] ≤ 0 (12)
σ(k) = yc (k) − ψr (k) (21)
where W denotes any diagonal positive definite matrix.
Note that the sector condition (12) is indeed useful to develop v(k + 1) = Im v(k) + σ(k) (22)
the convex stability conditions of the constrained linear time-
where yc (k) = Csc xc (k)
+ v(k) ] , s ∈ I, and
Dsc C[x(k)T T T
invariant system with anti-windup strategy [14], [15].
C = diag{C, Im }. Also, by the fuzzy inference mentioned in
Lemma 3 ([16]): Consider the following system composed of
Section II-A, (17) becomes
m integrators

L

 v(k + 1) = Im v(k) + σ(k) xc (k + 1) = µl [Bsc C[x(k)T v(k)T ]T + Acls xc (k)
σ(k) = satr (ω(k)) (13) l=1


u(k) = sata (v(k)) − Els
c
ψa (k) − Fsc ψr (k)], s ∈ I. (23)
And under the control input (20), the fuzzy system (3) reads (k))T · Wr · [ψ(Ks x
ψ(Ks x (k)) − Gr x
(k)] ≤ 0 (27)

L
for any matrix Ga , Gr ∈ m ×2(n +m ) satisfying (L x, Ga x ) ∈
x(k + 1) = µl [(Al + ∆Al )x(k) + (Bl + ∆Bl )u(k)
Ψ and (Ks x , Gr x) ∈ Ψ, respectively.
l=1
Therefore, the control objective becomes equivalent to find-
+ (Dl + ∆Dl )w(k)] ing an anti-windup dynamic output compensator (17) to ensure

L the stability of the closed-loop system (25) and minimize its dis-
= l x(k) + B
µl [A l v(k) − B
l ψa (k) turbance attenuation performance parameter γ. Now we present
l=1 the corresponding controller design method in the following
l w(k)], theorem.
+D s ∈ I. (24)
Theorem 1: For the system (1), the piecewise fuzzy dy-
Defining an augmented state vector namic output feedback controller with anti-windup loop (17)
can guarantee the stability of the closed-loop system and achieve
x(k) x(k) a disturbance rejection level γ, if there exist several sets of
x(k) = ∈ n + m , (k) =
x ∈ 2(n + m )
v(k) xc (k) symmetric positive definite matrices Ps , Vs ∈ (n + m )×(n +m ) ,
positive definite diagonal matrices Sa , Sr ∈ m ×m , and matri-
and with (22)–(24), the closed-loop system can be formulated
ces X, Y, W, Us , Als ∈ (n + m )×(n +m ) , Za1 , Za2 , Zr1 , Zr2 , Cs ∈
as follows:
m ×(n +m ) , Qals , Qrs ∈ (n + m )×m , Bs ∈ (n + m )×(p+ m ) , Ds ∈

L
m ×(p+ m ) , s ∈ I, such that the following optimization prob-
(k + 1) =
x ls x
µl [A (k) − Fs ψr (k) − B
ls ψa (k)
lem is solvable
l=1
l w(k)], s ∈ I
+D (25) min η (28)
X
c c

ls Al + FDs C FCs ∆Al O subject to the LMIs (29), as shown at the bottom of the page and
A = Als + ∆Al = +
Bsc C Acls O O  
Ps
Fs =
F
, Al =
Al Bl
, F=
O  Us Vs  ≥ 0,
Fsc O Im Im (L0 X)(i) − Za(i)
1
(L0 )(i) − Za(i)
2
ρ2a(i)

ls Bl ∆Bl Bl s ∈ I,i = 1, . . . , m (30)
B = B ls + ∆B l = + , B l =
c
Els O O  
Ps
l Dl ∆Dl Dl  Us Vs  ≥ 0,
D = Dl + ∆Dl = + , Dl = .
O O O Cs(i) − Zr1(i) (Ds C)(i) − Zr (i) ρr (i)
2 2
s ∈ I, i = 1, . . . , m (31)
Moreover, denoting L = [L0 O], L0 = [O Im ], and
Ks = [ Dsc C Csc ] yields v(k) = L · x(k) and yc (k) = Ks · where X denotes a set of decision variables
(k); then with (18) and (19), the sector conditions (12) for
x
amplitude and rate saturations become η, {Ps }, {Us }, {Vs }, Za1 , Za2 , Zr1 , Zr2 , Sa , Sr ,
x(k))T · Wa · [ψ(L
ψ(L x(k)) − Ga x
(k)] ≤ 0 (26) X, Y, W, {Als }{Bs }, {Cs }, {Ds }, {Qals }, {Qrs }
 
ηIq
 O Ps 
 
 O Us Vs 
 
 O Zr1 Zr2 2Sr 
 
 O Za1 Za2 O 2Sa 
 
 −D Al X + FCs Al + FDs C B l Sa X + X T − Pt 
 l FSr 
 T 
−Y Dl Al s Y Al + Bs C
T r
Qs Ql sa
I + W − Ut Y + Y − Vt
T T

  > 0,
 O 1/2 1/2 
 Q X Q O O O O In + m 
 O 
 O O O O O O O In + m 
 
 O 1/2
Ra L0 X
1/2
Ra L0 O
1/2
Ra Sa O O O O Im 
 
 O 1/2
Rr Cs
1/2
R r Ds C
1/2 
 Rr Sr O O O O O O Im 
 1 T 1 T 
 O O O O O (M l ) (M l ) Y O O −
O O ε 1 In 
3 1 1 2
−N l Nl X Nl O N l Sa O O O O O O O εIn
s ∈ I, t ∈ Φ(s), l ∈ Υ(s), (29)
and where Φ(s) = {t ∈ I: s, t ∈ Ω}, Υ(s) = {l ∈ {1, . . . , L}: With the closed-loop system (25), (33) is equivalent to (34), as
∃y ∈ Ss , µl (y) > 0}, and shown at the bottom of the page, where the first inequality holds
due to Lemma 3 and 2X̄ T R̄Ȳ ≤ inf R̄ > 0 {X̄ T R̄X̄ + Ȳ T R̄Ȳ }
Al Bl O Bl as in [20]. Then, one has
Al = , F= , Bl =
O Im Im O x(k + 1)) − V (
V ( x(k)) + y(k)T Qy y(k) + u(k)T Ra u(k)
T
Dl C Qy C O + ∆u(k)T Rr ∆u(k) − γ 2 w(k)T w(k) < 0 (35)
Dl = , Q=
O O O
if the terms in the bracket of (34) are negative definite; that is,
 2 
where Qy , Ra , and Rr are given controller design parameters, γ Iq
and γ = η 1/2 . Moreover, the controller gains of the piece-  O Ps 
 
wise fuzzy anti-windup dynamic compensator (17) can be ob-  O −W G 2W 
 r r r 
tained by  O −Wa Ga O 2Wa 
 
 −D ls Pt−1 
 l ls Fs
−A >0
Dsc = Ds , M N T = W − XT Y B 
 1/2 
 O −Q O O O I 
Csc = (Cs − Dcs CX)(M T )−1  
 O −Ra L
1/2
O Ra
1/2
O O I 
c
Els = N −1 (Qals Sa−1 − Y T Bl ) O
1/2
−Rr Ks Rr
1/2
O O O O I
(36)
Fsc = N −1 (Qrs Sr−1 − Y T F) or equivalently
 2 
Bsc = N −1 (Bs − Y T FDsc ) γ Iq
Acls = N −1 (Als − Y T Al X − Bs CX − Y T FCsc M T )M −T .
 O
 Ps  
 O
 −Wr Gr 2Wr  
(32)  O
 −Wa Ga O 2Wa  
Proof: For the closed-loop system (25), consider the
 −D
 l −Als Fs B ls Pt−1  
 1/2 
piecewise quadratic Lyapunov function given by V ( x(k)) =  O −Q O O O I 
 
(k)T Ps x
x (k), s ∈ I. If the sector condition (26) exists, and  O −Ra L
1/2
O Ra
1/2
O O I 
with Q = diag{C T Qy C, O}, it follows that O −Rr Ks
1/2
Rr
1/2
O O O O I
x(k + 1)) − V (
V ( x(k)) + y(k)T Qy y(k)
O 
 O O 
+ u(k)T Ra u(k) + ∆u(k)T Rr ∆u(k) − γ 2 w(k)T w(k)  O
 O O 
 O O O O  > 0.
(k + 1)T Pt x
≤x (k)T Ps x
(k + 1) − x (k) + x x(k)
(k)T Q +
 −∆Dl
 −∆Als O ∆B l O 
+ u(k)T Ra u(k) + ∆u(k)T Rr ∆u(k) − γ 2 w(k)T w(k)  O
 O O O O O 
O O O O O O O
− 2ψa (k)T Wa [ψa (k) − Ga x
(k)] O O O O O O O O
− 2ψr (k)T Wr [ψr (k) − Gr x
(k)], s, t ∈ Ω. (33) (37)

L
L
ls x
µl [A ls ψ(k) + D
(k) − B l w(k)]T · Pt · j s x
µj [A j s ψ(k) + D
(k) − B j w(k)] − x
(k)T Ps x
(k) + x x(k)
(k)T Q
l=1 j =1
+ u(k) Ra u(k) + ∆u(k) Rr ∆u(k) − γ w(k)T w(k) − 2ψa (k)T Wa [ψa (k) − Ga x
T T 2
(k)] − 2ψr (k)T Wr [ψr (k) − Gr x
(k)]

L
≤ ls x
µl [A ls ψ(k) + D
(k) − B l w(k)]T Pt [A
ls x ls ψ(k) + D
(k) − B l w(k)] − x
(k)T Ps x
(k) + x x(k)
(k)T Q
l=1
+ u(k)T Ra u(k) + σ(k)T Rr σ(k) − γ 2 w(k)T w(k) − 2ψa (k)T Wa [ψa (k) − Ga x
(k)] − 2ψr (k)T Wr [ψr (k) − Gr x
(k)]
      2 
w(k) T DT γ Iq
l
 x   T  
(k)  
Ps − Q 
L
µl    A
 lsT  Pt [ D l A ls ] −  O
ls −Fs −B 
=  ψr (k)  
 −Fs   O −Wr Gr 2Wr 
l=1
ψa (k) −BT O −Wa Ga O 2Wa
ls
     
O O w(k)
 LT   T   (k) 
+  · Ra · [ O L O −I ] +  Ks  · Ra · [ O Ks −I O ]  x  (34)
 O   −I    ψr (k) 
−I O ψa (k)

Also with the assumption (2) for the uncertain parts of fuzzy Ps UsT Pt UtT
model, let ΠT Ps Π = , ΠT Pt Π =
Us Vs Ut Vt
l1 ∆[N
l1 , N
l2 , N
l3 ] c
[∆Al , ∆B l , ∆Dl ] = M Al X + FCs Al + FDs C
1 ΠT H T Als Π =
Als Y T Al + Bs C
l1 = M l , N
M l1 = N 1 O , N l2 = N 2 , N
l3 = N 3l .
l l
O Als = Y T Al X + Bs CX + Y T FCsc M T + N Acls M T
Then the second term in (37) is equivalent to X̄ T ∆Ȳ + Ȳ T Bs = Y T FDsc + N Bsc , Cs = Dsc CX + Csc M T ,
∆X̄ ≤ inf ∆ ≤I {εX̄ T X̄ + ε−1 Ȳ T Ȳ }, where X̄ = [O O O
1 )T O O O] and Ȳ = [N 3 N 1 O −N 2 O O Ds = Dsc
O (M l l l l
O O], which is similar to that in [7]. So, (36) holds if the T T Bl Sa
Π H B ls Sa = , Qals = (Y T Bl + N Els c
)Sa
matrix inequality (38), as shown at the bottom of the page, Qals
holds. Motivated by the variable change method in [6], now we
define several matrices to deal with piecewise dynamic output T T FSr
Π H Fs S r = , Qals = (Y T F + N Fsc )Sr
feedback Qrs

Y ∗ −1 X ∗ X I −Dl
H= , H = , Π= . − Π H Dl =
T T
, Gr Π = [ Zr1 Zr2 ]
NT ∗ MT ∗ MT O −Y T Dl
1/2 1/2 T
Thus, multiplying (38) on both sides by 1/2 Π = Q X Q
Q , Q=
C Qy C O
diag{I, −Π, Sr , Sa , HΠ, I, I, I, I, I}, O O O O

Sr = Wr−1 , Sa = Wa−1 Ra1/2 LΠ = Ra1/2 L0 X Ra1/2 L0 , Ga Π = [ Za1 Za2 ]

and its transpose yields (39), as shown at the bottom of the page. Rr1/2 Ks Π = Rr1/2 Cs Rr1/2 Ds C , η = γ 2
From the definition of Π and H, it follows that
(Ml1 )T HΠ = (M 1 )T (M 1 )T Y , N l2 Sa = N 2 Sa
T l l l
T X W
Π HΠ = , W = XT Y + M N T 1 Π = N X N , −N
N
1 1 3 = −N . 3
I Y l l l l l
 
γ 2 Iq
 
 O Ps 
 
 O −Wr Gr 2Wr 
 
 
 O −Wa Ga O 2Wa 
 
 −Dl −Als Fs B ls
Pt−1 
 
 1/2  > 0. (38)
 O −Q O O O I 
 
 1/2 1/2 
 O −Ra L O Ra O O I 
 
 O 1/2
−Rr Ks
1/2
Rr O O O O I 
 
 
 O O O O (Ml1 )T O O O ε−1 I 
3
−N −N1 O
N2 O O O O O εI
l l l
 
γ 2 Iq
 
 O ΠT Ps Π 
 
 O Gr Π 2Sr 
 
 
 O Ga Π O 2Sa 
 
 −ΠT H T Dl T
Π H Als ΠT
Π H T F s Sr
T
Π H T B ls Sa
T
Π H Pt−1 HΠ
T T

 
 1/2 Π >0 (39)
 O Q O O O I 
 
 1/2 1/2 
 O Ra LΠ O Ra S a O O I 
 
 O
1/2
Rr Ks Π
1/2
Rr S r O O O O I 
 
 1 )T HΠ 
 O O O O (M l O O O ε−1 I 
−N3 1Π
N O N 2 Sa O O O O O εI
l l l
With H T Pt−1 H ≥ H + H T − Pt as in [8], substituting these the piecewise fuzzy anti-windup compensator (17) can be ob-
variable changes into (39) yields the condition (29) in the tained as in (32), and γ = η 1/2 . The proof is thus complete.
theorem, and from (35), it can be concluded that (29) can Remark 1: As a special case, choosing Gr = Ks , Ga = L,
guarantee Ks x − Gr x = 0 < ρr , and L x − Ga x = 0 < ρa implies that
(Ks x, Gr x
) and (L x, Ga x ) always belong to Ψ; thus, the sector
V ( x(k)) < −y(k)T Qy y(k) − u(k)T Ra u(k)
x(k + 1)) − V ( conditions (26) and (27) are verified for any x ∈ n +n . So the
−∆u(k)T Rr ∆u(k) + γ 2 w(k)T w(k). (40) condition (30) and (31) in Theorem 1 can be dropped, and the
related terms Zr1 , Zr2 , Za1 , and Za2 in the condition (29) become
It thus follows that when w(k) ≡ 0, the closed-loop system Cs , Ds C, L0 X, and L0 , respectively.
is asymptotically stable. By summing this inequality from Remark 2: The piecewise quadratic Lyapunov function (PLF)
k = 0 to ∞, with V ( x(0)) = 0 from the assumption of zero has been considered for dynamic output feedback compensator
(0) = 0 and V (
initial condition x x(∞)) ≥ 0 in the presence of design in Theorem 1. Alternatively, if a common Lyapunov func-
disturbance, it implies that tion (CLF) is used in the design of fuzzy anti-windup dynamic
∞
compensator like that for linear systems in [16], Ps and Pt in
[y(k)T Qy y(k) + u(k)T Ra u(k) the conditions (29)–(31) would be replaced by one positive def-
k =0 inite matrix X, and Vs and Vt by one positive definite block Y ,
∞
and all the matrices Us , Ut , and W by In + m ; moreover, the de-
+ ∆u(k)T Rr ∆u(k)] < γ 2 [w(k)T w(k)]. (41) cision variables Ps , Us , and Vs would also be eliminated in the
k =0 theorem. In fact, the CLF is just one special case of the piece-
Meanwhile, if the condition (30) in the theorem is satisfied, wise Lyapunov function [27], [28]; thus, PLF provides much
with the change of variables Ps , L, and Ga , it becomes more freedom of controller design than CLF. And the resulting
PLF-based anti-windup controller in Theorem 1 is expected to
ΠT Ps Π (LΠ − Ga Π)T(i) be less conservative.
≥ 0. (42)
(LΠ − Ga Π)(i) ρ2a(i)
IV. CASE STUDY: ROBUST DYNAMIC OUTPUT FEEDBACK
Pre- and postmultiplying (42), respectively, by diag{Π−T , I}
CONTROL OF CSTR WITH INPUT SATURATIONS
and its transpose, one gets
In this section, we consider the case study of a highly nonlin-
Ps (L − Ga )T(i) ear CSTR [2] to evaluate the validity of the proposed piecewise
≥0 (43)
(L − Ga )(i) ρ2a(i) fuzzy H∞ dynamic output feedback control approach.
or equivalently
A. Dynamics of Chemical CSTR Plant
(k)|2 ≤ ρ2a(i) .
|(L − Ga )(i) x
With constant liquid volume, the CSTR for an exothermic
Obviously, with the condition (30), (L ) ∈ Ψ ensures
x, Ga x irreversible reaction A → B is described by the following dy-
the satisfaction of the sector condition (26). Similarly, if the namic model based on a component balance for reactant A and
condition (31) in the theorem is satisfied, with the change of on an energy balance
variables Ps , Ks , and Gr , it becomes
F E
ΠT Ps Π (Ks Π − Gr Π)T(i) ĊA = (CA f − CA ) − k0 exp − CA
V RT
≥ 0. (44)
(Ks Π − Gr Π)(i) ρ2r (i) F (−∆H) E
Ṫ = (Tf − T ) + k0 exp − CA
Pre- and postmultiplying (44), respectively, by diag{Π−T , I} V ρCp RT
and its transpose, one gets UA
+ (Tc − T )
Ps (Ks − Gr )T(i) V ρCp
≥0 (45)
(Ks − Gr )(i) ρ2r (i) y = T, u = Tc (46)
or equivalently where CA is the concentration of A in the reactor, T is the
reactor temperature, and Tc is the temperature of the coolant
|(Ks − Gr )(i) x
(k)| ≤ 2
ρ2r (i) .
stream. The objective is to control T by manipulating Tc .
Obviously, with the condition (31), (Ks x , Gr x ) ∈ Ψ ensures Table I contains nominal operating conditions, which corre-
the satisfaction of the sector condition (27). So, when both the spond to an unstable steady state. The resulting steady-state
conditions (29), and (30) and (31) are satisfied, the closed-loop Tc –T curve of CSTR is given in Fig. 1. The open-loop response
system (25) can be guaranteed to be stable with H∞ control per- in Fig. 2 demonstrates that the reactor exhibits highly nonlinear
formance. Moreover, less γ in (41) implies better disturbance behavior in this operating regime. By defining the state vec-
attenuation performance. Hence, with the decision variables tor as x = [CA , T ]T , the manipulated input as u = Tc , and the
X, Y, W, Sr , Sa , Als , Bls , Cs , Ds , Qals , Qrs , and η by solving external disturbance as the temperature deviation of feed reac-
(28)–(31), the controller gains Acls , Bsc , Els c
, Fsc , Csc , and Dsc in tant A, w = Tf − Tf s , (46) can be expressed in the form of
TABLE I
NOMINAL OPERATING CONDITIONS FOR THE CSTR
Fig. 3. Membership functions of fuzzy model and region partition for CSTR.
As shown in Fig. 1, the CSTR (46) has the following three

equilibrium states at steady input u = 300 with w = 0
x1s = [0.875, 324.4]T ,x2s = [0.5, 350]T ,x3s = [0.2, 370.65]T.
B. H∞ Output Feedback Control of Constrained CSTR via

Piecewise Fuzzy Anti-Windup Dynamic Compensator
Using the fuzzy modeling method described in [22], and
choosing the membership functions for the premise variable
x2 or T as

 1, if x2 ≤ 324.4

1 − (x2 − 324.4)/(350 − 324.4),
“low”: M 1 =

 if 324.4 < x2 < 350
0, if x2 ≥ 350
Fig. 1. Steady-state input–output curve of CSTR. !
2 1 − M 1 , if x2 ≤ 350
“middle”: M =
1 − M 3 , if x2 ≥ 350

 0, if x2 ≤ 350

3 (x2 − 350)/(370.65 − 350),
“high”: M =

 if 350 < x2 < 370.65
1, if x2 ≥ 370.65
which are shown in Fig. 3, we can easily obtain the following
local models of fuzzy system (1) for the nonlinear CSTR (46)
under the sampling time Ts = 0.05 min

0.9447 0.0001 0.00001
A1 = , B1 =
1.3442 0.9043 0.09951

0.9055 0.0001 0.00001
A2 = , B2 =
9.4439 0.9014 0.09934

0.7801 0.0001 0.00001
A3 = , B3 =
35.2608 0.9006 0.09927
Fig. 2. Open-loop response for +5 K (a) and −5 K (b) step changes in T c of C1 = C2 = C3 = C = [ 0 1 ]

CSTR.
0.00000 0.00000 0.00000
D1 = , D2 = , D3 = .
0.04757 0.04749 0.04745
nonlinear state–space model ẋ = f (x) + g(x)u + h(x)w with
The satisfactory approximation ability of fuzzy model is demon-
E
F
V (CA f − x 1 ) − k0 exp − R x2 x1 strated as shown in Fig. 4. For the sake of simplicity, we assume
f (x) = F (−∆ H ) E that
V (Tf s − x2 ) + ρC p k0 exp − R x 2 x1 − V ρC p x2
UA

1 1 1 1
M l = 0.004 · In , N l = 0.01 ·
0 0 1 1
g(x) = UA , h(x) = F .
V ρC p V 2 0 3 0
N l = 0.001 · , N l = 0.001 · , l = 1, 2, 3
1 1
tude bound ρa = 40 and the rate bound ρr = 5. The space of

output y = x2 is partitioned into three regions, S1 = {x2 |x2 <
324.4}, S2 = {x2 |324.4 < x2 < 370.65}, and |S3 = {x2 |x2 >
370.65}, which are shown in Fig. 3. Let Ls denote the set of
firing fuzzy rules in region s for the piecewise fuzzy controller
(17), then one has L1 = {1}, L2 = {1, 2, 3}, and L3 = {3}. By
solving the problem (28)–(31) in Theorem 1, we can obtain the
piecewise fuzzy anti-windup dynamic output compensator with
γ = 0.0669 and the controller gains in (47), as shown at the
bottom of this page. The corresponding characteristic matrices
for piecewise Lyapunov functions in controller design are also
given in (48), as shown at the bottom of the next page.
In order to compare the performance of the proposed piece-
wise fuzzy anti-windup dynamic controller, other two control
approaches are also applied to the CSTR plant under the same
design parameters Qy , Ra , and Rr :
1) CLF-based fuzzy anti-windup dynamic controller de-
Fig. 4. Validation of fuzzy model for CSTR plant (solid: original CSTR; scribed in Remark 2; and
dashed: fuzzy model). 2) fuzzy dynamic output feedback controller without anti-
windup compensator.
and ε = 0.01 for all the uncertain terms of the fuzzy model. Also, When the desired output reference of the closed-loop system
suppose that there exists a finite energy disturbance w(t) = is 350 K, the responses of the closed-loop control system based
Tf − Tf s = 0.1Tf s exp(−1.2(t)) sin(6t) since the fifth minute. on three approaches are shown in Figs. 5–7, respectively. In com-
The parameters for controller design are chosen as Qy = parison with these results, it is easily observed that the proposed
0.1 · I1 , Ra = 0.1 · I1 , Rr = 0.1 · I1 , and the input constraints piecewise fuzzy approach can achieve satisfactory output feed-
are um ax = 340, um in = 260, ∆um ax = 5, that is, the ampli- back control performance with closed-loop stability ensured,
   
0.25308 −1156.3 4.8748 × 107 2.3728 × 108 −5.8733 × 107
Ac11 =  1.8069 × 10−7 0.75259 2381.1  , B11 c
=  −115.67 22.431 ,
−12 −5
2.5375 × 10 8.7628 × 10 0.34201 −2.5285 −0.45264
   
0.0144 −4.2433 × 108
c
E11 =  11.6198  , C1c = [ 0.000 0.000 −0.1149 ] , D1c = [ −0.5592 −0.7235 ] , F1c =  114.91 ;
−0.3081 −0.51984
   
0.09095 −44613 −3.3247 × 105 1.0035 × 109 4.0061 × 107
c 
A12 = 2.2969 × 10 −7
0.76352 −389.84  , B2 =c  −308.33 −0.54291  ,
−11 −5
1.8868 × 10 3.3803 × 10 −0.04833 −1.5795 −0.28583
     
0.0094 0.09095 −44613 −3.3247 × 10 5
0.0094
c
E12 =  12.1908  , Ac22 =  2.2486 × 10−7 0.73194 −373.61  , E22c
=  10.0010  ,
−11 −5
−0.3062 3.4676 × 10 7.8613 × 10 −0.06547 −0.3066
   
0.09095 −44613 −3.3247 × 105 0.0092
A32 =  2.0266 × 10
c −7
0.63089 −303.85  , E32c
=  5.8839  ,
1.8868 × 10−11 0.00021814 −0.14878 −0.3031
 
−4.2433 × 108
C2c = 10−3 [ 0.0000 0.1051 0.7835 ] , D2c = [ −2.3648 −1.0396 ] , F2c =  115.42 ;
−0.51896
   
0.24685 −20281 −2.2063 × 107 2.3343 × 108 −5.9776 × 107
Ac33 =  1.3964 × 10−7 0.62458 12445  , B3c =  −109.69 19.709 ,
7.8232 × 10−11 0.00025 5.445 −2.5254 −0.45062
   
0.0145 −4.2433 × 108
E3c =  8.7016  , C3c = [ 0.0000 0.0000 0.0520 ] , D3c = [ −0.5501 −0.7211 ] , F3c =  110.52 . (47)
−0.3043 −0.5169
Fig. 5. Piecewise fuzzy anti-windup dynamic output feedback control of

CSTR (solid: closed-loop response; dashed: output reference). Fig. 7. Fuzzy dynamic output feedback control of CSTR without anti-windup
(solid: closed-loop response; dashed: output reference).
and the performance shown in Fig. 5 is much better than those

in Figs. 6 and 7 based on the other two approaches.
V. CONCLUSION
This paper has extended the sector condition for constrained
discrete-time linear systems [15] to the anti-windup controller
design of nonlinear processes with amplitude and rate actua-
tor saturation. Based on uncertain fuzzy models, the proposed
piecewise fuzzy anti-windup dynamic output feedback compen-
sator was formulated as a convex optimization problem, which
can be solved easily by efficient semidefinite programming
tools. The resulting closed-loop system is guaranteed to be sta-
ble with robust H∞ control performance. The simulation results
on a nonlinear CSTR plant have demonstrated the advantages
Fig. 6. Fuzzy anti-windup dynamic output feedback control of CSTR (solid: of this novel dynamic controller. Time delay is commonly met
closed-loop response; dashed: output reference).
in practical process control, so one interesting future research
   
−0.0027 −0.0380 −0.0076 1.4694 −0.1142 1.7611
P1 =  −0.0380 1.2861 −0.9181  , U1 =  −0.0136 1.0218 −0.0128  ,
−0.0076 −0.9181 6.2321 −0.0029 0.0044 1.0017
   
−0.0025 −0.0383 −0.0084 0.8519 0.0409 1.7417
P2 =  −0.0383 1.2866 −0.9179  , U2 =  0.0004 0.9999 −0.0157  ,
−0.0084 −0.9179 6.2751 −0.0000 0.0000 1.0016
   
−0.0024 −0.0328 −0.0069 3.7733 −1.1131 1.4259
P3 =  −0.0328 1.0072 −0.9488  , U3 =  0.0135 0.4275 −0.0818  ,
−0.0069 −0.9488 6.2076 0.0025 −0.1156 0.9838
   
0.0065 −0.0000 0.0000 0.0065 −0.0000 0.0000
V1 = 108 ·  −0.0000 0.0000 0.0000  , V2 = 108 ·  −0.0000 0.0000 0.0000  ,
0.0000 0.0000 3.2817 0.0000 0.0000 3.5348
 
0.0065 −0.0000 0.0000
V3 = 108 ·  −0.0000 0.0000 0.0000  . (48)
0.0000 0.0000 3.2817
topic is to investigate novel robust piecewise fuzzy anti-windup [24] C. Lin, Q.-G. Wang, and T. H. Lee, “Improvement on observer-based H ∞
controller for constrained nonlinear processes with time delay. control for T–S fuzzy systems,” Automatica, vol. 41, pp. 1651–1656,
2005.
Another appealing research topic is the study of sampled-data [25] G. Feng, “Stability analysis of discrete-time fuzzy dynamic systems based
anti-windup dynamic output feedback control systems when the on piecewise Lyapunov functions,” IEEE Trans. Fuzzy Systems, vol. 12,
effect of the sampling period on digital control performance is no. 1, pp. 22–28, Feb. 2004.
[26] G. Feng, C.-L. Chen, D. Sun, and Y. Zhu, “H ∞ controller synthesis
considered. of fuzzy dynamic systems based on piecewise Lyapunov functions and
bilinear matrix inequalities,” IEEE Trans. Fuzzy Syst., vol. 35, no. 1,
pp. 94–103, Feb. 2005.
REFERENCES [27] M. Johansson, Piecewise Linear Control Systems: A Computational Ap-
proach. Berlin, Germany: Springer-Verlag, 2003.
[1] B. W. Bequette, Process Control: Modeling, Design, and Simulation.
[28] G. Feng, “Stability analysis of piecewise discrete-time linear systems,”
Upper Saddle River, NJ: Prentice-Hall, 2003.
IEEE Trans. Autom. Control, vol. 47, no. 7, pp. 1108–1112, Jul. 2002.
[2] M. A. Henson and D. E. Seborg, Nonlinear Process Control. Upper
[29] B. S. Chen, Y. Y. Chen, and C. L. Lin, “Nonlinear fuzzy H ∞ guidance law
Saddle River, NJ: Prentice-Hall, 1997.
with saturation of actuators against maneuvering targets,” IEEE Trans.
[3] K. M. Hangos, J. Bokor, and G. Szederkenyi, Analysis and Control of
Control Syst. Tech., vol. 10, no. 6, pp. 769–779, Nov. 2002.
Nonlinear Process Systems. London: Springer-Verlag, 2004.
[30] Y. Y. Cao and Z. L. Lin, “Robust stability analysis and fuzzy-scheduling
[4] A. Swarnakar, H. J. Marquez, and T. Chen, “Robust stabilization of nonlin-
control for nonlinear systems subject to actuator saturation,” IEEE Trans.
ear interconnected systems with application to an industrial utility boiler,”
Fuzzy Syst., vol. 11, no. 1, pp. 57–67, Feb. 2003.
Control Eng. Practice, vol. 15, no. 6, pp. 639–654, Jun. 2007.
[31] C. S. Tseng and B. S. Chen, “H ∞ fuzzy control design for nonlinear
[5] P. C. Chen and J. S. Shamma, “Gain-scheduled l 1 -optimal control for
systems subject to actuator saturation,” in Proc. IEEE Int. Conf. Fuzzy
boiler-turbine dynamics with actuator saturation,” J. Process Control,
Syst., 2006, pp. 783–788.
vol. 14, pp. 263–277, 2004.
[32] J. C. Lo and M. L. Lin, “Feedback control via Popov for fuzzy systems with
[6] C. Scherer, P. Gahinet, and M. Chilali, “Multiobjective output-feedback
input saturations,” in Proc. IEEE Int. Conf. Fuzzy Syst., 2004, pp. 1221–
control via LMI optimization,” IEEE Trans. Autom. Control, vol. 42,
1226.
no. 7, pp. 896–911, Jul. 1997.
[33] Y. Y. Cao, Z. Lin, and D. G. Ward, “Anti-windup design of output tracking
[7] L. Yu and F. Gao, “Output feedback guaranteed cost control for uncertain
systems subject to actuator saturation and constant disturbances,” Auto-
discrete-time systems using linear matrix inequalities,” J. Optim. Theory
matica, vol. 40, pp. 1221–1228, 2004.
Appl., vol. 113, no. 3, pp. 621–634, Jun. 2002.
[8] J. Daafouz and J. Bernussou, “Robust dynamic output feedback control
for switched systems,” in Proc. 41st IEEE Conf. Decision Control, Las
Vegas, NV, 2002.
[9] R. Yang, L. Lu, and L. Xie, “Robust H 2 and H ∞ control of discrete-time
systems with polytopic uncertainties via dynamic output feedback,” Int. Tiejun Zhang (S’04–M’08) received his B.Eng. and
J. Control, vol. 78, no. 16, pp. 1285–1294, Nov. 2005. M.Sc. degrees in thermal power engineering from the
[10] V. Kapila and K. M. Grigoriadis, Eds., Actuator Saturation Control. Southeast University, Nanjing, China, in 2001 and
New York: Marcel Dekker, 2002. 2004, respectively, and his Ph.D. degree in systems
[11] T. Hu and Z. Lin, Control Systems With Actuator Saturation: Analysis and and control from the City University of Hong Kong,
Design. Boston, MA: Birkhauser, 2001. Kowloon, Hong Kong, in 2008.
[12] Y.-Y Cao, Z. Lin, and D. G. Ward, “An antiwindup approach to enlarg- He is currently a Postdoctoral Research Associate
ing domain of attraction for linear systems subject to actuator satura- in the Rensselaer Polytechnic Institute, Troy, NY. His
tion,” IEEE Trans. Autom. Control, vol. 47, no. 1, pp. 140–145, Jan. current research interests focus on modeling, con-
2002. trol, and optimization of advanced energy and power
[13] I. E. Kose and F. Jabbari, “Scheduled controllers for linear systems with systems.
bounded actuators,” Automatica, vol. 39, pp. 1377–1387, 2003.
[14] J. M. Gomes da Silva, Jr. and S. Tarbouriech, “Antiwindup design with
guaranteed regions of stability: An LMI-based approach,” IEEE Trans.
Autom. Control, vol. 50, no. 1, pp. 106–111, Jan. 2005.
[15] J. M. Gomes da Silva, Jr. and S. Tarbouriech, “Anti-windup design with
guaranteed regions of stability for discrete-time linear systems,” Syst.
Control Lett., vol. 55, pp. 184–192, 2006. Gang Feng (S’90–M’92–SM’95–F’09) received the
[16] J. M. Gomes da Silva, Jr., D. Limon, and T. Alamo, “Dynamic output B.Eng. and M.Eng. degrees in automatic control (of
feedback for discrete-time systems under amplitude and rate actuator electrical engineering) from Nanjing Aeronautical In-
constraints,” in Proc. 44th IEEE Conf. Decision Control, 2005, pp. 5588– stitute, Nanjing, China, in 1982 and 1984, respec-
5593. tively, and the Ph.D. degree in electrical engineering
[17] T. Takagi and M. Sugeno, “Fuzzy identification of systems and its applica- from the University of Melbourne, Melbourne, Vic.,
tion to modelling and control,” IEEE Trans. Syst., Man, Cybern., vol. 15, Australia, in 1992.
no. 1, pp. 116–132, Jan. 1985. During 1997–1998, he was a Lecturer/Senior Lec-
[18] T. A. Johansen, R. Shorten, and R. Murray-Smith, “On the interpretation turer with the School of Electrical Engineering, Uni-
and identification of dynamic Takagi-Sugeno fuzzy models,” IEEE Trans. versity of New South Wales, Australia. He was a Vis-
Fuzzy Syst., vol. 8, no. 3, pp. 297–313, Jun. 2000. iting Fellow at the National University of Singapore
[19] S. G. Cao, N. W. Rees, and G. Feng, “Analysis and design for a class of (1997) and Aachen Technology University, Germany (1997–1998). Since 2000,
complex control systems, Part I–II,” Automatica, vol. 33, pp. 1017–1039, he has been with the City University of Hong Kong, Kowloon, Hong Kong,
1997. where he is currently a Professor. His current research interests include robust
[20] K. Tanaka and H. O. Wang, Fuzzy Control Systems Design and Analysis: adaptive control, signal processing, piecewise linear systems, and intelligent
A Linear Matrix Inequality Approach. New York: Wiley, 2001. systems and control.
[21] G. Feng, “A survey on analysis and design of model-based fuzzy control Prof. Feng is an Associate Editor of the IEEE TRANSACTIONS ON AUTOMATIC
systems,” IEEE Trans. Fuzzy Syst., vol. 14, no. 5, pp. 676–697, Oct. CONTROL and the IEEE TRANSACTIONS ON FUZZY SYSTEMS, and was an Asso-
2006. ciate Editor of the IEEE TRANSACTIONS ON SYSTEMS, MAN, & CYBERNETICS,
[22] M. C. M. Teixeira and S. H. Zak, “Stabilizing controller design for un- PART C. He is also an Associate Editor of the Journal of Control Theory and
certain nonlinear system using fuzzy models,” IEEE Trans. Fuzzy Syst., Applications. He has also been a member of the Conference Editorial Board of
vol. 7, no. 2, pp. 133–142, Apr. 1999. the IEEE Control System Society. He was the recipient of the Alexander von
[23] X. Liu and Q. Zhang, “New approaches to H ∞ controller designs based Humboldt Fellowship in 1997–1998, and the 2005 IEEE TRANSACTIONS ON
on fuzzy observers for T–S fuzzy systems via LMI,” Automatica, vol. 39, FUZZY SYSTEMS Outstanding Paper Award in 2007.
pp. 1571–1582, 2003.
Huaping Liu (M’07) was born in Sichuan Province, Jianhong Lu received the B.S. degree in power en-
China, in 1976. He received the Ph.D. degree from gineering and the Ph.D. degree in process control
the Department of Computer Science and Technol- engineering from the Southeast University, Nanjing,
ogy, Tsinghua University, Beijing, China, in 2004. China, in 1985 and 1990, respectively.
From 2004 to 2006, he was a Postdoctoral Fellow He is currently a Professor in the Department
in the Department of Automation, Tsinghua Univer- of Energy Information and Automation, School of
sity, where he is currently an Associate Professor in Energy and Environment, Southeast University. His
the Department of Computer Science and Technol- current research interests include advanced control
ogy. His current research interests include intelligent theory and technology for complex thermal systems.
control and robotics.
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 17, NO. 2, APRIL 2009 265
H∞ Fuzzy Control of Nonlinear Systems Under

Unreliable Communication Links
Huijun Gao, Member, IEEE, Yan Zhao, and Tongwen Chen, Fellow, IEEE
Abstract—This paper investigates the problem of H∞ fuzzy con-

trol of nonlinear systems under unreliable communication links.
The nonlinear plant is represented by a Takagi–Sugeno (T–S) fuzzy
model, and the control strategy takes the form of parallel dis-
tributed compensation. The communication links existing between
the plant and controller are assumed to be imperfect (that is, data
packet dropouts occur intermittently, which appear typically in
a network environment), and stochastic variables satisfying the
Bernoulli random binary distribution are utilized to model the un-
reliable communication links. Attention is focused on the design
of H∞ controllers such that the closed-loop system is stochasti-
cally stable and preserves a guaranteed H∞ performance. Two
approaches are developed to solve this problem, based on the
quadratic Lyapunov function and the basis-dependent Lyapunov
function, respectively. Several examples are provided to illustrate
the usefulness and applicability of the developed theoretical results. Fig. 1. NCS prototype.
Index Terms—Basis-dependent Lyapunov function, H∞ control,

nonlinear systems, Takagi–Sugeno (T–S) fuzzy systems, unreliable shown their interest in T–S fuzzy models, and a great number of
communication links. results have been reported in the literature. To mention a few, the
problem of stability analysis is investigated in [4], [25] and [38],
I. INTRODUCTION stabilizing and H∞ control designs are reported in [5], [8], [26],
and [50], state estimation is addressed in [1], and reliable con-
N THE REAL WORLD, most physical systems and pro-
I cesses are not linear, and the inherent nonlinearity makes
their control and regulation a very hard task. This facilitates a
trol strategies are presented in [44] and [46]. These results are
concerned with various engineering problems, as well as many
classes of T–S fuzzy systems, including T–S fuzzy systems with
rapidly growing interest in fuzzy control of nonlinear systems. parameter uncertainties [26], T–S fuzzy systems with state de-
The main reason is that Takagi–Sugeno (T–S) fuzzy models can lays [7], [28], [47], T–S fuzzy systems with actuator satura-
be used to describe nonlinear systems effectively as T–S fuzzy tion [6], T–S fuzzy systems with singular perturbations [2],
systems with affine terms can approximate any smooth nonlin- [29], [30], T–S sampled-data fuzzy systems [23], and so on.
ear function to any specified accuracy within any compact set, Most of the aforementioned results concerning T–S fuzzy
and the control design can be carried out on the basis of the fuzzy systems are based on an implicit assumption that the commu-
model. The main feature of T–S fuzzy models is to express the nications between the physical plant and controller are perfect.
local dynamics of each fuzzy rule by a linear system model, In many practical situations, however, there may be a nonzero
and the overall fuzzy model of the system is achieved by fuzzy probability that some measurements or control inputs will be
“blending” of the linear system models. For this local linearity lost during their transmission. This is more obvious and impor-
of the T–S fuzzy model, the stability and synthesis analysis can tant for the emerging networked control systems (NCSs) where
be simplified based on the well-known framework of Lyapunov feedback control loops are closed via digital communication
function method [13]. In recent years, many researchers have channels. In a prototype of NCS shown in Fig. 1, several com-
ponents communicate over a shared network, and information
Manuscript received April 10, 2007; revised August 7, 2007. First pub- flows among the sensor, actuator, and controller according to
lished April 30, 2008; current version published April 1, 2009. This work was
supported in part by the Natural Science and Engineering Research Coun- some rules. NCSs are becoming increasingly popular because
cil of Canada, in part by the National Natural Science Foundation of China they have several advantages over traditional systems, such as
(60825303, 60834003), in part by the 973 Project (2009CB320600), in part by low cost, reduced weight and power requirements, simple instal-
the Research Fund for the Doctoral Programme of Higher Education of China
(20070213084), in part by the Heilongjiang Outstanding Youth Science Fund lation and maintenance, and high reliability [22]. However, the
(JC200809), in part by Postdoctoral Science Foundation of China (200801282), insertion of the communication network in the feedback loops
and in part by the Fok Ying Tung Education Foundation (111064). makes the analysis and synthesis problems much more com-
H. Gao and Y. Zhao are with the Space Control and Inertial Technology Re-
search Center, Harbin Institute of Technology, Harbin 150001, China (e-mail: plex. In particular, among a few others, data packet dropout is
hjgao@hit.edu.cn). an important issue to be taken into consideration for the analysis
T. Chen is with the Department of Electrical and Computer Engineer- and synthesis of communication-based control systems.
ing, University of Alberta, Edmonton, AB T6G 2R3, Canada (e-mail:
tchen@ece.ualberta.ca). In view of the importance of unreliable communications
Digital Object Identifier 10.1109/TFUZZ.2008.924315 in modern control systems, many researchers have begun to
1063-6706/$25.00 © 2009 IEEE
study how to design control systems with the simultaneous

consideration of communication issues. Concerning the issue
of data packet dropout (or data missing), a number of important
results have been reported. For control problems, readers are
referred to [33], [42], and [49]. In [48], the stability of NCSs
with finite data packet dropout is analyzed; in [33], a packet loss
model for the network is proposed based on which the results for
discrete-time linear systems with Markovian jumping parame-
ters can be applied; in [3], the uncertainty threshold principle is
utilized for stability analysis; and in [21], stochastic variables Fig. 2. T–S fuzzy system with unreliable communication links.
are utilized to model the data missing phenomenon. The issue
of missing measurements has also been investigated for sig- pectation of x and expectation of x conditional on y. Matrices,
nal processing problems (see, for instance, [32] and [41]). It is if their dimensions are not explicitly stated, are assumed to be
worth mentioning that most of these results are concerning lin- compatible for algebraic operations.
ear control systems, and to the best of the authors’ knowledge,
up to now, no effort has been made about how to analyze and II. PROBLEM FORMULATION
synthesize T–S fuzzy systems with unreliable communication Consider a T–S fuzzy system with unreliable communication
links, and control of T–S fuzzy systems in the presence of packet links shown in Fig. 2, where the physical plant is represented by
dropouts still remains open and unsolved. a T–S fuzzy model, and the unreliable communications exist in
In this paper, we make an attempt to investigate the prob- both the downlink (from physical plant to controller) and uplink
lem of H∞ fuzzy control of nonlinear systems under unreliable (from controller to physical plant). Now, we model the physical
communication links. More specifically, the nonlinear plant is plant, controller, and communication links mathematically.
represented by a T–S fuzzy model, and the control strategy
takes the form of parallel distributed compensation. The com- A. Physical Plant
munication links, existing between the plant and controller, are
Many nonlinear systems can be expressed as a set of linear
assumed to be imperfect (that is, data packet dropouts occur
systems in local operating regions. According to fuzzy modeling
intermittently), and stochastic variables satisfying the Bernoulli
in [34], a nonlinear discrete-time system can be represented by
random binary distribution are utilized to model the unreliable
the T–S fuzzy model as follows.
communication links. Attention is focused on the design of
Plant Rule i: IF θ1 (k) is Mi1 and θ2 (k) is Mi2 and · · · and
state feedback H∞ controllers such that the closed-loop sys-
θp (k) is Mip , THEN
tem is stochastically stable and preserves a guaranteed H∞
performance. Two approaches are developed to solve this prob- xk +1 = Ai xk + B2i uk + B1i wk
lem, based on quadratic Lyapunov function and basis-dependent zk = Ci xk + D2i uk + D1i wk (1)
Lyapunov function, respectively. These two approaches provide
alternatives for designing H∞ controllers with different degree i = 1, . . . , r.
of conservativeness and computational complexity, and all the Here, Mij are the fuzzy sets; xk ∈ Rn is the state vector;
conditions are formulated in the framework of linear matrix in- uk ∈ Rm is the input vector; wk ∈ Rp is the exogenous distur-
equalities (LMIs). Several examples substantially illustrate the bance input which belongs to l2 [0, ∞); zk ∈ Rq is the controlled
usefulness and applicability of the developed theoretical results. output; Ai , B2i , B1i , Ci , D2i , and D1i are known constant ma-
The remainder of the paper is organized as follows. Section II trices with appropriate dimensions; r is the number of IF–THEN
formulates the problem under consideration. H∞ analysis and rules; and θk = [ θ1 (k), θ2 (k), . . . , θp (k) ] is the premise vari-
synthesis results are presented in Section III based on a quadratic able vector. Throughout the paper, it is assumed that the premise
approach, and parallel results are given in Section IV through variables do not depend on the input variable uk explicitly. Given
a basis-dependent approach. Some extensions and illustrative a pair of (xk , uk ), the final output of the fuzzy system is inferred
examples are given in Sections V and VI, respectively. Finally, as [34]
concluding remarks are given in Section VII. r
Notations: The notation used throughout the paper is fairly xk +1 = hi (θk ) [Ai xk + B2i uk + B1i wk ]
i=1
standard. The superscript “T ” stands for matrix transposition;
Rn denotes the n-dimensional Euclidean space, and the notation
r
P > 0 (≥ 0) means that P is real symmetric and positive def- zk = hi (θk ) [Ci xk + D2i uk + D1i wk ] (2)
inite (semi-definite). l2 [0, ∞) is the space of square-integrable i=1
where
vector functions over [0, ∞); the notation | · | refers to the
Euclidean vector norm, and · 2 stands for the usual l2 [0, ∞) ωi (θ(k))
hi (θ(k)) = r
norm. In symmetric block matrices or complex matrix expres- i=1 ωi (θ(k))
sions, we use an asterisk (∗) to represent a term that is induced
p
by symmetry, and diag{. . .} stands for a block diagonal matrix. ωi (θ(k)) = Mij (θj (k))
In addition, E{x} and E{ x | y} will, respectively, mean ex- j =1
GAO et al.: H ∞ FUZZY CONTROL OF NONLINEAR SYSTEMS UNDER UNRELIABLE COMMUNICATION LINKS 267
with Mij (θj (k)) representing the grade of membership of θj (k) D. Closed-Loop System
in Mij . Then, it can be seen that We introduce another Bernoulli process {e(k)} with

r
e(k) = α(k)β(k). Thus, we have that e(k) = 1 when both
ωi (θk ) ≥ 0, i = 1, 2, . . . , r, ωi (θk ) > 0 α(k) = 1 and β(k) = 1, and e(k) = 0 otherwise. Then, we have
i=1
for all k. In what follows, we will drop the argument of hi (θk ) Prob {e(k) = 1} = E {e(k)} = ē, Prob {e(k) = 0} = 1 − ē.
for brevity. Therefore, for all k, we have
Then, from (2) and (4), the closed-loop system is given by

r
hi ≥ 0, i = 1, 2, . . . , r, hi = 1.
i=1

r
r
xk +1 = hi hj [Āij xk + B1i wk ]
i=1 j =1
B. Controller

r
r
In this paper, we consider the state feedback control problem, zk = hi hj [C̄ij xk + D1i wk ] (5)
and the control strategy is the parallel distributed compensation, i=1 j =1
which is described as follows.
Controller Rule i: IF θ1 (k) is Mi1 and θ2 (k) is Mi2 and · · · where Āij = Ai + ēB2i Kj + ẽ(k)B2i Kj , C̄ij = Ci + ēD2i Kj
and θp (k) is Mip , THEN + ẽ(k)D2i Kj , and ẽ(k) = e(k) − ē. It is clear that E{ẽ(k)} =
uck = Ki xck . 0 and E{ẽ(k)ẽ(k)} = ē(1 − ē).
Before proceeding further, we first introduce the following
Here, xck ∈ Rn is the input of the controller; uck ∈ Rm is the definition.
output of the controller; and Ki is the gain matrix of the state Definition 1: The closed-loop system (5) is said to be stochas-
feedback controller. Thus, the controller can be represented by tically stable in the mean square if, when w(t) ≡ 0 for any initial
the following input–output form: condition x0 , there exists a finite W > 0 such that

r ∞

uck = hi Ki xck . (3) 2
E |xk | x0 < xT0 W x0 .
i=1
k =0
C. Communication Links Then, the problem to be addressed in this paper is expressed

Due to the existence of the communication links and the data as follows.
loss between the physical plant and controller, the measurement Problem H∞ fuzzy control under unreliable links (HFCULs):
of the plant is no longer equivalent to the input of the con- Consider the problem shown in Fig. 2, and suppose the commu-
troller (that is, xk = xck ), and the output of the controller is no nication link parameter ē is given. Given a scalar γ > 0, design
longer equivalent to the input of the plant (that is, uk = uck ). In a state feedback controller in the form of (3) such that
this paper, we model the data loss phenomena via a stochastic 1) (stochastic stability) the closed-loop system in (5) is
approach. That is, we model them as stochastically stable in the sense of Definition 1;
2) (H∞ performance) under zero initial conditions, the con-
uck = α(k)xk , uk = β(k)uck trolled output zk satisfies
where {α(k)} and {β(k)} are two independent Bernoulli pro-
cesses. The process {α(k)} models the unreliable nature of the zE ≤ γw2 (6)
link from the sensor to the controller, and β(k) models that from ∞

the controller to the actuator. More specifically, α(k) = 0 when where zE = E{ k =0 |zk | }. If the previous two
2
the link fails (that is, data are lost), and α(k) = 1 means success- conditions are satisfied, the closed-loop system is called
ful transmission [the same goes to β(k)]. A natural assumption stochastically stable with a guaranteed H∞ performance
on {α(k)} and {β(k)} can be made as γ.
Remark 1: The stochastic variable e(k) characterizes the pos-
Prob {α(k) = 1} = E {α(k)} = ᾱ
sibility of the link from the sensor to controller and then to
Prob {α(k) = 0} = 1 − ᾱ actuator being successful. It is easy to understand that the larger
the value of ē, the higher the chance of successful transmission.
Prob {β(k) = 1} = E {β(k)} = β̄
This model to characterize data missing phenomenon was first
Prob {β(k) = 0} = 1 − β̄. introduced in [31].
Based on this, we have

III. QUADRATIC APPROACH

r
uk = hi β(k)α(k)Ki xk . (4) In this section, Problem HFCUL formulated earlier will be
i=1 solved via a quadratic approach.

A. H∞ Performance Analysis 1
r
r
≤ xTk hi hj (Ai + Aj )T P (Ai + Aj )
In this section, a condition guaranteeing the stochastic sta- 4 i=1 j =1
bility and H∞ performance is established for the closed-loop
system in (5). + 2 (Ai + Aj )T P ē (B2i Kj + B2j Ki )
Theorem 1: Consider the fuzzy system in (2) and suppose the
gain matrices Ki , i = 1, . . . , r, of the subsystem controllers in T
+ ē(B2i Kj +B2j Ki ) P (B2i Kj +B2j Ki) xk
(3) are given. The closed-loop fuzzy system in (5) is stochasti-
cally stable with a guaranteed H∞ performance γ, if there exists
a matrix P > 0 satisfying − xTk P xk
r −1

ΠTii P̄ Πii + ΛTii Λii − L < 0, i = 1, . . . , r (7)
r
r
= xTk h2i ΓTii P̄ Γii − P xk + xTk hi hj
(Πij + Πj i )T P̄ (Πij + Πj i ) i=1 i=1 j =i+1

T 1
+ (Λij + Λj i ) (Λij + Λj i ) − 4L < 0 × T
(Γij + Γj i ) P̄ (Γij + Γj i ) − 2P xk
2
1≤i<j≤r (8)
where
Ai + ēB2i Kj
where Γij = . (11)
ē(1 − ē)B2i Kj
Ai + ēB2i Kj B1i P 0
Πij = , L= By the Schur complement, from inequalities (7) and (8), we
ē(1 − ē)B2i Kj 0 0 γ2 I
know that
Λij =
Ci + ēD2i Kj D1i
, P̄ =
P 0
. ΓTii P̄ Γii − P < 0, i = 1, . . . , r (12)
ē(1 − ē)D2i Kj 0 0 P
1
(9) (Γij + Γj i )T P̄ (Γij + Γj i ) − 2P < 0
2
Proof: We first prove the stochastic stability of the closed-loop 1 ≤ i < j ≤ r. (13)
system. To this end, assume wk ≡ 0, and choose a Lyapunov Define Φij = (Γij + Γj i )T P̄ (Γij + Γj i ) − 4P ; thus
1
function as r r
Φ= hi hj [(Γij + Γj i )T P̄ (Γij + Γj i ) − 4P ]
V (xk ) = xTk P xk (10) 4 i=1 j =1
r −1

where P > 0 is a matrix to be determined. Then, along the r
r
closed-loop fuzzy system in (5), we have = h2i ΓTii P̄ Γii − P xk + hi hj (14)

i=1 i=1 j =i+1
∆Vk = E{ Vk +1 | xk } − Vk
 1 T
T × (Γij + Γj i ) P̄ (Γij + Γj i ) − 2P (15)

 r r 2
=E  hi hj Āij xk  and we get

 i=1 j =1
E xTk +1 P xk +1 xk − xTk P xk ≤ −λm in (−Φ)xTk xk .
   Taking mathematical expectation of both sides and for any d ≥

r 
r

×P  hi hj Āij xk  xk − xTk P xk 1, summing up the inequality on both sides from k = 0, . . . , d,
i=1 j =1  we have
d
T T

r
r
r
r
Āij + Āj i E xd+1 P xd+1 − xT0 P x0 ≤ −λm in (−Φ)E |xk |2
= E xTk hi hj hs ht
i=1 j =1 s=1 t=1
2 k =0
which implies
Āst + Āts d
×P xk xk − xTk P xk
2 E |xk | ≤ (λm in (−Φ))−1 xT0 P x0 − E xTd+1 P xd+1
2
r r r r k =0
1
= xTk hi hj hs ht (Ai + Aj )T ≤ (λm in (−Φ))−1 xT0 P x0
4 i=1 j =1 s=1 t=1
where x0 is the initial condition. When T = 1, . . . , ∞, consid-
×P (As + At ) + (Ai + Aj )T P ē (B2s Kt + B2t Ks ) ering E{xT∞ P x∞ } ≥ 0, we have
∞
+ ē (B2i Kj + B2j Ki )T P (As + At )
2
E |xk | x0 ≤ (λm in (−Φ))−1 xT0 P x0

k =0
T
+ ē(B2i Kj +B2j Ki) P (B2s Kt +B2t Ks) xk −xTk P xk
= xT0 (λm in (−Φ))−1 P x0 = xT0 W x0
1
where W = (λm in (−Φ))−1 P . By substituting (12) and (13) (Gij + Gj i )T P̄ (Gij + Gj i )
into (14), we know Φ < 0, and therefore, λm in (−Φ) > 0. 2
Thus, W = (λm in (−Φ))−1 P > 0. According to Definition 1, 1
+ (Mij + Mj i )T (Mij + Mj i ) − 2L < 0
the closed-loop system is stochastically stable in the mean 2
square. 1≤i<j≤r
Next, the H∞ performance criterion for the closed-loop sys-
tem will be established. To this end, assume zero initial condi- where
tions. An index is introduced as
Gij = [ Ai + B2i Kj B1i ]

J = E xTk +1 P xk +1 ξk − xTk P xk Mij = [ Ci + D2i Kj D1i ] .

+E zkT zk ξk −γ 2 wkT wk .
This is the standard H∞ performance criterion for a discrete-
Defining ξk = [ xTk wkT ]T
, Ḡij = [ Āij B1i ] , and H̄ij = time fuzzy closed-loop system reported in [5, Th. 2]. When
[ C̄ij D1i ] , we have ē = 0 (that is, all the control signals are lost, and the closed-
loop system is nothing but the open-loop system), inequalities
J = E xTk +1 P xk +1 ξk + E zkT zk ξk − ξkT Lξk (7) and (8) reduce to
 T T

 r r Ai P Ai + CiT Ci − P ATi P B1i + CiT D1i
Ḡ + Ḡ −L<0
≤E  ξk  P
ij ji
hi hj

 2 ∗ T
B1i P B1i + D1iT
D1i
i=1 j =1
   i = 1, . . . , r.


r
r
Ḡij + Ḡj i
× hi hj ξk  ξk This is the standard H∞ performance criterion for an unforced
i=1 j =1
2  system reported in [5, Th. 1].
 T

 r r B. Controller Design
H̄ + H̄
+E  ξk 
ij ji
hi hj

 i=1 j =1 2 In this section, we aim to design a state feedback controller
in the form of (3) such that the closed-loop system in (5) is
  
 stochastically stable and the controlled output zk satisfies (6).

r
r
H̄ij + H̄j i
× hi hj ξk  ξk − ξkT Lξk Theorem 2: Consider the fuzzy system in (2). There exists a
i=1 j =1
2  state feedback controller in the form of (4) such that the closed-
loop system in (5) is stochastically stable with a guaranteed H∞
r −1

r
r
performance γ, if there exist matrices X > 0 and Mi , i = 1,
= ξkT h2i ΠTii P̄ Πii + ΛTii Λii −L ξk + ξkT hi hj . . . , r, satisfying
i=1 i=1 j = i+1
 
L̄ ∆Tii χTii
× [(Πij + Πj i )T P̄ (Πij + Πj i )  
 ∗ X̄ 0  < 0, i = 1, . . . , r (16)
+ (Λij + Λj i )T (Λij + Λj i ) − 4L]ξk .
∗ ∗ I¯
From inequalities (7) and (8), we know J ≤ 0, that is  
4L̄ ∆Tij + ∆Tji χTij + χTji

E xTk +1 P xk +1 ξk ) − xTk P xk  
 ∗ X̄ 0  < 0, 1≤i<j≤r

+ E{ zkT zk ξk } − γ 2 wkT wk ≤ 0. ∗ ∗ ¯
I
Taking mathematical expectation on both sides, we have (17)
E{xTk +1 P xk +1 } − E{xTk P xk } + E{zkT zk } − γ 2 wkT wk ≤ 0. where

For k = 0, 1, 2, . . . , summing up both sides, under zero initial Ai X + ēB2i Mj B1i
condition and E{xT∞ P x∞ } ≥ 0, we obtain ∆ij =
∞ ē(1 − ē)B2i Mj 0
∞

E z k zk −
T
γ 2 wkT wk ≤ 0. Ci X + ēD2i Mj D1i
χij =
k =0 k =0 ē(1 − ē)D2i Mj 0
The proof is completed.
−I 0 −X 0
Remark 2: Theorem 1 is obtained for ē ∈ [0, 1]. When ē = 1 I¯ = , L̄ =
(that is, perfect communication links exist between the plant and 0 −I 0 −γ 2 I
controller), inequalities (7) and (8) reduce to
−X 0
X̄ = . (18)
GTii P̄ Gii + MiiT Mii − L < 0, i = 1, . . . , r 0 −X
Furthermore, if the previous conditions have feasible solutions, Pj > 0, j = 1, . . . , r, satisfying

the gains Ki of the subsystem controllers in (4) are given by
ΠTii P̄j Πii + ΛTii Λii − Li < 0, i = 1, . . . , r, j = 1, . . . , r
Ki = Mi X −1 .
(23)
Proof: According to Theorem 1, if there exists P > 0 satis- T
fying (7) and (8), then the closed-loop system is stochastically Πil + Πli
P̄j (Πil + Πli )
stable with a guaranteed H∞ performance γ. By the Schur com- 2
plement, the following inequalities are obtained: T
Λil + Λli
  + (Λil + Λli ) − Li − Ll < 0
−L ΠTii ΛTii 2
 
 ∗ −P̄ −1 0  < 0, i = 1, . . . , r (19) 1 ≤ i < l ≤ r, j = 1, . . . , r (24)
∗ ∗ I¯
  where Πij , Λij are given in (9), and
−4L ΠTij + ΠTji ΛTij + ΛTji
  Pi 0 Pj 0
 ∗ −P̄ −1 0  < 0, 1 ≤ i < j ≤ r (20) Li = , P̄j = . (25)
0 γ2 I 0 Pj
∗ ∗ ¯
I
Proof: We first prove the stability of the closed-loop system.
where Πij , Λij , P̄ , and L are the same as those in Theorem 1.
To this end, assume wk ≡ 0 and choose a Lyapunov function as
Performing congruence transformations to inequalities (19) and
(20) by diag{P −1 , I, I, I, I, I}, we have r

 ˜T  Vk = xTk hi Pi xk (26)
−L̃ ∆ ii χ̃Tii
  i=1
 ∗ −P̄ −1 0  < 0, i = 1, . . . , r (21)
¯ where Pi > 0 are to be determined. Define h+ j = hj (θk +1 ).
∗ ∗ I
 Then, along the closed-loop system in (5), we have
−4L̃ ∆ ˜ +∆
T ˜ T χ̃T + χ̃T 
ij ji ij ji
 −1 
 ∗ −P̄ 0  < 0, 1≤i<j≤r
∆Vk = E { Vk +1 | xk } − Vk
∗ ∗ I¯ 
 r
(22) = E xTk h+
 j
where j =1

A P −1 + ēB2i Kj P −1
r
r
r
r
˜ ij = i
B1i
∆ × hi hl hp hq ĀTip Pj Ālq xk xk
ē (1 − ē)B2i Kj P −1 0
i=1 l=1 p=1 q =1
−1
P 0
r
L̃ = − xTk hi Pi xk
0 γ2 I

i=1
Ci P −1 + ēD2i Kj P −1 D1i
χ̃ij = .  r
ē (1 − ē)D2i Kj P −1 0 = E xTk h+
 j
j =1
Defining X = P −1 , Mi = Ki P −1 , we readily obtain (16) and T

(17), and the proof is completed.
r
r
r
r
Āip + Āpi
× hi hl hp hq
i=1 l=1 p=1 q =1
2
IV. BASIS-DEPENDENT APPROACH r
#
Ālq + Āq l
To reduce the conservativeness in the quadratic framework, ×Pj
xk xk − xk T
hi Pi xk
in this section, we present another approach based on a basis- 2 i=1
dependent Lyapunov function. r r r r
1 T +
r
= xk hj hi hl hp hq (Ai + Al )T
A. H∞ Performance Analysis 4 j =1 i=1 p=1 q =1
l=1
The following theorem presents a condition guaranteeing the × Pj (Ap +Aq )+(Ai + Al )T Pj ē (B2p Kq + B2q Kp )
H∞ performance for the closed-loop system in (5).
Theorem 3: Consider the fuzzy system in (2), and suppose the + ē (B2i Kl + B2l Ki )T Pj (Ap + Aq )
gain matrices Ki , i = 1, . . . , r of the subsystem controllers (3)
are given. The closed-loop system in (5) is stochastically stable T
+ē (B2i Kl + B2l Ki ) Pj (B2p Kq + B2q Kp ) xk
with a guaranteed H∞ performance γ, if there exist matrices
r

r
r
− xTk hi Pi xk ≤ ξkT h+
j h2i ΠTii P̄j Πii + ΛTii Λii − Li
i=1 j =1 i=1
r −1 T

r
r
r r −1

r
Πil + Πli
≤ xTk h+
j h2i ΓTii P̄j Γii − Pi + + hi hl P̄j (Πil + Πli )
j =1 i=1 i=1 i< l i=1 l=i+1
2
T T
Γil + Γli Λil + Λli
×hi hl P̄j (Γil + Γli ) − Pi − Pl xk . + (Λil + Λli ) − Li − Ll ξk .
2 2
From inequalities (23) and (24), we know that From inequalities (23) and (24), we have J¯ ≤ 0, and thus
ΓTii P̄j Γii − Pi < 0, i = 1, . . . , r, j = 1, . . . , r

r
T E {Vk+1 | ξk} +E zkT zk ξk − γ 2 wkT wk − ξkT hi Pi ξk ≤ 0.

Γil + Γli i=1
P̄j (Γil + Γli ) − Pi − Pl < 0
2
Taking mathematical expectation on both sides, we obtain
1 ≤ i < l ≤ r, j = 1, . . . , r
E{Vk +1 } − E{Vk } + E{zkT zk } − γ 2 wkT wk ≤ 0.
where Γij is defined in (11). By defining
For k = 0, 1, 2, . . . , summing up both sides, under zero initial
r −1
condition and E{V∞ } > 0, we obtain

r r
r
Φ= h+ h2i ΓTii P̄j Γii − Pi +
j ∞ ∞
j =1 i=1 i=1 i< l
T E z k zk −
T
γ 2 wkT wk ≤ 0.
Γil + Γli
× h i hl P̄j (Γil + Γli ) − Pi − Pl k =0 k =0
2
The proof is completed.
the stochastic stability of the closed-loop system can be proved Remark 3: Theorem 3 is obtained by using the basis-
by following similar lines as the proof of Theorem 1. dependent Lyapunov function defined in (26). It is clear that
Next, the H∞ performance criteria for the closed-loop system when Pi = P for any i ∈ {1, . . . , r}, (26) becomes the quadratic
will be established. To this end, assume zero initial conditions. Lyapunov function that is widely used in the literature. When
An index is introduced as Pi = P , Theorem 3 reduces to Theorem 1.

r
¯
J=E { Vk +1 | ξk } +E zkT zk ξk −γ 2 wkT wk − ξkT hi Pi ξk B. Controller Design
i=1
It is worth mentioning that it is usually difficult to obtain
where ξk is the same as that in the proof of Theorem 1. LMI conditions for the controller design in the basis-dependent
Since framework, unless additional overdesign is introduced. In the

r
r
r following, we present a controller design procedure based on
Vk +1 ≤ ξkT h+
j hi hp
the cone complementarity linearization (CCL) algorithm.
j =1 i=1 p=1 Theorem 4: Consider the fuzzy system in (2). A state feed-
back controller in the form of (4) exists, such that the closed-loop
ĀTip system in (5) is stochastically stable with a guaranteed H∞ per-
× P̄j [ Āip B1i ] ξk ,
T
B1i formance γ, if there exist matrices Xi > 0, Pi > 0, i = 1, . . . , r
satisfying

r
r
zkT zk ≤ ξkT hi hp  
−X̄j 0 Πii
i=1 p=1  
T
 ∗ −I Λii  < 0, i = 1, . . . , r, j = 1, . . . , r
C̄ip
× [ C̄ip D1i ] ξk , ∗ ∗ −Li
T
D1i
(27)
 
we have −2X̄j 0 Πil + Πli
 
 ∗ −2I Λil + Λli  < 0, 1≤i<l≤r

r
r
r
∗ ∗ −(Li + Ll )
J¯ ≤ ξkT h+
j hi hp
j =1 i=1 p=1 j = 1, . . . , r (28)
$ %
× ΠTip P̄j Πip + ΛTip Λip − Li ξk Pi Xi = I, i = 1, . . . , r. (29)

where isfied, with |tr ( ri=1 Pi Xi ) − rn| < δ, for some suf-
ficiently small scalar δ > 0, then output the feasible
Xj 0
X̄j = solutions (Pi , Xi , Ki ) . EXIT.
0 Xj
Step 4: If K > N where N is the maximum number of itera-
Πij , and Λij are given in (9), and Li is defined in (25). tions allowed, EXIT.
Proof: According to the Schur complement, (23) and (24) can (t) (t) (t)
Step 5: Set t = t + 1, Pi , Xi , Ki = (Pi , Xi , Ki ) , and
be converted to go to Step 2.
 
−P̄j−1 0 Πii
  V. EXTENSIONS AND DISCUSSIONS
 ∗ −I Λii  < 0, i = 1, . . . , r, j = 1, . . . , r
∗ ∗ −Li Parameter uncertainties often exist in practical applications
  [16], [17], [43]. In this section, the results obtained previously
−2P̄j−1 0 Πil + Πli are extended to fuzzy systems with time-varying uncertainties.
 
 ∗ −2I Λil + Λli  < 0, 1≤i<l≤r The robust H∞ performance criterion of fuzzy systems with un-
∗ ∗ − (Li + Ll ) certainties is presented and the robust H∞ controller is designed.
All these are considered based on the quadratic approach, and the
j = 1, . . . , r. results based on the basis-dependent approach can be obtained
By defining by similar procedures. Moreover, the advantages and drawbacks
of the quadratic approach and the basis-dependent approach are
Xj 0 Pj−1 0 discussed.
X̄j = =
0 Xj 0 Pj−1
A. Extensions
(27) and (28) can be obtained.
It should be noted that the condition in Theorem 4 is not a In this section, the robust H∞ control of fuzzy system with
convex set due to the existence of the matrix inequalities in (29). uncertainties is considered. Under control law (3), the closed-
Several approaches have been proposed to solve such nonconvex loop fuzzy system with parameter uncertainties becomes
feasibility problems, among which the CCL method [10] is the
r
r
most commonly used one (for instance, the CCL algorithm has xk +1 = hi hj [Ãij xk + B̃1i wk ]
been used for solving controller design problems [15]). The i=1 j =1
basic idea in the CCL algorithm is that if the LMI [ PI LI ] ≥ 0

r
r
is feasible in the n × n matrix variables L > 0 and P > 0, then zk = hi hj [C̃ij xk + D̃1i wk ] (31)
tr(P L) ≥ n, and tr(P L) = n if and only if P L = I. i=1 j =1
Now, using a cone complementarity approach [10], we sug- where
gest the following nonlinear minimization problem involving
LMI conditions instead of the original nonconvex feasibility Ãij = (Ai + ∆Ai ) + ē(B2i + ∆B2i )Kj
problem formulated in Theorem 4: + ẽ(k)(B2i + ∆B2i )Kj , B̃1i = B1i + ∆B1i
& r '

min tr Pi Xi subject to (27), (28), and C̃ij = (Ci + ∆Ci ) + ē(D2i + ∆D2i )Kj
i=1 + ẽ(k)(D2i + ∆D2i )Kj , D̃1i = D1i + ∆D1i .

Pi I
≥ 0, i = 1, . . . , r. (30) Here, Ai , B2i , B1i , Ci , D2i , and D1i are the same as those in (2),
I Xi and ∆Ai , ∆B2i , ∆B1i , ∆Ci , ∆D2i , and ∆D1i are real-valued
According to [10], if the solution
of the previous minimization matrices with appropriate dimensions satisfying [34]
problem is rn, that is, min tr( ri=1 Pi Xi ) = rn, then the con-
∆Ai ∆B2i ∆B1i E1i
ditions in Theorem 4 are solvable. The following algorithm is = Fi (k) [ H1i H2i H3i ]
suggested to solve the previous problem. ∆Ci ∆D2i ∆D1i E2i
Algorithm 1 where E1i , E2i , H1i , H2i , and H3i are known constant matri-
(0) (0) (0) ces with compatible dimensions; Fi (k) are unknown nonlinear
Step 1: Find a feasible set (Pi , Xi , Ki ) satisfying (27),
time-varying matrix functions satisfying FiT (k)Fi (k) ≤ I. It is
(28), and (30). Set t = 0.
assumed that the elements of Fi (k) are Lebesgue measurable.
Step 2: Solve the following LMI problem:
The following theorems give the results of robust H∞ per-
& r ' formance analysis and controller design for fuzzy systems with
( (t) (t)
) parameter uncertainties, which can be proved by following sim-
min tr Pi X i + P i Xi
ilar lines in [14].
i=1
1) Robust H∞ Performance Analysis:
subject to (27), (28), and (30). Theorem 5: Suppose the gain matrices Ki , i = 1, . . . , r of the
Step 3: Substitute the obtained matrix variables (Pi , Xi , Ki ) subsystem controllers in (3) are given. The closed-loop fuzzy
into (23) and (24). If conditions (23) and (24) are sat- system in (31) is robustly stochastically stable with a guaranteed
H∞ performance γ, if there exist a matrix P > 0, scalars εi , εij , B. Discussions

satisfying
This paper is targeted to solve the problem of H∞ fuzzy
 −P̄ 0 P̄ Π i i P̄ Ē 1 i 0  control with unreliable communication links, which has not
 ∗ −I Λi i Ē 2 i 0  been solved in the previous literature. Data loss has been rec-
  ognized as an important issue for the analysis and synthesis of
 
 ∗ ∗ −L 0 εi T iTi  < 0, i = 1, . . . , r (32) communication-based control systems. Although there are a few
 
 ∗ ∗ ∗ εi I 0  attempts toward this problem, most of the existing results in this
∗ ∗ ∗ ∗ εi I direction are limited to linear systems. On the other hand, the
 −P̄  T–S fuzzy model has been recognized to be an effective means
0 P̄ Π i j + Π j i P̄ Ē 1 i P̄ Ē 1 j 0 0
for nonlinear systems, while most of the results on fuzzy con-
 ∗ −I Λi j + Λj i Ē 2 i Ē 2 j 0 0 
  trol are based on the perfect communication assumption. To fill
 ∗ ∗ −4L
 the gap, the problem of H∞ fuzzy control of nonlinear systems
 0 0 εi j T iTj εj i T jTi 
  under unreliable communication links is investigated in this pa-
 ∗ ∗ ∗ εi j I 0 0 0 <0
  per, and the results are good complements to the analysis and
 
 ∗ ∗ ∗ ∗ εj i I 0 0 
  synthesis of communication-based control systems.
 ∗ ∗ ∗ ∗ ∗ εi j I 0  To solve the problem, we have developed two approaches
∗ ∗ ∗ ∗ ∗ ∗ εj i I based on the quadratic Lyapunov function and the basis-
dependent Lyapunov function, respectively. These approaches
1≤i<j≤r
provide alternatives for designing H∞ controllers with different
where Πij , L, Λij , and P̄ are defined in (9), and degrees of conservativeness and computational complexity. It is
shown that with the use of basis-dependent Lyapunov function,
H1i + ēH3i Kj H2i less conservative results can be obtained than with the use of sin-
Tij = gle quadratic Lyapunov function, while the computational cost
ē(1 − ē)H3i Kj 0
of the former is higher than the latter. Most results in this paper

E1i 0 E2i 0 are formulated in the form of LMIs, which can be solved eas-
Ē1i = , Ē2i = . (33) ily by numerical software. It is usually difficult to obtain strict
0 E1i 0 E2i
LMI conditions for the controller design in the basis-dependent
2) Controller Design: framework, and a CCL algorithm is introduced to solve the non-
Theorem 6: There exists a state feedback controller in the form convex feasibility problem, which provides an alternative for
of (4) such that the closed-loop fuzzy system in (31) is robustly basis-dependent fuzzy control.
stochastically stable with a guaranteed H∞ performance γ, if The quadratic Lyapunov function approach has been widely
there exist matrices X > 0 and Mi , scalars εi , εij , satisfying used in the literature [5], [20], [25]–[27], [36], [38]–[40] since
 T T T  it has some advantages: The definition of Lyapunov function is
L̄ ∆i i χi i 0 εi H 1 i i
simple, and the computation cost in the design procedure is low.
∗ X̄ 0 Ē 1 i 0 
∗  However, the common quadratic Lyapunov functions tend to be
 ∗ I¯ Ē 2 i 0 <0
∗ ∗ ∗ εi I¯ 0
 conservative and might not exist for some highly nonlinear sys-
∗ ∗ ∗ ∗ εi I¯ tems. To reduce the conservatism, basis-dependent Lyapunov
functions have been adopted in [9], [11], [12], [18], [19], [35],
i = 1, . . . , r (34)
[37], [45], [50], and [51]. It can be seen that the quadratic
 4L̄ ∆ T + ∆ T χT + χT 0 0 ε H T
ε H T 
ij ji ij ji ij 1ij ij 1j i Lyapunov function is a special case of the basis-dependent Lya-
 ∗ X̄ 0 Ē 1 i Ē 1 j 0 0  punov function, and thus, the latter is more general and less
 ∗ ∗ ¯ 
 I Ē 2 i Ē 2 j 0 0  conservative. Meanwhile, the basis-dependent Lyapunov func-
 ∗ ∗ ∗ εi j I¯ 0 0 0 <0
  tion also has some disadvantages; the design procedures become
 ∗ ∗ ∗ ∗ εj i I¯ 0 0 
  more complex, and the computational requirement is usually
∗ ∗ ∗ ∗ ∗ ¯
εi j I 0
demanding.
∗ ∗ ∗ ∗ ∗ ∗ εj i I¯
1≤i<j≤r (35) VI. ILLUSTRATIVE EXAMPLE
¯ and L̄ are defined in (18), and
where ∆ij , χij , X̄, I, In this section, we present three examples to illustrate the
theoretical results developed earlier.
H1i X + ēH3i Mj H2i Example 1: In this example, our aim is to compare Theo-
H1ij = . rems 1 and 3 to show that the basis-dependent approach is less
ē(1 − ē)H3i Mj 0
conservative. Consider the following system:
Furthermore, if the previous conditions have feasible solutions, Model Rule 1 : IF xk (1) is h1 (xk (1))
the gains Ki of the subsystem controllers in (4) are given by *
xk +1 = A1 xk + B21 uk + B11 wk
Ki = Mi X −1 . THEN
zk = C1 xk + D21 uk + D11 wk
TABLE I
COMPARISON OF THE MINIMUM H ∞ PERFORMANCE γ ∗ BY DIFFERENT
APPROACHES (THEOREMS 1 AND 3)
Model Rule 2 : IF xk (1) is h2 (xk (1))

*
xk +1 = A2 xk + B22 uk + B12 wk
THEN
zk = C2 xk + D22 uk + D12 wk
(36)
where

1.5 −0.5 1
A1 = , B21 =
1 0 1

Fig. 3. Closed-loop state variables when ē = 1.
0.2
B11 = , C1 = [ 1 0.5 ]
0.3
where
−1 −0.5 −2
A2 = , B22 =
1 0 1 1 −0.5 1
A1 = , B21 =
1 0 1
0.5
B12 = , C2 = [ 0.5 1 ]
−0.1 −1 −0.5 −2
A2 = , B22 = .
D11 = 0.4, D12 = 0.2, D21 = 1, D22 = 0.5. 1 0 1
The controller is given by
The fuzzy controller is given by
Model Rule 1 : IF xk (2) is h1 (xk (2))
Model Rule 1 : IF xk (1) is h1 (xk (1)) THEN uck = K1 xck
THEN uck = K1 xck Model Rule 2 : IF xk (2) is h2 (xk (2))
Model Rule 2 : IF xk (1) is h2 (xk (1)) THEN uck = K2 xck (38)
THEN uck = K2 xck
where K1 = [0.7460 0.1240], and K2 = [−0.9988 0.8724] .
We assume the membership function to be
where K1 = [ −0.65 0.5 ] , and K2 = [ −0.87 0.11 ] . 
The aim is to compare the H∞ performances of the closed- 
 1, xk (2) ≤ −1
loop system obtained by Theorems 1 and 3. The minimum H∞ h1 (xk (2)) = 0.5 − 0.5xk (2), −1 ≤ xk (2) ≤ 1


performances obtained by using Theorems 1 and 3 are listed in 0, xk (2) ≥ 1
Table I, where the communication link parameter ē is selected
from 0.7 to 0.9. h2 (xk (2)) = 1 − h1
Table I shows that the minimum H∞ performance γ ∗ obtained
and assume the initial condition to be x0 = [ 0.9 −0.7 ]T .
by Theorem 3 is much smaller than that by Theorem 1, showing
First, when ē = 1 (corresponding to perfect communication
that Theorem 3 is less conservative. It is interesting to point out
links), Fig. 3 shows the state response of the closed-loop system,
that when ē = 0.6833, Theorem 1 fails to find solutions, while
from which we can see that all the state variables converge to
Theorem 3 still yields feasible solutions.
zero. However, when ē = 0.8 (corresponding to the case that
Example 2: In this example, we aim to motivate the problem
20% data are lost during the communication), the state variables
considered in this paper, and to show the usefulness of our
no longer converge to zero, as shown in Fig. 4. The data missing
results. Consider the following fuzzy system:
is randomly generated according the probability shown in Fig. 5.
Model Rule 1 : IF xk (2) is h1 (xk (2)) Now, for ē = 0.8, we apply Theorem 2, and the following
feasible solutions are obtained:
THEN xk +1 = A1 xk + B21 uk
2.0920 0.0089
Model Rule 2 : IF xk (2) is h2 (xk (2)) X=
0.0089 5.5432
THEN xk +1 = A2 xk + B22 uk (37) K1 = [ −0.8093 −0.0075 ] , K2 = [ −0.6742 −0.0128 ] .
Fig. 6. Closed-loop state variables when ē = 0.8.

Fig. 4. Closed-loop state variables when ē = 0.8 with controller gain matrices
in (38).
Fig. 7. Inverted pendulum controlled by a dc motor.
where
 
1.0002 0.02 0.02
 
A1 =  0.196 1.0001 0.0181 
−0.0184 −0.1813 0.8170
 
1 0.02 0.0002
 
A2 =  0 0.9981 0.0181 
−0.1811
0 0.8170
   
0 0
Fig. 5. Random data loss.    
B21 =  0.0019  , B22 =  0.0019 
0.1811 0.1811
Fig. 6 shows that the controller designed by Theorem 2 makes F1 = [ 1 0 0], F2 = [ 1 0 0].
the closed-loop state variables converge to zero. This well mo- For simulation, we add some disturbance terms and a controlled
tivates the necessity to consider the communication issue in the output
control design.
Example 3: Consider an inverted pendulum controlled by a
2
xk +1 = hi (x1k ) (Ai xk + B2i uk + B1i wk )
dc motor via a gear train shown in Fig. 7, whose fuzzy modeling
i=1
was done in [24]. With a sampling time T = 0.02, we have the
discrete-time model as
2
zk = hi (x1k ) (Ci xk + D2i uk + D1i wk ) (39)

2 i=1
xk +1 = hi (x1k ) (Ai xk + B2i uk ) where
i=1    
−0.054 0.054

2
B12 =  −0.094  , B11 =  0.094 
yk = hi (x1k )Fi xk
0 0
i=1
Fig. 8. Closed-loop state variables when w(t) ≡ 0. Fig. 9. Closed-loop state variables with w(k) in (41).
 T  T
0.054 0.054
C1 =  0.005  , C2 =  0.005 
0.1 0.1
D11 = 0.1, D12 = 0.1.
Our aim is to design a state feedback paralleled controller in
the form of (4) such that the system (39) is stochastically stable
with a guaranteed H∞ performance γ.
Without loss of generality, the condition ē = 0.8 is discussed.
By solving LMIs (16) and (17), the minimum H∞ performance
γ ∗ = 0.9332 is obtained, and the feasible solutions are
 
0.4021 −0.3768 −1.0765
X =  −0.3768 0.5258 −0.3634 
−1.0765 −0.3634 24.7426
K1 = [ −75.3905 −52.8595 −6.8464 ]
Fig. 10. Signals w(k) and z(k).
K2 = [ −60.9557 −49.9968 −4.8232 ] . (40)
To show the effectiveness of the obtained results, we assume the
membership function to be Fig. 9 shows the changing curves of the state variables, and
 Fig. 10 shows the signals w(k) and z(k). It is found that z2 =
 1, x0 (1) = 0 0.4164 and w2 = 0.9381, which yields γ = 0.4439 (below
h1 = sin(x0 (1))
 the minimum γ ∗ = 0.9332), showing the effectiveness of the
, else
x0 (1) controller design.
h 2 = 1 − h1 .
VII. CONCLUDING REMARKS
First, we assume the initial condition to be x0 =
[ 0.8 −0.4 −0.8 ]T . Fig. 8 shows the state responses of the In this paper, the problem of H∞ fuzzy control of nonlinear
closed-loop system when the external disturbance w(k) ≡ 0, systems under unreliable communication links has been inves-
from which we can see that the two states converge to zero. tigated. The T–S fuzzy model is utilized to model the nonlin-
Next, to illustrate the disturbance attenuation performance, ear plant, and the communication link failure is modeled via a
assume zero initial condition, and the external disturbance w(k) stochastic variable satisfying the Bernoulli random binary dis-
is assumed to be tribution. Two approaches have been developed to design H∞
 controllers such that the closed-loop system is stochastically

 0.2, 20 ≤ k ≤ 30
stable and preserves a guaranteed H∞ performance. Three ex-
w(k) = −0.2, 40 ≤ k ≤ 50 (41) amples have been given to illustrate the developed theoretical


0, else. results.
REFERENCES [24] S. Kawamoto, “Nonlinear control and rigorous stability analysis based on
fuzzy system for inverted pendulum,” in Proc. FUZZ-IEEE, 1996, vol. 2,
[1] W. Assawinchaichote and S. K. Nguang, “H ∞ filtering for fuzzy sin- pp. 1427–1432.
gularly perturbed systems with pole placement constraints: An LMI ap- [25] E. Kim and H. Lee, “New approaches to relaxed quadratic stability con-
proach,” IEEE Trans. Signal Process., vol. 52, no. 6, pp. 1659–1667, Jun. dition of fuzzy control systems,” IEEE Trans. Fuzzy Syst., vol. 8, no. 5,
2004. pp. 523–534, Oct. 2000.
[2] W. Assawinchaichote and S. K. Nguang, “Fuzzy H ∞ output feedback [26] H. J. Lee, J. B. Park, and G. R. Chen, “Robust fuzzy control of nonlinear
control design for singularly perturbed systems with pole placement con- systems with parametric uncertainties,” IEEE Trans. Fuzzy Syst., vol. 9,
straints: An LMI approach,” IEEE Trans. Fuzzy Systems, vol. 14, no. 3, no. 2, pp. 369–379, Apr. 2001.
pp. 361–371, Jun. 2006. [27] S. Y. Li and X. X. Zhang, “Fuzzy logic controller with interval-valued
[3] B. Azimi-Sadjadi, “Stability of networked control systems in the presence inference for distributed parameter system,” Int. J. Innov. Comput., Inf.
of packet losses,” in Proc. 42th IEEE Conf. Decis. Control, 2003, pp. 676– Control, vol. 2, no. 6, pp. 1197–1206, 2006.
682. [28] C. Lin, Q. G. Wang, and T. H. Lee, “Stability and stabilization of a class of
[4] X. Ban, X. Gao, X. Huang, and H. Yin, “Stability analysis of the simplest fuzzy time-delay descriptor systems,” IEEE Trans. Fuzzy Syst., vol. 14,
Takagi–Sugeno fuzzy control system using Popov criterion,” Int. J. Innov. no. 4, pp. 542–551, Aug. 2006.
Comput., Inf. Control, vol. 3, no. 5, pp. 1087–1096, 2007. [29] H. Liu, F. Sun, and Y. N. Hu, “H ∞ control for fuzzy singularly perturbed
[5] Y. Y. Cao and P. M. Frank, “Robust H ∞ disturbance attenuation for a systems,” Fuzzy Sets Syst., vol. 155, pp. 272–291, 2005.
class of uncertain discrete-time fuzzy systems,” IEEE Trans. Fuzzy Syst., [30] H. Liu, F. Sun, and Z. Q. Sun, “Stability analysis and synthesis of fuzzy
vol. 8, no. 4, pp. 406–415, Aug. 2000. singularly perturbed systems,” IEEE Trans. Fuzzy Syst., vol. 13, no. 2,
[6] Y. Y. Cao and Z. L. Lin, “Robust stability analysis and fuzzy-scheduling pp. 273–284, Apr. 2005.
control for nonlinear systems subject to actuator saturation,” IEEE Trans. [31] N. Nahi, “Optimal recursive estimation with uncertain observation,” IEEE
Fuzzy Syst., vol. 11, no. 1, pp. 57–67, Feb. 2003. Trans. Inf. Theory, vol. IT-15, no. 4, pp. 457–462, Jul. 1969.
[7] B. Chen and X. P. Liu, “Delay-dependent robust H ∞ control for T–S [32] M. Sahebsara, T. Chen, and S. L. Shah, “Optimal filtering with random
fuzzy systems with time delay,” IEEE Trans. Fuzzy Syst., vol. 13, no. 4, sensor delay, multiple packet dropout and uncertain observations,” Int. J.
pp. 544–556, Aug. 2005. Control, vol. 80, no. 2, pp. 292–301, 2007.
[8] C. Chen, G. Feng, D. Sun, and X. Guan, “H ∞ output feedback control [33] P. Seiler and R. Sengupta, “An H ∞ approach to networked control,” IEEE
of discrete-time fuzzy systems with application to chaos control,” IEEE Trans. Autom. Control, vol. 50, no. 3, pp. 356–364, Mar. 2005.
Trans. Fuzzy Syst., vol. 13, no. 4, pp. 531–543, Aug. 2005. [34] T. Takagi and M. Sugeno, “Fuzzy identification of systems and its ap-
[9] D. J. Choi and P. Park, “H ∞ state-feedback controller design for discrete- plications to modeling and control,” IEEE Trans. Syst., Man, Cybern.,
time fuzzy systems using fuzzy weighting-dependent Lyapunov func- vol. SMC-15, no. 1, pp. 116–132, 1985.
tions,” IEEE Trans. Fuzzy Systems, vol. 11, no. 2, pp. 271–278, Apr. [35] K. Tanaka, H. Ohtake, and H. O. Wang, “A descriptor system approach to
2003. fuzzy control system design via fuzzy Lyapunov functions,” IEEE Trans.
[10] L. El Ghaoui, F. Oustry, and M. Ait Rami, “A cone complementar- Fuzzy Syst., vol. 15, no. 3, pp. 333–341, Jun. 2007.
ity linearization algorithm for static output-feedback and related prob- [36] K. Tanaka and M. Sugeno, “Stability analysis and design of fuzzy sys-
lems,” IEEE Trans. Autom. Control, vol. 42, no. 8, pp. 1171–1176, Aug. tems,” Fuzzy Sets Syst., vol. 45, no. 2, pp. 135–156, Jan. 1992.
1997. [37] T. Tanaka, T. Hori, and H. O. Wang, “A fuzzy Lyapunov approach to
[11] G. Feng, “Approaches to quadratic stabilization of uncertain fuzzy dy- fuzzy control system design,” in Proc. Amer. Control Conf., Arlington,
namic systems,” IEEE Trans. Circuits Syst. I, vol. 48, no. 6, pp. 760–769, VA, 2001, pp. 4790–4795.
Jun. 2001. [38] T. Tanaka, T. Ikeda, and H. O. Wang, “Fuzzy regulators and fuzzy ob-
[12] G. Feng, “H ∞ controller design of fuzzy dynamic systems based on servers: Relaxed stability conditions and LMI-based designs,” IEEE
piecewise Lyapunov functions,” IEEE Trans. Syst., Man, Cybern., vol. 34, Trans. Fuzzy Syst., vol. 6, no. 5, pp. 250–265, May 1998.
no. 1, pp. 283–292, Feb. 2004. [39] M. C. M. Teixeira and S. H. Zak, “Stabilizing controller design for un-
[13] G. Feng, “A survey on analysis and design of model-based fuzzy control certain nonlinear systems using fuzzy models,” IEEE Trans. Fuzzy Syst.,
systems,” IEEE Trans. Fuzzy Syst., vol. 14, no. 5, pp. 676–697, Oct. vol. 7, no. 4, pp. 133–142, Apr. 1999.
2006. [40] H. O. Wang, K. Tanaka, and M. F. Griffin, “An approach to fuzzy control
[14] H. Gao, J. Lam, and C. Wang, “Robust energy-to-peak filter design for of nonlinear systems: Stability and design issues,” IEEE Trans. Fuzzy
stochastic time-delay systems,” Syst. Control Lett., vol. 55, no. 2, pp. 101– Syst., vol. 4, no. 2, pp. 14–23, Feb. 1996.
111, 2006. [41] Z. Wang, D. W. C. Ho, and X. Liu, “Variance-constrained filtering for
[15] H. Gao and C. Wang, “Comments and further results on ‘A de- uncertain stochastic systems with missing measurements,” IEEE Trans.
scriptor system approach to H ∞ control of linear time-delay sys- Autom. Control, vol. 48, no. 7, pp. 1254–1258, Jul. 2003.
tems’,” IEEE Trans. Autom. Control, vol. 48, no. 3, pp. 520–525, Mar. [42] Z. Wang, D. W. C. Ho, and X. Liu, “Variance-constrained control for
2003. uncertain stochastic systems with missing measurement,” IEEE Trans.
[16] H. Gao and C. Wang, “A delay-dependent approach to robust L 2 -L ∞ Syst., Man, Cybern. A, Syst., Humans, vol. 35, no. 5, pp. 746–753, Sep.
filtering for a class of uncertain nonlinear time-delayed systems,” IEEE 2005.
Trans. Autom. Control, vol. 48, no. 9, pp. 1661–1666, Sep. 2003. [43] Z. Wang, B. Huang, and H. Unbehauen, “Robust H ∞ observer design of
[17] H. Gao and C. Wang, “A delay-dependent approach to robust H ∞ filtering linear state delayed systems with parametric uncertainty: The discrete-
for uncertain discrete-time state-delayed systems,” IEEE Trans. Signal time case,” Automatica, vol. 35, pp. 1161–1167, 1999.
Process., vol. 52, no. 6, pp. 1631–1640, Jun. 2004. [44] H. Wu, “Reliable LQ fuzzy control for continuous-time nonlinear systems
[18] H. Gao, Z. Wang, and C. Wang, “Improved H ∞ control of discrete-time with actuator faults,” IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 34,
fuzzy systems: A cone complementarity linearization approach,” Inf. Sci., no. 4, pp. 1743–1752, Aug. 2004.
vol. 175, no. 1–2, pp. 57–77, 2005. [45] H. Wu and H. Li, “New approach to delay-dependent stability analysis and
[19] T. M. Guerra and L. Vermeiren, “LMI-based relaxed nonquadratic stabi- stabilization for continuous-time fuzzy systems with time-varying delay,”
lization conditions for nonlinear systems in the Takagi–Sugeno’s form,” IEEE Trans. Fuzzy Syst., vol. 15, no. 3, pp. 482–493, Jun. 2007.
Automatica, vol. 40, no. 5, pp. 823–829, 2004. [46] H. Wu and H. Y. Zhang, “Reliable H ∞ fuzzy control for continuous-
[20] H. Han, “Fuzzy controller for a class of nonlinear systems,” Int. J. Innov. time nonlinear systems with actuator failures,” IEEE Trans. Fuzzy Syst.,
Comput., Inf. Control, vol. 1, no. 4, pp. 727–742, 2005. vol. 14, no. 5, pp. 609–618, Oct. 2006.
[21] O. C. Imer, S. Yuksel, and T. Basar, “Optimal control of LTI systems over [47] S. Xu and J. Lam, “Robust H ∞ control for uncertain discrete-time-delay
unreliable communication links,” Automatica, vol. 42, no. 9, pp. 1429– fuzzy systems via output feedback controllers,” IEEE Trans. Fuzzy Syst.,
1439, 2006. vol. 13, no. 1, pp. 82–93, Feb. 2005.
[22] H. Ishii and B. A. Francis, Limited Data Rate in Control Systems With Net- [48] M. Yu, L. Wang, T. Chu, and F. Hao, “Stabilization of networked control
works (Lecture Notes in Control and Information Sciences 275). Berlin, systems with data packet dropout and transmission delays: continuous-
Germany: Springer-Verlag, 2002. time case,” Eur. J. Control, vol. 11, no. 1, pp. 40–55, 2005.
[23] H. Katayama and A. Ichikawa, “H ∞ control for sampled-data nonlinear [49] D. Yue, Q. L. Han, and J. Lam, “Network-based robust H ∞ con-
systems described by Takagi–Sugeno fuzzy systems,” Fuzzy Sets Syst., trol of systems with uncertainty,” Automatica, vol. 41, pp. 999–1007,
vol. 148, no. 3, pp. 431–452, 2004. 2005.
[50] S. Zhou, G. Feng, J. Lam, and S. Xu, “Robust H ∞ control for discrete- Yan Zhao received the B.S. degree in chemical engi-
time fuzzy systems via basis-dependent Lyapunov functions,” Inf. Sci., neering and equipment control and the M.S. degree
vol. 174, no. 3/4, pp. 197–217, 2005. in mechanical engineering from Inner Mongolia Uni-
[51] S. Zhou, J. Lam, and W. X. Zheng, “Control design for fuzzy systems based versity of Technology, Hohhot, China, in 2002 and
on relaxed nonquadratic stability and H ∞ performance conditions,” IEEE 2005, respectively. She is currently working toward
Trans. Fuzzy Syst., vol. 15, no. 2, pp. 188–199, Apr. 2007. the Ph.D. degree in control science and engineer-
ing with the Harbin Institute of Technology, Harbin,
China.
Her current research interests include fuzzy con-
trol systems, robust control, and networked control
systems.
Huijun Gao (M’06) was born in Heilongjiang,

China, in 1976. He received the M.S. degree in elec-
trical engineering from Shenyang University of Tech-
nology, Shenyang, China, in 2001, and the Ph.D. de- Tongwen Chen (S’86–M’91–SM’97–F’06) received
gree in control science and engineering from Harbin the B.Eng. degree in automation and instrumentation
Institute of Technology, Harbin, China, in 2005. from Tsinghua University, Beijing, China, in 1984
From November 2003 to August 2004, he was a and the M.A.Sc. and Ph.D. degrees in electrical engi-
Research Associate in the Department of Mechan- neering from the University of Toronto, Toronto, ON,
ical Engineering, University of Hong Kong, Hong Canada, in 1988 and 1991, respectively.
Kong. In November 2004, he joined Harbin Insti- From 1991 to 1997, he was an Assistant/Associate
tute of Technology, where he is currently a Professor. Professor with the Department of Electrical and Com-
From October 2005 to October 2007, he was a Postdoctoral Researcher in the puter Engineering, University of Calgary, Calgary,
Department of Electrical and Computer Engineering, University of Alberta, AB, Canada. Since 1997, he has been with the De-
Edmonton, AB, Canada. He is an Associate Editor of the Journal of Intelligent partment of Electrical and Computer Engineering,
and Robotics Systems, Circuits, System and Signal Processing, etc. His current University of Alberta, Edmonton, AB, where he is currently a Professor. He
research interests include network-based control, robust control, and time-delay held visiting positions at the Hong Kong University of Science and Technology,
systems and their industrial applications. Tsinghua University, and Kumamoto University. He was the coauthor of the
Prof. Gao is an Associate Editor of the IEEE TRANSACTIONS ON SYS- book Optimal Sampled-Data Control Systems (Springer-Verlag, 1995). He was
TEMS, MAN AND CYBERNETICS PART B: CYBERNETICS and the IEEE TRANS- a McCalla Professor for 2000–2001 and a Killam Professor for 2006–2007 at
ACTIONS ON INDUSTRIAL ELECTRONICS. He was an outstanding reviewer for the the University of Alberta. He was an Associate Editor for several international
IEEE TRANSACTIONS ON AUTOMATIC CONTROL AND AUTOMICA in 2008 and journals, including Automatica and Systems and Control Letters. His current re-
2007, respectively, and an appreciated reviewer for the IEEE TRANSACTIONS ON search interests include computer- and network-based control systems, process
SIGNAL PROCESSING in 2006. He was the recipient of the University of Alberta control, multirate digital signal processing, and their applications to industrial
Dorothy J. Killam Memorial Postdoctoral Fellow Prize in 2005, the National problems.
Outstanding Youth Science Fund in 2008, and the National Outstanding Doc- Prof. Chen was an Associate Editor for the IEEE TRANSACTIONS ON
toral Thesis Award in 2007. He was the corecipient of the National Natural AUTOMATIC CONTROL. He is a registered Professional Engineer in Alberta.
Science Award of China in 2008. He received the Fellowship from the Japan Society for the Promotion of Sci-
ence for 2004.
An Adaptive Consensus Support Model for Group

Decision-Making Problems in a Multigranular Fuzzy
Linguistic Context
Francisco Mata, Luis Martı́nez, and Enrique Herrera-Viedma
Abstract—Different consensus models for group decision-

making (GDM) problems have been proposed in the literature.
However, all of them consider the consensus reaching process a
rigid or inflexible one because its behavior remains fixed in all
rounds of the consensus process. The aim of this paper is to im-
prove the consensus reaching process in GDM problems defined in
multigranular linguistic contexts, i.e., by using linguistic term sets
with different cardinality to represent experts’ preferences. To do
that, we propose an adaptive consensus support system model for
this type of decision-making problem, i.e., a process that adapts
its behavior to the agreement achieved in each round. This adap-
tive model increases the convergence toward the consensus and,
therefore, reduces the number of rounds to reach it.
Index Terms—Consensus, fuzzy preference relation, group deci-
sion making (GDM), linguistic modeling. Fig. 1. Resolution process of a GDM problem.
I. INTRODUCTION
making problems [1], [3], [5], [20], [27], [38], [42], [49], [58],
NE of the reasons why decision-making processes have
O been widely studied in the literature is the increasing com-
plexity of the social–economic environment [12], [19]. Many
[59].
In GDM problems, there are cases where experts have differ-
ent levels of knowledge about the alternatives, and as a conse-
organizations have moved from a single decision maker to a quence, they might use linguistic term sets with different car-
group of experts to accomplish this task successfully. A group dinality to express their preferences. In such cases, we say that
decision-making (GDM) problem may be defined as a deci- the GDM problem is defined in a multigranular fuzzy linguistic
sion problem with several alternatives and experts that try to context [11], [13], [21], [28], [32], [35], [50], [52], [57].
achieve a common solution taking into account their opinions Usually, GDM problems are solved by carrying out selection
or preferences. processes to obtain a solution set of alternatives from the prefer-
Our interest is focused on GDM problems in which the experts ences given by the experts [19], [23], [25], [53]. However, it may
have to express their preferences on qualitative aspects that happen that, after the selection process, some experts consider
cannot be assessed by means of quantitative values. In these that their preferences have not been taken into account properly
cases, the use of linguistic terms instead of precise numerical to obtain the solution, and hence, they might reject it. One way
values seems to be more appropriate. For example, to evaluate to avoid this situation would be the application of a consensus
the “comfort” of a car, linguistic terms like “good,” “fair,” or process (see Fig. 1) so that the experts discuss and modify their
“poor” could be preferred by the experts instead of numerical preferences in order to reach a sufficient agreement, before ap-
values [48]. plying the selection process [5], [7], [9], [26], [31], [33], [41].
The use of the fuzzy linguistic approach [60]–[62] to assess Selection processes for GDM problems defined in multigranular
qualitative aspects by using linguistic variables, i.e., variables linguistic contexts were introduced in [21] and [28]; hence, here
whose values are not numbers but words or sentences in a natural we focus on the consensus process.
or an artificial language, has proven successful in decision- Consensus modeling is an important area of research in deci-
sion analysis [5], [7]–[9], [14], [16]–[18], [26], [31], [33], [35],
Manuscript received April 27, 2007; revised November 14, 2007, April 20, [37], [39], [40], [46], [47], [54], [55]. Consensus is defined as a
2008, and December 2, 2008; accepted December 15, 2008. First published state of mutual agreement among members of a group where all
January 27, 2009; Current version published April 1, 2009. This work was opinions have been heard and addressed to the satisfaction of
supported in part by the Research Projects TIN2006-02121 and TIN2007-61079.
F. Mata and L. Martı́nez are with the Department of Computer Science, Uni- the group [54]. A consensus reaching process is a dynamic and
versity of Jaén, Jaén 23700, Spain (e-mail: fmata@ujaen.es; martin@ujaen.es). iterative process composed by several rounds where the experts
E. Herrera-Viedma is with the Department of Computer Science and express, discuss, and modify their preferences. Normally, this
Artificial Intelligence, University of Granada, Granada 18071, Spain (e-mail:
viedma@decsai.ugr.es). process is guided by the figure of a moderator, who helps the
Digital Object Identifier 10.1109/TFUZZ.2009.2013457 experts to make their preferences closer to each other [40], [54].
1063-6706/$25.00 © 2009 IEEE
In each consensus round, the moderator evaluates the current S should be small enough so as not to impose useless precision
agreement among the experts’ preferences. If the agreement is levels on the users but large enough to allow a discrimination of
not acceptable, i.e., it is lower than a specified threshold, the the assessments in a limited number of degrees. Additionally,
moderator would then recommend to modify the furthest pref- the following properties are assumed.
erences from the collective ones in order to make them closer. 1) The set S is ordered: si ≥ sj , if i ≥ j.
Otherwise, when the agreement is acceptable, the moderator 2) There is the negation operator: Neg(si ) = sj such that
would apply the selection process in order to obtain the final j = g − i.
solution for the GDM problem. The semantics of S can be given by fuzzy numbers defined on
The main interest in consensus research has been the devel- the unit interval [0,1]. One way to characterize a fuzzy number
opment of new processes with different structures and method- is by using a representation based on parameters of its member-
ologies to achieve its aim [7], [10], [33], [35], [44]. However, ship function [4]. For example, the following semantics can be
the enhancement of these processes has not been the focus in assigned to a set of seven terms via triangular fuzzy numbers:
this research field yet. For instance, it is easy to check that
if the agreement is “very low” (initial rounds), then the num- N = None = (0, 0, 0.17)
ber of changes of preferences should be greater than when the V L = Very Low = (0, 0.17, 0.33)
agreement is “high” (final rounds). Thus, adapting the consen-
sus reaching process to the level of agreement achieved in each L = Low = (0.17, 0.33, 0.5)
discussion round could significantly improve its performance. M = Medium = (0.33, 0.5, 0.67)
The aim of this paper is to propose an adaptive consensus
support system (ACSS) model to support consensus processes H = High = (0.5, 0.67, 0.83)
in GDM problems with multigranular linguistic information, V H = Very High = (0.67, 0.83, 1)
which improves the consensus reaching process by adapting
the search for preferences in disagreement to the current level P = Perfect = (0.83, 1, 1).
of consensus at each round. To do so, three different methods A GDM problem is classically defined as a decision situ-
to identify the preferences that each expert should modify, in ation where a set of experts, E = {e1 , e2 , . . . , em } (m ≥ 2),
order to increase the agreement in the next consensus round, are express their preferences about a set of feasible alternatives,
defined. The result is a model that improves the convergence X = {x1 , x2 , . . . , xn } (n ≥ 2). In many decision situations, it
rate toward the consensus, and therefore decreases the number is assumed that each expert ei provides his or her preferences
of rounds to achieve it. by means of a fuzzy preference relation [19] Pe i = [plk i ], l, k ∈
The rest of the paper is set out as follows. In Section II, pre- {1, . . . , n}, with plk = µ (x , x ) assessed in the unit inter-
i P ei l k
liminaries about the multigranular fuzzy linguistic GDM prob- val [0,1] and being interpreted as the preference degree of the
lems and the consensus process are presented. The proposed alternative xl over xk according to the expert ei . In this paper,
ACSS model is described in detail in Section III. In Section IV, we use linguistic preference relations to represent the experts’
the application of the proposed ACSS model is given, while preferences as in [23] and [24], i.e., with plk i = µP e i (xl , xk )
in Section V, we draw our conclusions. Finally, the Appendix assessed in a linguistic term set S = {s0 , s1 , . . . , sg }.
introduces the meaning and features of the measurements used The ideal situation for GDM problems defined in linguistic
to evaluate the agreement. contexts would be that all the experts use the same linguistic
term set S to express their preferences about the alternatives.
II. PRELIMINARIES However, in some cases, experts may belong, e.g., to distinct
In order for this paper to be as self-contained as possible, research areas, and therefore could have different background
we include in this section a brief review of the fuzzy linguistic and levels of knowledge. A consequence of this is that they need
approach, focusing on GDM problems defined in multigranular to express their preferences by using linguistic term sets with
fuzzy linguistic contexts, and the main elements and features of different granularity Si = {si0 , . . . , sig i }, i ∈ {1, 2, . . . , m}. In
the consensus processes. these cases, the GDM problem is defined in a multigranular
fuzzy linguistic context [11], [13], [21], [28], [32], [35], [50],
A. Multigranular Fuzzy Linguistic GDM Problems [52], [57].
The fuzzy linguistic approach assesses qualitative attributes
B. Consensus Process
by using linguistic assessments by means of linguistic vari-
ables [60]–[62]. This approach has been successfully applied to A consensus reaching process in a GDM problem is an iter-
different problems [2], [3], [6], [20], [22], [29], [30], [34], [36], ative process composed by several discussion rounds, in which
[43], [56], [63]. experts are expected to modify their preferences according to
In this approach, assessments of the preferences on pairs of the advice given by the moderator (see Fig. 1). The moderator
alternatives are provided in the form of linguistic terms or labels plays a key role in this process. Normally, the moderator is a
of a linguistic term set S = {s0 , s1 , . . . , sg }, #(S) = g + 1. An person who does not participate in the discussion but knows the
important issue to analyze is the “granularity of uncertainty,” preferences of each expert and the level of agreement during
i.e., the cardinality of the linguistic term set. The granularity of the consensus process. He/she is in charge of supervising and
MATA et al.: ADAPTIVE CONSENSUS SUPPORT MODEL FOR GDM PROBLEMS IN A MULTIGRANULAR FUZZY LINGUISTIC CONTEXT 281
2) pali is the proximity between the preference values of

expert ei and the collective ones on xl .
Level 3: Preference relation.
1) cr is the global agreement among all experts on all the
pairs of alternatives of a preference relation.
2) pri is the global proximity between the preferences given
by ei and the collective ones of a preference relation.
A further detailed description of the meaning as well as a
description of the computation of these measurements can be
found in the Appendix.
Fig. 2. Phases of the consensus process supervised by the moderator. III. ADAPTIVE CONSENSUS SUPPORT SYSTEM MODEL FOR
GDM IN A MULTIGRANULAR FUZZY LINGUISTIC CONTEXT
driving the consensus process toward success, i.e., to achieve the Several authors [7], [31], [35], [44] have proposed different
maximum possible agreement and reduce the number of experts models to carry out consensus processes where the human mod-
outside of the consensus in each new consensus round. erator’s role is assumed by the own model. In all of them, the
An overall scheme of the different phases carried out in a consensus reaching process is considered as a rigid or inflexi-
consensus process guided by a moderator is shown in Fig. 2. ble one because its behavior remains fixed in all rounds of the
1) Computing level of agreement: The moderator com- consensus process. However, it is obvious that when the level
putes the current agreement among all experts from their of agreement between the experts is “high”, a few number of
preferences. changes of opinions from some of the experts might lead to
2) Checking level of agreement: The moderator compares consensus in a few discussion rounds. On the contrary, when
the current level of agreement with a consensus threshold the level of agreement among the experts is “low,” a high num-
fixed previously. If the consensus threshold is achieved, ber of changes of opinions and many group discussion rounds
the selection process will be applied to obtain the final might be necessary for consensus to be achieved. In this sec-
solution. Otherwise, the consensus process will continue ond case, it seems reasonable that many experts’ preferences
its execution. should be changed if they try to achieve a common solution.
3) Search for preferences: The moderator searches for the As the level of agreement increases, less and less experts might
experts’ preferences furthest from the collective ones and need to change their opinions. In fact, in these cases, it might
suggests how to change them in order to improve the be expected that only those experts whose preference values
agreement in the next round. are furthest from the group ones should change them. In other
In order to evaluate the agreement, it is required to com- words, the number of changes in different stages of a consensus
pute similarity measures among the experts [7], [26], [27], [31], process is clearly related to the actual level of agreement. A con-
[35], [39]. Two types of measurements to guide the consensus sensus model that implements this idea will improve the GDM
reaching process were proposed in [26]: processes.
a) consensus degrees to evaluate the level of agreement In this section, following the aforementioned idea, we present
among all the experts. They will be used to identify the an ACSS model for multigranular fuzzy linguistic GDM prob-
preference values where the agreement is not sufficient; lems that improves the convergence rate toward the consensus,
b) proximity measures to evaluate the distance between the and therefore decreases the number of rounds to achieve it. It
experts’ individual preferences and the group or collective consists of four phases (see Fig. 3).
ones. They will be used to identify the experts who should 1) Making the linguistic information uniform: In this phase,
change their preferences in the next rounds. all experts’ multigranular linguistic preferences are unified
These measurements are computed at the three different levels into a single linguistic domain.
of representation of information of a preference relation: pairs 2) Computing the consensus degree and control of the con-
of alternatives, alternatives, and relation. sensus process: The consensus degree among all experts is
Level 1: Pairs of alternatives. Given a pair of alternatives calculated. If the consensus degree is high enough, the se-
(xl , xk ): lection process is then applied. Otherwise, the consensus
1) cplk is the agreement among all experts on the pair of process keeps going.
alternatives (xl , xk ); 3) Adaptive search for preferences: Different policies or pro-
2) pplk i is the proximity between the preference value of cedures for searching the preferences to be changed in
expert ei , plk
i , and the collective one on the pair of alter- each consensus round are applied based on a broad clas-
natives (xl , xk ). sification of the global consensus level as very low, low,
Level 2: Alternatives. Given the alternative xl ∈ X: and medium. Each preference search procedure (PSp) will
1) cal is the agreement among all experts on xl ; return the set of preferences each expert should change in
Fig. 3. ACSS model in a multigranular fuzzy linguistic context.
order to make his/her preferences closer to the collective

opinion.
4) Production of advice: Once the sets of preferences in dis-
agreement have been identified, an advice system sug-
gests the direction of the changes to be recommended to Fig. 4. Transforming l1 ∈ S into a fuzzy set on S T .
the experts in order to improve the agreement in the next
consensus round. Definition 1 [21]: If S = {l0 , . . . , lp } and ST = {c0 , . . . , cg }
In the following sections, the aforementioned phases are de- are two linguistic term sets, with g ≥ p, then a multigranu-
scribed in detail. lar transformation function τS S T : S → F (ST ) is defined as
follows:

A. Making the Linguistic Information Uniform
τS S T (li ) = {(ch , αh )αh = maxy min{µl i (y), µc h (y)},
To manage multigranular fuzzy linguistic information, we
h = 0, . . . , g}
need to make it uniform, i.e., experts’ preferences have to be
transformed (using a transformation function) into a single do- where F (ST ) is the set of fuzzy sets defined on ST , and µl i (y)
main or linguistic term set that we call the basic linguistic term and µc h (y) are the membership functions of the fuzzy sets as-
set (BLTS), denoted by ST [21]. To do this, it seems reasonable sociated with the linguistic terms li and ch , respectively.
to impose a cardinality high enough to maintain the uncertainty Example 1: Let S = {l0 , l1 , . . . , l4 } and ST =
degrees associated with each one of the possible domains to be {c0 , c1 , . . . , c6 } be two term sets with the following semantics:
unified. This means that the cardinality of the BLTS has to be
as high as possible. Therefore, in a general multigranular fuzzy l0 = (0, 0, 0.25) c0 = (0, 0, 0.16)
linguistic context, to select ST , we proceed as it was proposed l1 = (0, 0.25, 0.5) c1 = (0, 0.16, 0.34)
in [21].
1) If there is only one linguistic term set, from the set of dif- l2 = (0.25, 0.5, 0.75) c2 = (0.16, 0.34, 0.5)
ferent domains to be unified, with maximum cardinality, l3 = (0.5, 0.75, 1) c3 = (0.34, 0.5, 0.66)
then we choose that one as the BLTS, ST .
2) If there are two or more linguistic term sets with maximum l4 = (0.75, 1, 1) c4 = (0.5, 0.66, 0.84)
cardinality, then the selection of ST will depend on the c5 = (0.66, 0.84, 1)
semantics associated with them.
a) If all of them have the same semantics, i.e., the c6 = (0.84, 1, 1).
same fuzzy membership functions associated with
the linguistic terms but with different syntax, then The fuzzy set obtained when applying τS S T to l1 is (see Fig. 4)
any one of them could be selected as ST .
b) If two or more of them have different semantics, then τS S T (l1 ) = {(c0 , 0.39), (c1 , 0.85), (c2 , 0.85),
ST is defined as a generic linguistic term set with a (c3 , 0.39), (c4 , 0), (c5 , 0), (c6 , 0)}.
number of terms greater than the number of terms a
person is able to discriminate, which is normally 7 In order to unify all the experts’ preferences, different multi-
or 9 [51]. granular transformation functions τS i S T are defined. Each lin-
Once ST has been selected, the following multigranular trans- guistic preference value plk i ∈ Si will be transformed in a
i = τS i S T (pi ) = {(ch , αh )|h = 0, . . . , g} on ST .
fuzzy set p̃lk lk lk
formation function is applied to transform every linguistic value
into a fuzzy set defined on ST . To simplify, we will use the membership degrees (α0lk , . . . , αglk )
to denote each fuzzy set p̃lk i

 11 1n 
p̃i = (α0 , . . . , αg11 ) · · · p̃1n
11 1n
i = (α0 , . . . , αg )
 
P̃e i = 

..
.
..
.
..
.
.

p̃ni 1 = (α0n 1 , . . . , αgn 1 ) · · · p̃ni n = (α0n n , . . . , αgn n )
B. Computing the Consensus Degree and Control of the

Consensus Process
Once all the linguistic preferences have been unified by means
of fuzzy sets on the BLTS, the following two steps are applied.
1) Computing the consensus degree: The level of agreement
achieved in the current round is obtained. To do so, a global
consensus degree, called consensus on relation cr ∈ [0, 1], Fig. 5. Adaptive search for preferences.
is computed (see the Appendix).
2) Control of the consensus process: In this phase, both the
in disagreement of those experts furthest from the group. To
global consensus degree cr and the consensus threshold γ
do so, the system establishes three different PSps: “PSp for
are compared such that:
very low consensus,” “PSp for low consensus,” and “PSp for
a) if cr ≥ γ, the level of agreement is sufficient, the
medium consensus.” Each PSp will identify the preferences in
consensus process will stop, and the selection pro-
disagreement in a different way. This fact defines the adaptive
cess will be applied;
character of our model.
b) if cr < γ, a new consensus round is applied.
The adaptive search for preferences consists of two processes
Note that γ is fixed in advance and represents the necessary
(see Fig. 5).
level of agreement for a solution to be accepted by the group.
1) Choose the most suitable PSp: Two parameters θ1 and
A γ value too high may cause that the first condition will never
θ2 , whose values depend on the particular problem dealt
be satisfied, and in consequence, we have a never-ending con-
with, are fixed at the beginning of the consensus process to
sensus process. In order to avoid this situation, we define a
differentiate the three consensus situations: very low, low,
parameter Max rounds that limits the maximum number of con-
and medium consensus. Depending on both parameters,
sensus rounds. This parameter has already been used by other
we choose the most appropriate PSp to apply to each
authors in the control of consensus processes [7], [33].
particular consensus round: a) PSp for very low consensus
The value of γ will obviously depend on the particular prob-
if cr ≤ θ1 ; b) PSp for low consensus if cr ≤ θ2 ; and c)
lem we are dealing with. When the consequences of the decision
PSp for medium consensus otherwise.
to be made are of utmost importance, the minimum level of con-
2) Apply the PSp: Each PSp finds out a set of pref-
sensus required to make that decision should be logically as high
erences, PREFECH i = {(l, k), l, k ∈ {1, 2, . . . , n}, l =
as possible, and it is not unusual if a minimum value of 0.8 or
k}, to be changed by each expert ei in order to improve the
higher is imposed. At the other extreme, we have cases where
agreement in the next round. In each PSp, the agreement
the consequences are not so serious (but are still important),
is analyzed in a different preference representation level.
and it is urgent to obtain a solution to the problem, and thus, a
a) In PSp for very low consensus, the level of pairs of
minimum consensus value as close as possible to 0.5 could be
alternatives is considered.
required.
b) In PSp for low consensus, the level of alternative is
considered.
C. Adaptive Search for Preferences c) In PSp for medium consensus, the level of preference
If the agreement among all experts is low, then there exist a lot relation is considered.
of experts’ preferences in disagreement. In such a case, in order The three PSps are described in detail below:
to bring the preferences closer to each other and so to improve 1) PSp for Very Low Consensus (P SpV L ): Usually, at the
the consensus situation, the number of changes in the experts’ beginning of the consensus process, experts’ preferences are
preferences should be high. However, if the agreement is high, quite far from each other, and therefore, the agreement will be
the majority of preferences is close and only a low number of very low. In these situations, it seems reasonable to require many
experts’ preferences are in disagreement; it seems reasonable to changes in order to make the preferences closer to one another.
change only these particular preferences. We distinguish three To do this, the procedure suggests modifying the preference val-
levels of consensus: very low, low, and medium consensus. Each ues on all the pairs of alternatives where the agreement is not
level implies a different search policy to identify the preferences high enough. These changes may be carried out either by some
with low agreement degree. When the level of consensus is very experts, for example, the experts furthest from the group as pro-
low, all experts will be advised to modify all the preferences posed in [35], or by all experts. We consider the second option
values identified in disagreement, while if the level of consensus more appropriate because it prevents some experts imposing
is greater, the search will be limited to the preference values their preferences in the first rounds, and as a consequence, the
the alternatives and only the preference values in disagreement

of those alternatives where agreement is not sufficient will be
considered.
Another important difference with respect to the P SpV L is
the number of experts involved in the change of preferences.
While in the P SpV L , all experts are required to modify the
identified preference values, in the P SpL , the following re-
striction is added: the experts required to modify the identi-
fied preference values will be those with proximity value at
level of alternatives, for those identified alternatives in dis-
agreement, smaller than an alternative proximity threshold β,
i.e., {ei |pali < β, β ∈ [0, 1], i ∈ {1, . . . , m}}. As in the previ-
ous case, the value of β may be static or dynamic. Again, we
consider the second option more appropriate because it means
that the restriction adapts to the proximity values obtained in
each consensus round. A possible dynamic value in this case
Fig. 6. Chosen preferences by P Sp V L. could be thearithmetic mean of all proximity on alternatives
β = pal = m l
i=1 pai /m.
L
consensus process could be guided toward their own opinions, P Sp finds out the set of preferences to be changed by each
which is known as “tyranny of the majority,” and is a problem ei , PREFECH Li , as follows.
that should be avoided in consensus reaching processes [54]. 1) The consensus degrees at level of alternatives are obtained
Also, with the second option, all experts would be willing to (see the Appendix): {cal |l = 1, . . . , n}.
share the final solution because their preferences were taken 2) Alternatives to be changed X ch are identified. A dynamic
into account to obtain the solution. consensus threshold at level of alternatives is proposed in
P SpV L finds out the set of preferences to be changed by this case, such as the average of the consensus degrees
ei , PREFECH Vi L , as follows. at level of alternative ca = nl=1 cal /n, and then, X ch =
1) First, the pairs of alternatives with a consensus degree {l|cal < ca}.
smaller than a threshold ρ defined at level of pairs of alter- 3) Pairs of alternatives to be changed are identified: P =
natives, P = {(l, k)|cplk < ρ}, are identified. The value {(l, k)|l ∈ X ch ∧ cplk < cp}.
of ρ may be static and fixed before starting the consen- 4) The proximity of the alternatives that should be changed
sus process or dynamic with respect to the level of con- is computed for all experts (see the Appendix): {pali |l ∈
sensus reached in each round. The selection of such a X ch }∀ei ∈ E.
threshold plays a very important role in the identifica- 5) The proximity threshold β = pal used to identify the ex-
tion process because a static value too high may imply perts that will be required to modify the identified pairs of
many changes (leading to all experts having to change al- alternatives is computed.
most all their preference values), while a value too low 6) Then, the sets of preference values that are required to be
may imply very few changes. We consider that a dy- modified are
namic value that changes during the consensus process PREFECH Li = {(l, k) ∈ P | pali < pal }.
is better than static one fixed in advance. In this pa-
per, we have assigned to ρ the average of the consen- Clearly, the new restriction reduces the number of preferences
sus degree at level
of all pairs of alternatives, ρ = cp, and experts required to make changes. Consequently, we have
such that cp = nl=1 ( nk=1,l= k cplk )/(n2 − n). Then,

P = {(l, k)|cplk < cp, l, k = 1, . . . , n}. # PREFECH Li ≤ # PREFECH Vi L .

2) The set of preference values PREFECH Vi L to be changed i i
by each expert ei will be
Graphically, the behavior of this procedure is shown in Fig. 7.
PREFECH Vi L = P. By comparing this figure with the previous one, we can check
that, indeed, the number of changes required is reduced.
In Fig. 6, the characteristics and the behavior of this procedure 3) PSp for Medium Consensus (P SpM ): In the last con-
are graphically described. sensus rounds, the agreement will be close to the desired con-
2) PSp for Low Consensus (P SpL ): After several discussion sensus threshold, θ2 < cr < γ. Therefore, the agreement can
rounds, the agreement among all experts should be greater than be improved by suggesting fewer changes than in the previ-
at the beginning with θ1 < cr ≤ θ2 . In this situation, it seems ous two cases. Consequently, a new restriction is added to the
logical to reduce the number of changes and modify the point P SpM , which will reduce the number of experts required to
of view for the analysis of the agreement. While in the P SpV L modify their opinions: only those experts who have proximity
we focused on all the pairs of alternatives in disagreement, in values on the pairs of alternatives identified in disagreement
the P SpL the agreement is analyzed from the point of view of smaller than a specific proximity threshold at level of pairs of
TABLE I
SUMMARY TABLE OF PSPS
D. Production of Advice
Once the preferences to be changed have been identified, the
model shows the right direction of the changes in order to im-
prove the agreement. For each preference value to be changed,
Fig. 7. Chosen preferences by P Sp L . the model will suggest increasing or decreasing the current as-
sessment.
A guidance advice system based on several direction param-
eters was proposed in [35] to increase the agreement. However,
this system presented some difficulties. In this paper, we present
a new mechanism based on a set of direction rules to identify and
suggest the changes. These rules compare the central values of
the fuzzy sets on ST of the individual and collective preference
assessments cv(p̃lk lk
i ) and cv(p̃c ). The central value represents
the center of gravity of the information contained in the fuzzy
set (see the Appendix). The new direction rules, DR, in our case
are as follows.
i ) − cv(p̃c )) < 0, then the expert ei should in-

DR.1: If (cv(p̃lk lk
crease the assessment associated with the pair of alternatives

(xl , xk ).
i ) − cv(p̃c )) > 0, then the expert ei should de-
crease the assessment associated with the pair of alternatives

Fig. 8. Chosen preferences by P Sp M . (xl , xk ).
i ) − cv(p̃c )) = 0, then the expert ei should not
modify the assessment associated with the pair of alternatives

alternatives will have to change their opinions. This is illustrated (xl , xk ).
in Fig. 8.
The computation of the set of preferences to be changed by
IV. APPLICATION OF THE ACSS MODEL
ei , PREFECH M M
i in P Sp , is as follows.
1) Operations enumerated from 1 to 5 in P SpL are carried In this section, we apply the ACSS model presented in Sec-
out. tion III to a GDM problem with multigranular fuzzy linguistic
2) The proximity threshold to be used in identifying the ex- information.
perts required to modify the identified
lkpairsofmalternatives
in disagreement
is computed: pp = i=1 pplk i /m | A. GDM Framework
(l, k) ∈ P .
3) The sets of preference values that are required to be mod- Let us suppose that a supermarket wants to buy 10 000
ified are bottles of Spanish wine from among four possible brands
of wine or alternatives: {x1 = Marques de Caceres, x2 =
PREFECH M i = {(l, k) ∈ P | pal
i < pal
∧ pp lk
i < pp lk
}. Los Molinos, x3 = Somontano, x4 = Rene Barbier}.
The manager decided to inquire eight experts about their
Clearly,
we have #( i PREFECH M
i ) ≤ opinions E = {e1 , . . . , e8 }. The experts have to reach a high
#( i PREFECH Li ). level of agreement before choosing the best brand of wine. Due
Therefore, this adaptive search of preferences in disagreement to the fact that the experts involved in the problem have different
reduces the number of changes as the consensus increases. levels of knowledge about wine, three linguistic term sets with
The main features of the PSps are shown in the Table I. different cardinalities may be used to provide their preferences.
1) Experts e3 and e7 use label set A TABLE II

FUZZY SETS OBTAINED FOR τ A S T AND τ B S T
a0 = (0, 0, 0.13) a1 = (0, 0.13, 0.25)
a2 = (0.13, 0.25, 0.38) a3 = (0.25, 0.38, 0.5)
a4 = (0.38, 0.5, 0.63) a5 = (0.5, 0.63, 0.75)
a6 = (0.63, 0.75, 0.88) a7 = (0.75, 0.88, 1)
a8 = (0.88, 1, 1).
2) Experts e4 , e5 , and e8 use label set B
TABLE III
b0 = (0, 0, 0.17) b1 = (0, 0.17, 0.33) FUZZY SETS OBTAINED FOR τ C S T
b2 = (0.17, 0.33, 0.5) b3 = (0.33, 0.5, 0.67)

b4 = (0.5, 0.67, 0.83) b5 = (0.67, 0.83, 1)
b6 = (0.83, 1, 1).
3) Experts e1 , e2 , and e6 use label set C
i) Pairs of alternatives
c0 = (0, 0, 0.25) c1 = (0, 0.25, 0.5)  
− 0.6 0.69 0.68
c2 = (0.25, 0.5, 0.75) c3 = (0.5, 0.75, 1)  0.58 − 0.58 0.66 
CM =  .
0.71 0.58 − 0.69
c4 = (0.75, 1, 1). 0.61 0.61 0.63 −
Initially, the experts provide the following linguistic prefer- ii) Alternatives
ence relations: 1
    ca = 0.642, ca2 = 0.6, ca3 = 0.645, ca4 = 0.646 .
− c0 c0 c2 − c2 c0
c4
 c4 − c3 c4   c1 − c1 
c1 iii) Relation: cr = 0.633.
Pe 1 =   Pe 2 =  
c3 c0 − c1 c3 c3 −
c1 b) Control of the consensus process: Because cr =
c2 c1 c3 − c0 c4 −
c3 0.63 < γ = 0.75, the adaptive search of preference values in
    disagreement is activated.
− a1 a4 a3 − b 0 b4 b5
3) Adaptive Search for Preferences:
 a5 − a8 a4   b6 − b 1 b6 
Pe 3 =   Pe 4 =   a) Choose the most suitable PSp: Given that cr =
a4 a1 − a2 b3 b4 − b 2
a5 a5 a7 − b0 b1 b4 − 0.633 ≤ θ1 = 0.65, the level of agreement is very low, and
    therefore, P SpV L is applied.
− b 4 b1 b6 − c2 c3 c1 b) Apply the P SpV L :
 b2 − b 3 b2   c2 − c0 c1  i) Identification of pairs of alternatives in disagree-
Pe 5 =   Pe 6 =  
b4 b3 − b2 c0 c4 − c4 ment
b0 b5 b3 − c4 c4 c0 −
    P = {(1, 2), (2, 1), (2, 3), (3, 2), (4, 1), (4, 2), (4, 3)}.
− a0 a3 a7 − b 6 b1 b3
 a8 − a0 a4   b0 − b 0 b5  ii) Set of preferences to be changed by all experts
Pe 7 =   Pe 8 =  .
a4 a8 − a5 b6 b6 − b 5 PREFECH Vi L = P, i = 1, . . . , 8.
a1 a4 a3 − b4 b1 b0 −
4) Production of Advice:
The parameters applied to the ACSS model are
a) According to rule DR1, the experts are required to in-
1) γ = 0.75;
crease the following preference assessments:
2) θ1 = 0.65 and θ2 = 0.72;
3) Max rounds = 10. 1 = c0 → c1
p12 3 = a4 → a5
p12
5 = b0 → b 1
p41 7 = a0 → a1
p23
B. First Round
1 = c0 → c1
p32 3 = a1 → a2
p32
1) Making the Linguistic Information Uniform: According
to conditions set out in Section III-A, ST = A. To unify the 5 = b3 → b 4
p43 7 = a1 → a2
p41
different linguistic term sets, three multigranular transformation
1 = c1 → c2
p42 4 = b0 → b 1
p12
functions {τA S T , τB S T , τC S T } are used (see Tables II and III).
2) Computing the Consensus Degree and Control of the Con- 6 = c2 → c3
p21 7 = a4 → a5
p42
sensus Process:
2 = c1 → c2
p21 4 = b1 → b 2
p23
a) Computing consensus degree: The consensus degree
obtained at the different levels is as follows (see the Appendix). 6 = c0 → c1
p23 7 = a3 → a4
p43
2 = c1 → c2
p23 4 = b0 → b 1
p41 TABLE IV
PROXIMITY AT LEVEL OF ALTERNATIVES
6 = c0 → c1
p43 8 = b0 → b 1
p21
2 = c0 → c1
p41 4 = b1 → b 2
p42
7 = a0 → a1
p12 8 = b0 → b 1
p23
5 = b2 → b 3
p21 8 = b1 → b 2
p42
5 = b3 → b 4
p32 8 = b0 → b 1 .
p43
b) According to rule DR2, the experts are required to de-

crease the following preference assessments:
Note: In the remaining rounds, we shall only show the most
1 = c4 → c3
p21 2 = c3 → c2
p43 relevant information, the evolution of the consensus degrees and
the operation of PSps.
4 = b4 → b 3
p32 6 = c4 → c3
p41
2) Computing the Consensus Degree and Control of the Con-
1 = c4 → c3
p23 3 = a8 → a7
p23 sensus Process:
a) Consensus degree:
4 = b4 → b 3
p43 6 = c4 → c3
p42
i) Pairs of alternatives
1 = c2 → c1
p41 3 = a5 → a4
p42  
− 0.77 0.69 0.68
5 = b4 → b 3
p12 7 = a8 → a7
p21  0.73 − 0.73 0.66 
CM =  .
0.72 0.71 − 0.69
1 = c3 → c2
p43 3 = a5 → a4
p21
0.77 0.8 0.78 −
5 = b3 → b 2
p23 7 = a8 → a7
p32
ii) Alternatives
2 = c2 → c1
p12 3 = a5 → a4
p41
1
5 = b5 → b 4
p42 8 = b6 → b 5
p12 ca = 0.725, ca2 = 0.73, ca3 = 0.719, ca4 = 0.728 .
2 = c3 → c2
p32 3 = a7 → a6
p43 iii) Relation: cr = 0.726.
6 = c2 → c1
p12 8 = b6 → b 5
p32 b) Control of the consensus process: Because cr =
0.726 < γ = 0.75, the adaptive search of preference values in
2 = c4 → c3
p42 4 = b6 → b 5
p21 disagreement is activated.
6 = c4 → c3
p32 8 = b4 → b 3 .
p41 3) Adaptive Search for Preferences:
a) Choose the most suitable PSp: Given that θ2 = 0.72 <
cr = 0.726 < γ = 0.75, the level of agreement is medium, and
C. Second Round therefore, P SpM is applied.
1) Gathering Information (New Preferences): According to b) Apply the P SpM :
the previous advices, the experts implemented all suggested i) Identifying the alternatives with consensus degree
changes, and the new provided preferences are not high enough
   
− c1 c0 c2 − c1 c0 c4 X ch = {l|cal < 0.726} = {x1 , x3 }.
 c3 − c2 c4   c2 − c2 c1 
Pe 1 =  Pe 2 =  
c3 c1 − c1 c2 c3 − c1 ii) For each one of the aforementioned alternatives, the
c1 c2 c2 − c1 c3 c2 − preference values in disagreement are identified
   
− a2 a4 a3 − b1 b4 b5
P = {(l, k)|l ∈ X ch , cplk < 0.726}
 a4 − a7 a4   b5 − b2 b6 
Pe 3 =  Pe 4 =  
a4 a2 − a2 b3 b3 − b2 = {(1, 3), (1, 4), (3, 1), (3, 2), (3, 4)}.
a4 a4 a6 − b1 b2 b3 −
    iii) Computing the proximity values at level of alterna-
− b3 b1 b6 − c1 c3 c1
tives for these elements in X ch (Table IV).
 b3 − b2 b2   c3 − c1 c1 
Pe 5 =  Pe 6 =   iv) Computing the proximity thresholds used to iden-
b4 b4 − b2 c0 c3 − c4
tify the experts required to modify their preferences
b1 b4 b4 − c3 c3 c1 −
    pa1 = 0.8 pa3 = 0.81.
− a1 a3 a7 − b5 b1 b3
 a7 − a1 a4  b1 − b1 b5 
Pe 7 =  Pe 8 =  .
a4 a7 − a5 b6 b5 − b5 v) Computing the proximity thresholds used to select
a2 a5 a4 − b3 b2 b1 − the experts required to modify the identified pairs
of alternatives in disagreement
pp12 = 0.83 pp13 = 0.78 pp14 = 0.76
pp21 = 0.81 pp23 = 0.82 pp24 = 0.76
pp31 = 0.82 pp32 = 0.79 pp34 = 0.77
pp41 = 0.83 pp42 = 0.86 pp43 = 0.85.
vi) Sets of preferences to be changed
1 = {(3, 1), (3, 2)}

PREFECH M Fig. 9. Experts’ preferences behavior during the consensus reaching process.
PREFECH M
6 = {(1, 3), (1, 4), (3, 1), (3, 4)}
E. Graphic Description of the Experts’ Preferences Behavior
PREFECH M
3 = {(3, 2)}
Fig. 9 illustrates graphically the experts’ preferences in the
8 = {(3, 1), (3, 2), (3, 4)}.
PREFECH M first and third rounds. Each point represents the preference value
4) Production of Advice: given by each expert (in a different color) on a pair of alterna-
a) According to rule DR1, the experts are required to tives. In this figure, the movements of the experts’ preferences
increase the following preference assessments: on the pairs (1, 3) and (3, 2) are highlighted by means of a
box. We note that the preferences come closer each other, form-
1 = c1 → c2
p32 3 = a2 → a3
p32 ing a group (some experts’ preferences cannot be seen because
they are hidden by some other equal assessment), and therefore
6 = c0 → c1
p31 6 = c1 → c2 .
p14
increases the level of agreement.
b) According to rule DR2, the experts are required to
decrease the following preference assessments: V. CONCLUSION
In this paper, we have analyzed the consensus processes in
1 = c3 → c2
p31 6 = c3 → c2
p13
GDM problems under multigranular linguistic information, and
6 = c4 → c3
p34 8 = b6 → b 5
p31 have proposed an ACSS model to guide it and reduce the number
of consensus rounds. To do so, different procedures to find out
8 = b5 → b 4
p32 8 = b5 → b 4 .
p34
the experts’ preference values furthest from the collective ones
have been defined. These procedures are applied according to
D. Third Round
the level of agreement achieved in each consensus round. With
1) Gathering Information (New Preferences): this proposal, we overcome the convergence problems detected
    in other consensus approaches existing in the literature [3], [7],
− c1 c0 c2 − a2 a4 a3
[26], [33], [35], [44].
 c3 − c2 c4   a4 − a7 a4 
Pe 1 =   Pe 3 =  
c 2 c 2 − c1 a4 a3 − a2
APPENDIX
c1 c2 c2 − a4 a4 a6 −
   
This appendix contains the measurements to evaluate the
− c1 c2 c2 − b5 b1 b3
agreement, i.e., consensus degrees and proximity measures.
 c3 − c1 c1   b1 − b 1 b5 
Pe 6 =   Pe 8 =  .
Experts might use linguistic term sets with different car-
c1 c3 − c3 b5 b4 − b4
dinality and semantics. To unify this information, each ex-
c3 c3 c1 − b3 b2 b1 −
pert’s linguistic preference plk i is transformed in a fuzzy set
2) Computing the Consensus Degree and Control of the Con- p̃lk i = (α lk
i0 , . . . , α lk
ig ). Some drawbacks related to the use of tra-
sensus Process: ditional distance measurements were pointed out in [35], and an
a) Consensus degree: alternative similarity function s(·) was proposed to overcome
i) Pairs of alternatives them. The similarity function takes as its arguments the cen-
  tral values of the fuzzy sets to compare. Given a fuzzy set
− 0.77 0.8 0.77
p̃lk lk
i = (αi0 , . . . , αiglk
), its central value defined as [15], [45],
 0.73 − 0.73 0.73  g
CM =   . lk
cv(p̃i ) = h=0 hαih lk
/ gh=0 αih lk
represents the center of grav-
0.82 0.8 − 0.76
ity of the information contained in the fuzzy set.
0.77 0.8 0.78 −
i , p̃j ) ∈
s(p̃lk lk
The similarity between two preference values,
ii) Alternatives [0, 1], is defined as s(p̃i , p̃j ) = 1 − (cv(p̃i ) − cv(p̃j ))/g .
lk lk lk lk
1 The closer s(p̃lk , p̃lk ) is to 1, the more similar p̃lk and p̃lk are,
ca = 0.759, ca2 = 0.746, ca3 = 0.771, ca4 = 0.749 . i j i j
while the closer s(p̃lk lk lk
i , p̃j ) is to 0, the more distant p̃i and p̃j
lk
iii) Relation: cr = 0.756. are.

b) Control of the consensus process: Because cr = Consensus degrees measure the agreement between ex-
0.756 > γ = 0.75, the desired level of consensus is achieved, perts’ preferences. For each pair of experts i and j (i <
and the selection process is applied. j), a similarity matrix SMij = smlk ij is calculated with
smlk lk lk
ij = s(p̃i , p̃j ), l, k = 1, . . . , n ∧ l = k. The consen- [4] P. P. Bonissone and K. S. Decker, “Selecting uncertainty calculi and
granularity: An experiment in trading-off precision and complexity,” in
sus matrix, CM = (cmlk ), is obtained by aggregating all the Uncertainty in Artificial Intelligence. Amsterdam, The Netherlands:
similarity matrices. This aggregation is carried out at level of North Holland/Elsevier, 1986, pp. 217–247.
pairs of alternatives: cmlk = φ(smlk ij ), i, j = 1, . . . , m, l, k =
[5] G. Bordogna, M. Fedrizzi, and G. Pasi, “A linguistic modeling of con-
sensus in group decision making based on OWA operators,” IEEE Trans.
1, . . . , n ∧ i < j. In this paper, we use the arithmetic mean Syst., Man, Cybern. A, Syst. Humans, vol. 27, no. 1, pp. 126–132, Jan.
as aggregation function φ, although different aggregation op- 1997.
erators could be used according to the particular properties to [6] G. Bordogna and G. Pasi, “A fuzzy linguistic approach generalizing
Boolean information retrieval: A model and its evaluation,” J. Amer.
implement. Soc. Inf. Sci., vol. 44, pp. 70–82, 1993.
Consensus degrees are obtained from the consensus matrix [7] N. Bryson, “Group decision-making and the analytic hierarchy process:
CM and in each one of the three different levels of the prefer- Exploring the consensus-relevant information content,” Comput. Oper.
Res., vol. 1, no. 23, pp. 27–35, 1996.
ence relation Pe i . [8] F. J. Cabrerizo, S. Alonso, and E. Herrera-Viedma, “A consensus model
for group decision making problems with unbalanced fuzzy linguistic
Level 1 (Consensus on pairs of alternatives): The consensus information,” Int. J. Inf. Technol. Decis. Making, to be published.
degree on a pair of alternatives (xl , xk ), called cplk , measures [9] C. Carlsson, D. Ehrenberg, P. Eklund, M. Fedrizzi, P. Gustafsson, P. Lind-
the agreement among all experts on that pair of alternatives: holm, G. Merkuryeva, T. Riissanen, and A. G. S. Ventre, “Consensus in
distributed soft environments,” Eur. J. Oper. Res., vol. 61, pp. 165–185,
cplk = cmlk ∀l, k = 1, . . . , n ∧ l = k. 1992.
Level 2 (Consensus on alternatives): The consensus de- [10] W. L. Cats-Baril and G. P. Huber, “Decision support systems for ill-
gree on an alternative xl , called cal , measures the structure problems: An empirical study,” Decis. Sci., vol. 18, pp. 350–372,
1987.
agreement
n among all experts on that alternative: cal = [11] S. L. Chang, R. C. Wang, and S. Y. Wang, “Applying a direct multi-
( k =1,l= k (cplk + cpk l ))/2(n − 1). granularity linguistic and strategy-oriented aggregation approach on the
Level 3 (Consensus on the relation): The consensus degree on assessment of supply performance,” Eur. J. Oper. Res., vol. 177, no. 2,
pp. 1013–1025, 2007.
the relation, called cr, measuresthe global agreement among [12] S. J. Chen and C. L. Hwang, Fuzzy Multiple Attribute Decision-Making
the experts’ preferences: cr = nl=1 cal /n. Methods and Applications. New York: Springer-Verlag, 1992.
[13] Z. Chen and D. Ben-Arieh, “On the fusion of multi-granularity linguistic
Proximity measurements evaluate the proximity between label sets in group decision making,” Comput. Ind. Eng., vol. 51, no. 3,
the individual experts’ preferences and the collective ones. pp. 526–541, 2006.
[14] A. K. Choudhury, R. Shankar, and M. K. Tiwari, “Consensus-based intel-
The collective preference P̃e c = (p̃lk c ) is calculated by aggre- ligent group decision-making model for the selection of advanced tech-
gating the set of (uniformed) individual preference relations nology,” Decis. Support Syst., vol. 42, no. 3, pp. 1776–1799, 2006.
{P̃e 1 , . . . , P̃e m }: p̃lk lk lk
c = ψ(p̃1 , . . . , p̃m ), with ψ, an “aggrega-
[15] M. Delgado, F. Herrera, E. Herrera-Viedma, and L. Martı́nez, “Combining
numerical and linguistic information in group decision making,” Inf. Sci.,
tion operator.” Then, for each expert ei , we calculate a proxim- vol. 107, pp. 177–194, 1998.
ity matrix P Mi = (pmlk lk lk lk
i ), pmi = s(p̃i , p̃c ), which measures [16] P. Eklund, A. Rusinowska, and H. de Swart, “Consensus reaching in
the distance between P̃e i and P̃e c . committees,” Eur. J. Oper. Res., vol. 178, pp. 185–193, 2007.
[17] E. Ephrati and J. S. Rosenschein, “Deriving consensus in multiagent sys-
The proximity measurements are computed in each one of tems,” Artif. Intell., vol. 87, no. 1, pp. 21–74, 1996.
the three levels of the relation, too. [18] M. Fedrizzi, M. Fedrizzi, and R. A. Marques Pereira, “Soft consensus and
network dynamics in group decision making,” Int. J. Intell. Syst., vol. 14,
Level 1 (Proximity on pairs of alternatives): Given an expert ei , no. 1, pp. 63–77, 1999.
his/her proximity measure on a pair of alternatives (xl , xk ), [19] J. Fodor and M. Roubens, Fuzzy Preference Modelling and Multicriteria
Decision Support. Norwell, MA: Kluwer, 1994.
called pplk
i , measures the proximity between his/her pref- [20] J. L. Garcia-Lapresta, “A general class of simple majority decision rules
erence and the collective one on that pair of alternatives: based on linguistic opinions,” Inf. Sci., vol. 176, no. 4, pp. 352–365, 2006.
i = pmi ∀l, k = 1, . . . , n ∧ l = k.
pplk lk [21] F. Herrera, E. Herrera-Viedma, and L. Martı́nez, “A fusion approach
for managing multi-granularity linguistic term sets in decision making,”
Level 2 (Proximity on alternatives): Given an expert ei , Fuzzy Sets Syst., vol. 114, pp. 43–58, 2000.
his/her proximity measure on an alternative xl , called [22] F. Herrera, E. Herrera-Viedma, and L. Martı́nez, “A fuzzy linguistic
pal , measures the proximity between his/her prefer- methodology to deal with unbalanced linguistic term sets,” IEEE Trans.
Fuzzy Syst., vol. 16, no. 2, pp. 354–370, Apr. 2008.
ence
n and the lkcollective one on that alternative: pali = [23] F. Herrera, E. Herrera-Viedma, and J. L. Verdegay, “A sequential selection
( k =1,k = l (ppi + ppi ))/2(n − 1).
kl
process in group decision making with linguistic assessment,” Inf. Sci.,
Level 3 (Proximity on the relation): Given an expert ei , his/her vol. 85, pp. 223–239, 1995.
[24] F. Herrera, E. Herrera-Viedma, and J. L. Verdegay, “Direct approach
proximity measure on the relation, called pri , measures the processes in group decision making using linguistic OWA operators,”
global proximity between his/her individual preferences and Fuzzy Sets Syst., vol. 79, pp. 175–190, 1996.
the collective one: pri = ( nl=1 pali )/n. [25] F. Herrera, E. Herrera-Viedma, and J. L. Verdegay, “A linguistic decision
process in group decision making,” Group Decis. Negotiation, vol. 5,
pp. 165–176, 1996.
REFERENCES [26] F. Herrera, E. Herrera-Viedma, and J. L. Verdegay, “A model of consensus
in group decision making under linguistic assessments,” Fuzzy Sets Syst.,
[1] B. Arfi, “Fuzzy decision making in politics: A linguistic fuzzy-set ap- vol. 79, pp. 73–87, 1996.
proach,” Political Anal., vol. 13, no. 1, pp. 23–56, 2005. [27] F. Herrera, E. Herrera-Viedma, and J. L. Verdegay, “Linguistic measures
[2] B. Arfi, “Linguistic fuzzy-logic game theory,” J. Conflict Resolution, based on fuzzy coincidence for reaching consensus in group decision
vol. 50, no. 1, pp. 28–57, 2006. making,” Int. J. Approximate Reason., vol. 16, pp. 309–334, 1997.
[3] D. Ben-Arieh and C. Zhifeng, “Linguistic labels aggregation and con- [28] F. Herrera and L. Martı́nez, “A model based on linguistic 2-tuples for
sensus measure for autocratic decision-making using group recommenda- dealing with multigranularity hierarchical linguistic contexts in multi-
tions,” IEEE Trans. Syst., Man, Cybern. A, Syst. Humans, vol. 36, no. 3, expert decision-making,” IEEE Trans. Syst., Man, Cybern. B, Cybern.,
pp. 558–568, May 2006. vol. 31, no. 2, pp. 227–234, Apr. 2001.
[29] E. Herrera-Viedma, “An information retrieval model with ordinal linguis- [53] M. Roubens, “Fuzzy sets and decision analysis,” Fuzzy Sets Syst., vol. 90,
tic weighted queries based on two weighting elements,” Int. J. Uncer- pp. 199–206, 1997.
tainty, Fuzziness Knowl.-Based Syst., vol. 9, pp. 77–88, 2001. [54] S. Saint and J. R. Lawson, Rules for Reaching Consensus: A Modern
[30] E. Herrera-Viedma, “Modeling the retrieval process of an information Approach to Decision Making. San Francisco, CA: Jossey-Bass, 1994.
retrieval system using an ordinal fuzzy linguistic approach,” J. Amer. Soc. [55] E. Szmidt and J. Kacprzyk, “A consensus reaching process under intu-
Inf. Sci., vol. 52, no. 6, pp. 460–475, 2001. itionistic fuzzy preference relations,” Int. J. Intell. Syst., vol. 18, no. 7,
[31] E. Herrera-Viedma, S. Alonso, F. Chiclana, and F. Herrera, “A consen- pp. 837–852, 2003.
sus model for group decision making with incomplete fuzzy preference [56] J. H. Wang and J. Hao, “A new version of 2-tuple fuzzy linguistic rep-
relations,” IEEE Trans. Fuzzy Syst., vol. 15, no. 5, pp. 863–877, Oct. resentation model for computing with words,” IEEE Trans. Fuzzy Syst.,
2007. vol. 14, no. 3, pp. 435–445, Jun. 2006.
[32] E. Herrera-Viedma, O. Cordón, M. Luque, A. G. López, and A. N. Muñoz, [57] S. Y. Wang, “Applying 2-tuple multigranularity linguistic variables to
“A model of fuzzy linguistic IRS based on multi-granular linguistic in- determine the supply performance in dynamic environment based on
formation,” Int. J. Approximate Reason., vol. 34, no. 3, pp. 221–239, product-oriented strategy,” IEEE Trans. Fuzzy Syst., vol. 16, no. 1, pp. 29–
2003. 39, Feb. 2008.
[33] E. Herrera-Viedma, F. Herrera, and F. Chiclana, “A consensus model for [58] Z. S. Xu, “Deviation measures of linguistic preference relations in group
multiperson decision making with different preference structures,” IEEE decision making,” Omega, vol. 33, no. 3, pp. 249–254, 2005.
Trans. Syst., Man, Cybern. A, Syst. Humans, vol. 32, no. 3, pp. 394–402, [59] R. R. Yager, “An approach to ordinal decision making,” Int. J. Approxi-
May 2002. mate Reason., vol. 12, pp. 237–261, 1995.
[34] E. Herrera-Viedma and A. G. López-Herrera, “A model of information [60] L. A. Zadeh, “The concept of a linguistic variable and its applications to
retrieval system with unbalanced fuzzy linguistic information,” Int. J. approximate reasoning—Part I,” Inf. Sci., vol. 8, pp. 199–249, 1975.
Intell. Syst., vol. 22, no. 11, pp. 1197–1214, 2007. [61] L. A. Zadeh, “The concept of a linguistic variable and its applications to
[35] E. Herrera-Viedma, L. Martı́nez, F. Mata, and F. Chiclana, “A consensus approximate reasoning—Part II,” Inf. Sci., vol. 8, pp. 301–357, 1975.
support system model for group decision-making problems with multi- [62] L. A. Zadeh, “The concept of a linguistic variable and its applications to
granular linguistic preference relations,” IEEE Trans. Fuzzy Syst., vol. 13, approximate reasoning—Part III,” Inf. Sci., vol. 9, pp. 43–80, 1975.
no. 5, pp. 644–658, Oct. 2005. [63] S. Zadrozny and J. Kacprzyk, “Computing with words for text processing:
[36] E. Herrera-Viedma, G. Pasi, A. G. López-Herrera, and C. Porcel, “Evaluat- An approach to the text categorization,” Inf. Sci., vol. 176, no. 4, pp. 415–
ing the information quality of web sites: A methodology based on fuzzy,” 437, 2006.
J. Amer. Soc. Inf. Sci. Technol., vol. 57, no. 4, pp. 538–549, 2006.
[37] C.-L. Huang, C.-C. Lo, K.-M. Chao, and M. Younas, “Reaching consensus:
A moderated fuzzy Web services discovery method,” Inf. Softw. Technol., Francisco Mata was born in 1971. He received the
vol. 48, no. 6, pp. 410–423, 2006. M.Sc. degree in computer sciences from the Univer-
[38] V. N. Huynh and Y. Nakamori, “A satisfactory-oriented approach to mul- sity of Granada, Granada, Spain, in 1994, and the
tiexpert decision-making with linguistic assessments,” IEEE Trans. Syst., Ph.D. degree in computer sciences from the Univer-
Man, Cybern. B, Cybern., vol. 35, no. 2, pp. 184–196, Apr. 2005. sity of Jaén, Jaén, Spain, in 2006.
[39] J. Kacprzyk and M. Fedrizzi, “A “soft” measure of consensus in the setting He is currently a Lecturer in the Department of
of partial (fuzzy) preferences,” Eur. J. Oper. Res., vol. 34, pp. 316–325, Computer Science, University of Jaén. His current re-
1988. search interests include linguistic preference model-
[40] J. Kacprzyk, M. Fedrizzi, and H. Nurmi, “Group decision making and ing, consensus reaching process, group decision mak-
consensus under fuzzy preferences and fuzzy majority,” Fuzzy Sets Syst., ing, and multiagent systems.
vol. 49, pp. 21–31, 1992.
[41] J. Kacprzyk, M. Fedrizzi, and H. Nurmi, Consensus Under Fuzziness.
Norwell, MA: Kluwer, 1997.
[42] J. Kacprzyk and S. Zadrozny, “Computing with words in decision Luis Martı́nez was born in 1970. He received the
making: Through individual and collective linguistic choice rules,” M.Sc. and Ph.D. degrees in computer sciences, from
Int. J. Uncertainty, Fuzziness Knowl.-Based Syst., vol. 9, pp. 89–102, the University of Granada, Granada, Spain, in 1993
2001. and 1999, respectively.
[43] J. Kacprzyk and S. Zadrozny, “Linguistic database summaries and their He is currently a Senior Lecturer in the Depart-
protoforms: Towards natural language based knowledge discovery tools,” ment of Computer Science, University of Jaén, Jaén,
Inf. Sci., vol. 173, no. 4, pp. 281–304, 2005. Spain. His current research interests include linguis-
[44] N. I. Karacapilidis and C. P. Pappis, “A framework for group decision tic preference modeling, decision making, fuzzy-
making support systems: Combining AI tools and OR techniques,” Eur. logic-based systems, computer-aided learning, sen-
J. Oper. Res., vol. 103, pp. 373–388, 1997. sory evaluation, recommender systems, and elec-
[45] G. J. Klir and B. Yuan, Fuzzy Sets and Fuzzy Logic: Theory and Applica- tronic commerce. He was a coeditor of several journal
tions. Englewood Cliffs, NJ: Prentice-Hall, 1995. special issues on fuzzy preference modeling and fuzzy sets theory.
[46] L. I. Kuncheva and R. Krishnapuram, “A fuzzy consensus aggregation
operator,” Fuzzy Sets Syst., vol. 79, pp. 347–356, 1996.
[47] H. S. Lee, “Optimal consensus of fuzzy opinions under group decision
making environment,” Fuzzy Sets Syst., vol. 132, no. 3, pp. 303–315, Enrique Herrera-Viedma was born in 1969. He
2002. received the M.Sc. and Ph.D. degrees in computer
[48] E. Levrat, A. Voisin, S. Bombardier, and J. Bremont, “Subjective evalu- sciences from the University of Granada, Granada,
ation of car seat comfort with fuzzy set techniques,” Int. J. Intell. Syst., Spain, in 1993 and 1996, respectively.
vol. 12, pp. 891–913, 1997. He is currently a Professor in the Department of
[49] J. Ma, D. Ruan, Y. Xu, and G. Zhang, “A fuzzy-set approach to treat Computer Science and Artificial Intelligence, Uni-
determinacy and consistency of linguistic terms in multi-criteria deci- versity of Granada. He has authored or coauthored
sion making,” Int. J. Approximate Reason, vol. 44, no. 2, pp. 165–181, more than 50 papers published in international jour-
2007. nals. He has coedited one international book and 11
[50] L. Martı́nez, J. Liu, J. B. Yang, and F. Herrera, “A multi-granular hier- special issues in international journals on topics such
archical linguistic model for design evaluation based on safety and cost as computing with words and preference modeling,
analysis,” Int. J. Intell. Syst., vol. 20, no. 12, pp. 1161–1194, 2005. soft computing in information retrieval, and aggregation operators. He is a mem-
[51] G. A. Miller, “The magical number seven plus or minus two: Some limits ber of the Editorial Board of the Fuzzy Sets and Systems, Soft Computing, and
on our capacity of processing information,” Psychol. Rev., vol. 63, pp. 81– International Journal of Information Technology and Decision Making. His cur-
97, 1956. rent research interests include group decision making, decision support systems,
[52] C. Porcel, A. G. López-Herrera, and E. Herrera-Viedma, “A recommender consensus models, linguistic modeling, aggregation of information, information
system for research resources based on fuzzy linguistic modeling,” Expert retrieval, genetic algorithms, digital libraries, Web quality evaluation, and rec-
Syst. Appl., vol. 36, pp. 5173–5183, 2009. ommender systems.
H∞ Fuzzy Filtering of Nonlinear Systems

With Intermittent Measurements
Huijun Gao, Member, IEEE, Yan Zhao, James Lam, Senior Member, IEEE, and Ke Chen
Abstract—This paper is concerned with the problem of H∞ of linear matrix inequalities (LMIs), which is recognized to be
fuzzy filtering of nonlinear systems with intermittent measure- conservative, and for some highly nonlinear complex systems,
ments. The nonlinear plant is represented by a Takagi–Sugeno the common Lyapunov matrix even does not exist [45]. This
(T–S) fuzzy model. The measurements transmission from the plant
to the filter is assumed to be imperfect, and a stochastic variable has motivated the development of the more recent approach,
satisfying the Bernoulli random binary distribution is utilized to which employs basis-dependent Lyapunov functions. Results in
model the phenomenon of the missing measurements. Attention many papers have shown that this approach is less conserva-
is focused on the design of an H∞ filter such that the filter er- tive because the basis-dependent Lyapunov function is also a
ror system is stochastically stable and preserves a guaranteed H∞ “blending” of some piecewise Lyapunov functions [15], [44].
performance. A basis-dependent Lyapunov function approach is
developed to design the H∞ filter. By introducing some slack ma- Since the state variables in control systems are not always
trix variables, the coupling between the Lyapunov matrix and the available, state estimation is another important problem that
system matrices is eliminated, which greatly facilitates the filter- has been attracting attention from researchers around the world,
design procedure. The developed theoretical results are in the form and a great number of important results have been reported.
of linear matrix inequalities (LMIs). Finally, an illustrative exam- To mention a few, the filtering problem has been solved for
ple is provided to show the effectiveness of the proposed approach.
linear systems for uncertain systems [35], Markovian jumping
Index Terms—Basis-dependent Lyapunov function, H∞ filter systems [3], [17], [25], [38], sample-data systems [21], [23],
design, intermittent measurements, nonlinear systems, Takagi– [33], systems with singular perturbation [18], [20], and systems
Sugeno (T–S) fuzzy systems.
with time delays [11], [38]. Different norms have been used to
measure the filtering performance (see, for instance, the H∞
I. INTRODUCTION norm [12], [16], the L1 norm [1], and the L2 –L∞ norm [19]).
N RECENT years, there has been a growing interest in the There are also some results investigating the filtering problems
I Takagi–Sugeno (T–S) fuzzy model since it is a powerful so-
lution that bridges the gap between linear control and complex
for nonlinear systems [10], [20], [24].
Among the aforementioned references, H∞ filtering is one of
nonlinear systems [4], [36], [37]. The important advantage of the the most important strategies [14], [38], [41]. The advantage of
T–S fuzzy model is its universal approximation of any smooth H∞ filtering lies in that no statistical assumption on the noise
nonlinear function by a “blending” of some local linear system signals is needed, and thus, it is more general than classical
models. Based on that local linearity, many complex nonlinear Kalman filtering [12]. Due to the powerful approximation prop-
problems can be simplified by employing the Lyapunov function erty of T–S fuzzy model, recently, there have been a number of
approach [9]. The earlier approach employs quadratic Lyapunov results on H∞ filtering for T–S fuzzy systems [8], [39], [45].
functions, which has shown great effectiveness and has been A robust H∞ filter design for continuous T–S fuzzy models
widely used up until now [5]–[7], [27], [30]. This approach at- based on the notion of quadratic stability proposed in [8], [13],
tempts to find a common positive definite matrix to satisfy a set and [45] are concerned with the H∞ filtering problem for a class
of discrete-time fuzzy systems using basis-dependent Lyapunov
Manuscript received May 8, 2007; revised July 16, 2007. First published functions with reduced conservatism. It is worth noting that all
April 30, 2008; current version published April 1, 2009. This work was sup- these results are based on the implicit assumption that the com-
ported in part by the National Natural Science Foundation of China (60825303,
60834003), in part by the 973 Project (2009CB320600), in part by the Research munication between the physical plant and filter is perfect, that
Fund for the Doctoral Programme of Higher Education of China (20070213084), is, the signals transmitted from the plant will arrive at the filter
in part by the Heilongjiang Outstanding Youth Science Fund (JC200809), in part simultaneously and perfectly.
by the Postdoctoral Science Foundation of China (200801282), in part by the
Fok Ying Tung Education Foundation (111064), and in part by HKU CRCG On another research front, networked control systems have
200707176077. drawn much attention due to their great advantages over tra-
H. Gao is with the Space Control and Inertial Technology Research ditional systems such as low cost, reduced weight and power
Center, Harbin Institute of Technology, Harbin 150001, China (e-mail:
hjgao@hit.edu.cn). requirements, simple installation and maintenance, and high re-
Y. Zhao is with the Space Control and Inertial Technology Research Center, liability. However, the utilization of networks as communication
Harbin Institute of Technology, Harbin, Heilongjiang 150001, China (e-mail: channels brings us new challenges, and the analysis and syn-
zhaoresponsible@gmail.com).
J. Lam is with the Department of Mechanical Engineering, University of thesis problems become more difficult and complicated due to
Hong Kong, Hong Kong (e-mail: james.lam@hku.hk). their limited transmission capacity. Among a few other impor-
K. Chen is with the Department of Electrical and Computer Engineering, tant problems, data packet dropout is an important issue to be
University of Alberta, Edmonton, AB T6G 2R3, Canada (e-mail: kchen1@
ece.ualberta.ca). addressed. So far, there have been a number of results focusing
Digital Object Identifier 10.1109/TFUZZ.2008.924206 on stability analysis of networked systems [29], [43]. Recently,
1063-6706/$25.00 © 2009 IEEE
increasing attention has been paid to the synthesis problems.

For example, state-feedback control is investigated in [40], and
H∞ control is developed in [22] and [42]. It is noted that most
of these results focus on the control-related problems. More
recently, there have been a few results on the filtering problem
for networked systems: References [31] and [32] consider the
filtering problem for stochastic systems with missing measure-
ments, and [26] investigates the problem of performing Kalman
filtering with intermittent observations, while [11] and [34] dis-
cuss that for stochastic systems with time delays. To the best
Fig. 1. Filtering problem with intermittent measurements.
of the author’s knowledge, up until now, there has been no re-
search on the filter design for T–S fuzzy systems in the presence
A. Physical Plant
of intermittent measurements, which still remains important and
challenging. This motivates the present study. The plant under consideration is a nonlinear discrete-time
In this paper, we investigate the problem of H∞ filter design system that is represented by the T–S fuzzy model as follows:
for nonlinear systems with intermittent measurements. The non- 1) Plant Rule i: IF θ1 (k) is Mi1 and θ2 (k) is Mi2 and · · ·
linear plant is represented by a T–S fuzzy model. The measure- and θp (k) is Mip , THEN
ments transmitted between the plant and the filter are assumed to xk +1 = Ai xk + Bi wk
be imperfect, and the phenomenon of the missing measurements
is assumed to satisfy the Bernoulli random binary distribution. yk = Ci xk + Di wk
Given a T–S fuzzy system, our objective is to design an H∞ zk = Li xk
filter such that the filter error system is stochastically stable and
preserves a guaranteed H∞ performance. A basis-dependent i = 1, . . . , r. (1)
Lyapunov function approach is developed to design a desired Here, Mij are the fuzzy sets, xk ∈ R is the state vector, n
H∞ filter. The introduction of some slack matrix variables elim- wk ∈ Rp is the noise signal that is assumed to be arbitrary sig-
inates the coupling between the system matrices and Lyapunov nal in l2 [0, ∞), zk ∈ Rq is the signal to be estimated, yk ∈ Rm
matrix, which simplifies the filter design. The theoretical results is the measurement output, Ai , Bi , Ci , Di , and Li are known
are in the form of LMIs, which can be solved by standard nu- constant matrices with appropriate dimensions, r is the num-
merical software. An example shows the effectiveness of the ber of IF–THEN rules, and θ(k) = [ θ1 (k), θ2 (k), . . . , θp (k) ]
proposed approach. is the premise variable vector and measurable. The fuzzy basis
The rest of this paper is organized as follows. Section II functions are given by
formulates the problem under consideration. The stability con- p
j =1 Mij (θj (k))
dition and H∞ performance of the filter error system are given hi (θ(k)) = r p
in Section III. The filter design problem is solved in Section IV. i=1 j =1 Mij (θj (k))
An illustrative example is given in Section V, and we conclude with Mij (θj (k)) representing the grade of membership of θj (k)
the paper in Section VI. in Mij . In what follows, we will drop the argument of hi (θk )
The notation used in the paper is standard. The super- for brevity. Therefore, for all k, we have
script “T ” stands for matrix transposition; Rn denotes the n-
hi ≥ 0, i = 1, 2, . . . , r
dimensional Euclidean space, and the notation P > 0 (≥ 0)
means that P is real symmetric and positive definite (semidefi-
r
hi = 1. (2)
nite). l2 [0, ∞) is the space of square-integrable vector functions
i=1
over [0, ∞); the notation | · | refers to the Euclidean vector
norm, and · 2 stands for the usual l2 [0, ∞) norm. In sym- Let ρ be a set of basis functions satisfying (2). A more compact
metric block matrices or complex matrix expressions, we use presentation of the T–S discrete-time fuzzy model is given by
an asterisk (∗) to represent a term that is induced by symmetry, xk +1 = A(h)xk + B(h)wk
and diag{. . .} stands for a block-diagonal matrix. In addition,
yk = C(h)xk + D(h)wk
E{x} and E{ x| y} will, respectively, mean expectation of x
and expectation of x conditional on y. Matrices, if their dimen- zk = L(h)xk (3)
sions are not explicitly stated, are assumed to be compatible for
where
algebraic operations.

r
r
r
II. PROBLEM FORMULATION A(h) = hi Ai , B(h) = hi Bi , C(h) = hi Ci

i=1 i=1 i=1
The filtering problem with intermittent measurements is

r
r
shown in Fig. 1, where the physical plant is represented by D(h) = hi Di , L(h) = hi Li (4)
a T–S fuzzy model, and the data missing phenomenon occurs i=1 i=1
intermittently from the plant to the filter. In the following, we

model the whole problem mathematically. and h = (h1 , h2 , . . . , hr ) ∈ ρ.
GAO et al.: H ∞ FUZZY FILTERING OF NONLINEAR SYSTEMS WITH INTERMITTENT MEASUREMENTS 293
B. Filter Before proceeding further, we first introduce the following

In this paper, we consider the following fuzzy filter to estimate definition.
zk . Definition 1: The filter error system in (8) is said to be stochas-
1) Filter Rule i: IF θ1 (k) is Mi1 and θ2 (k) is Mi2 and · · · tically stable in the mean square when w(k) ≡ 0 for any initial
and θp (k) is Mip , THEN condition x0 if there exists a finite W > 0 such that
∞

x̂k +1 = Af i x̂k + Bf i yf k 2
E |xk | x0 < xT0 W x0 .
ẑk = Lf i x̂k
k =0
i = 1, . . . , r. (5) Then, the problem to be addressed in this paper is expressed

Here, x̂k ∈ R , and ẑk ∈ R , and Af i , Bf i , and Lf i are to be
n q as follows.
determined. Thus, the filter can be represented by the following Problem H∞ filtering with intermittent measurements
input–output form: (HFIM): Consider the filtering problem shown in Fig. 1, and
suppose the intermittent transmission parameter ē is known.
x̂k +1 = Af (h)x̂k + Bf (h)yf k Given a scalar γ > 0, design a fuzzy filter in the form of (7)
ẑk = Lf (h)x̂k . (6) such that
1) (stochastic stability) the filter error system in (8) is
C. Communication Link stochastically stable in the sense of Definition 1;
2) (H∞ performance) under zero initial condition, the error
It is assumed that measurements are intermittent, that is, the output z̄k satisfies
data may be lost during their transmission. In this case, the input
yf k of the filter is no longer equivalent to the output yk of the ||z̄||E ≤ γ||w||2 (11)
plant (that is, yk = yf k ). In this paper, the data loss phenomenon
is modeled via a stochastic approach: where
 
yf k = e(k)yk  ∞


where {e(k)} is Bernoulli process. {e(k)} models the intermit- ||z̄||E = E z̄kT z̄k .
 
k =0
tent nature of the link from the plant to the filter. More specif-
ically, e(k) = 0 when the link fails (that is, data is lost), and If the previous two conditions are satisfied, the filter error
e(k) = 1 means successful transmission. A natural assumption system is called stochastically stable with a guaranteed H∞
on {e(k)} can be made as performance γ.
Prob{e(k) = 1} = E{e(k)} = ē, Prob{e(k) = 0} = 1 − ē.
Based on this, we have III. FILTERING PERFORMANCE ANALYSIS
x̂k +1 = Af (h)x̂k + Bf (h)e(k)yk In this section, the filtering analysis problem is concerned.
More specifically, we assume that the filter matrices in (6) are
ẑk = Lf (h)x̂k . (7) known, and we will study the condition under which the filter
error system in (8) is stochastically stable in the mean square
D. Filter Error System with a given H∞ performance γ. The following theorem shows
From (3) and (7), the filter error system is given by that the H∞ performance of the filter error system can be guar-
anteed if there exist some fuzzy-basis-dependent matrices and
x̄k +1 = A1 (h)x̄k + ẽ(k)A2 (h)x̄k + B1 (h)wk + ẽ(k)B2 (h)wk additional matrices satisfying a certain linear matrix inequality
z̄k = L̄(h)x̄k (8) (LMI).
Theorem 1: Consider the fuzzy system in (3), and suppose
where that the filter in (6) is given. The filter error system in (8)

xk is stochastically stable with a given H∞ performance γ, if
x̄k = , z̄k = zk − ẑk (9)
x̂k there exist fuzzy-basis-dependent matrices P (h) > 0, Ω(h),

for any h ∈ ρ, h+ = (h1 (θk +1 ), h2 (θk +1 ), . . . , hr (θk +1 )) ∈ ρ,
A(h) 0 B(h) satisfying
A1 (h) = , B1 (h) =
ēBf (h)C(h) Af (h) ēBf (h)D(h)  
Θ 0 0 ΩT (h+ )A1 (h) ΩT (h+ )B1 (h)

0 0 0  ∗ Θ 0 f ΩT (h+ )A (h) f ΩT (h+ )B (h) 
A2 (h) = , B2 (h) =  2 2 
 
Bf (h)C(h) 0 Bf (h)D(h)  ∗ ∗ −I L̄(h) 0 <0
 
L(h) = [ L(h) −Lf (h) ] (10)  
∗ ∗ ∗ −P (h) 0 
and ẽ(k) = e(k) − ē. It is clear that E{ẽ(k)} = 0 and that ∗ ∗ ∗ ∗ −γ 2 I
E{ẽ(k)ẽ(k)} = ē(1 − ē). (12)
where Here, (15) implies

Θ = P (h+ ) − Ω(h+ ) − ΩT (h+ ) AT1 (h)P (h+ )A1 (h) + f 2 AT2 (h)P (h+ )A2 (h) − P (h) < 0

f = ē(1 − ē). and thus, we have
Proof: We first prove the stochastic stability of the filter er-
ror system in (8). To this end, assume wk ≡ 0, and choose a ∆Vk < 0.
Lyapunov function as Define
Vk = x̄Tk [P (h)] x̄k . (13)
Φ = AT1 (h)P (h+ )A1 (h) + f 2 AT2 (h)P (h+ )A2 (h) − P (h)
When wk ≡ 0, (8) becomes
and we get
x̄k +1 = A1 (h)x̄k + ẽ(k)A2 (h)x̄k
E { Vk +1 | x̄k } − Vk ≤ −λm in (−Φ)x̄Tk x̄k .
z̄k = L̄(h)x̄k .
Taking mathematical expectation of both sides, for any T ≥ 1,
Then, we have and summing up the inequality on both sides from k = 0, . . . , T ,
∆Vk = E { Vk +1 | x̄k } − Vk we have

= E x̄Tk AT1 (h) + ẽ(k)AT2 (h) P (h+ ) E {VT +1 } − V0 ≤ −λm in (−Φ)E{|x̄k |2 }

× (A1 (h) + ẽ(k)A2 (h))x̄k |x̄k − x̄Tk P (h)x̄k which implies
E{|x̄k |2 } ≤ (λm in (−Φ))−1 (V0 − E {VT +1 }).
= x̄Tk AT1 (h)P (h+ )A1 (h) + f 2 AT2 (h)P (h+ )A2 (h)
Considering E{V (k)} ≥ 0 for all k ≥ 0, we have
− P (h) x̄k .
∞

Note that the inequality 2
E |x̄k | x̄0 ≤ (λm in (−Φ))−1 x̄T0 max(P (h))x̄0
[P (h+ ) − Ω(h+ )]T P −1 (h+ )[P (h+ ) − Ω(h+ )] ≥ 0
k =0
implies that = x̄T0 (λm in (−Φ))−1 max(P (h))x̄0

P (h+ ) − (Ω(h+ ) + ΩT (h+ )) ≥ −ΩT (h+ )P −1 (h+ )Ω(h+ ) = x̄T0 W x̄0
which together with (12) yields
  where x0 is the initial condition, and W = (λm in (−Φ))−1
Θ̃ 0 0 ΩT (h+ )A1 (h) ΩT (h+ )B1 (h) max(P (h)). According to Definition 1, the filter error system
 ∗ Θ̃ 0 f ΩT (h+ )A (h) f ΩT (h+ )B (h)  is stochastically stable in the mean square.
 2 2 
  Next, the H∞ performance criteria for the filter error sys-
 ∗ ∗ −I L̄(h) 0 <0
  tem in (8) will be established. To this end, assume zero initial
 
∗ ∗ ∗ −P (h) 0  conditions. An index is introduced as
∗ ∗ ∗ ∗ −γ 2 I J¯ = E { Vk +1 | ξk } + z̄kT z̄k − γ 2 wkT wk − x̄Tk P (h)x̄k
(14)
where Θ̃ = −ΩT (h+ )P −1 (h+ )Ω(h+ ). Clearly, Ω(h+ ) is in- where
vertible. Diag {Ω−T (h+ ), Ω−T (h+ ), I, I, I} and postmultiply-
x̄k
ing diag {Ω−1 (h+ ), Ω−1 (h+ ), I, I, I} on the left and right sides ξk =
wk
.
of (14), we obtained the following inequality:
  Since
−P −1 (h+ ) 0 0 A1 (h) B1 (h)
 ∗ −P −1 (h+ ) 0 f A2 (h) f B2 (h)  E {Vk +1 |ξk }
 
  T
 ∗ ∗ −I L̄(h) 0  A1 (h) + ẽ(k)AT2 (h) P (h+ ) 0
 <0 = E ξkT
  B1T (h) + ẽ(k)B2T (h) P (h+ )
 ∗ ∗ ∗ −P (h) 0  0
∗ ∗ ∗ ∗ −γ 2 I × [ A1 (h) + ẽ(k)A2 (h) B1 (h) + ẽ(k)B2 (h) ] ξk | ξk }
T
by Schur complement, which leads to T
A1 (h)
  = E ξk P (h+ ) [ A1 (h) B1 (h) ]
T T T P (h )
+
0 0 B1T (h)
A1 (h) f A2 (h) L̄ (h)   T
 0 P (h+ ) 0  A2 (h)
T T
B1 (h) f B2 (h) 0 +f 2
T
P (h ) [ A2 (h) B2 (h) ] ξk ξk
+
0 0 I B2 (h)
 
A1 (h) B1 (h) and
  P (h) 0
×  f A2 (h) f B2 (h)  − < 0. (15) L̄T (h)
0 γ2 I z̄kT z̄k = ξkT [ L̄(h) 0 ] ξk
L̄(h) 0 0
we have and R, S, W, Āf (h), B̄f (h), L̄f (h), for any h, h+ ∈ ρ satisfying

AT1 (h) ϕ11 ϕ12
J¯ = ξkT P (h+ ) [ A1 (h) B1 (h) ] <0 (17)
B1T (h) ∗ ϕ22

AT2 (h)
+ f2 P (h+ ) [ A2 (h) B2 (h) ] where
B2T (h)
Q1 (h+ ) Q2 (h+ ) R + RT S+W
T Ξ= −
L̄ (h) P (h) 0 QT2 (h+ ) Q3 (h+ ) W T + ST W + WT
+ [ L̄(h) 0 ] − ξk .
0 0 γ2 I (18)
From inequality (15), we know that    (11) (12)

Ξ 0 0 ϕ12 ϕ12
   (21) 
J¯ ≤ 0 ϕ11 =  ∗ Ξ 0 , ϕ12 = ϕ (22)
ϕ12 
 12 
that is ∗ ∗ −I (31)
ϕ12 0
 
E { Vk +1 | ξk } + z̄kT z̄k −γ 2
wkT wk − x̄Tk P (h)x̄k ≤ 0. −Q1 (h) −Q2 (h)
 0 
Take mathematical expectation on both sides, we have
ϕ22 = −QT2 (h) −Q3 (h) (19)
0 −γ I 2
E{Vk +1 } − E{Vk } + E z̄kT z̄k − γ 2 wkT wk ≤ 0.

and
For k = 0, 1, 2, . . . , summing up both sides, considering
(11) RT A(h) + ēB̄f (h)C(h) Āf (h)
E{V (k)} ≥ 0 for all k ≥ 0, under zero initial condition, we ϕ12 =
obtain S T A(h) + ēB̄f (h)C(h) Āf (h)
∞ T
∞
(12) R B(h) + ēB̄f (h)D(h)
ϕ12 =
E z̄k z̄k −
T
γ 2 wkT wk ≤ 0 S T B(h) + ēB̄f (h)D(h)
k =0 k =0
(21) B̄f (h)C(h) 0 (22) B̄f (h)D(h)
which is equivalent to (11). The proof is completed. ϕ12 =f , ϕ12 = f
Remark 1: If there is no data dropout in the channel between B̄f (h)C(h) 0 B̄f (h)D(h)
the physical plant and the filter, that is, perfect communication (31)
ϕ12 = [ L(h) −L̄f (h) ] (20)
links exist between the plant and the filter, then we have the
following corollary, which can be proved by following similar then there exists a fuzzy filter in the form of (6) such that
arguments, as in the proof of Theorem 1. the filtering error system in (8) is stochastically stable with a
Corollary 1: Consider the fuzzy system in (3) and suppose prescribed H∞ norm bound γ. Moreover, if the aforementioned
that the filter in (6) is given. When ē = 1, the filter error system condition is satisfied, the matrices for the filter in (6) are given
in (8) is stochastically stable with a given H∞ performance γ, by
if there exist fuzzy-basis-dependent matrices P (h) > 0, Ω(h), −T
Af (h) Bf (h) Ω4 0 Āf (h) B̄f (h)
for any h, h+ ∈ ρ, satisfying =
  Lf (h) 0 0 I L̄f (h) 0
Θ 0 ΩT (h+ )A1 (h) ΩT (h+ )B1 (h) −1
 ∗ −I  Ω4 Ω3 0
 L̄(h) 0  × (21)
 <0 (16) 0 I
∗ ∗ −P (h) 0 
where Ω3 and Ω4 can be obtained by the decomposition on W .
∗ ∗ ∗ −γ 2 I
Proof: Suppose that there exist matrices Q(h) > 0, R, S,
where h+ and Θ are defined in (12). W, Āf (h), B̄f (h), and L̄f (h) satisfying (17). From (17), we
know that W > 0. One can always find square and nonsingular
IV. FILTER DESIGN matrices Ω3 and Ω4 that W = ΩT4 Ω−1 3 Ω4 . Let
In this section, we will design a fuzzy filter in the form of (6) R = Ω1 , S = Ω2 Ω−1
3 Ω4
based on Theorem 1, that is, to determine the filter matrices in
Ω1 Ω2 I 0
(6) such that the filter error system in (8) is stochastically stable Ω= , T = (22)
with a guaranteed H∞ performance. Since the condition in (12) Ω4 Ω3 0 Ω−1
3 Ω4
cannot be utilized to obtain the filter directly, we introduce some and
slack matrices, which will simplify the filter design procedure.
Theorem 2: Consider the fuzzy system in (3). For a given pos- −T
Q1 (h+ ) Q2 (h+ ) −1
P1 (h+ ) P2 (h+ )
T T =
itive constant γ, if there exist fuzzy-basis-dependent matrices QT2 (h+ ) Q3 (h+ ) P2T (h+ ) P3 (h+ )

Q1 (h) Q2 (h) Af (h) Bf (h)
Q(h) = >0
QT2 (h) Q3 (h) Lf (h) 0
 
Ω−T
4 0 Āf (h) B̄f (h) Ω−1
4 Ω3 0 T 0 0 0 0
= . (23) 0
0 I L̄f (h) 0 0 I  T 0 0 0

 
×
0 0 I 0 0
<0
By (22) and (23), one has  
0 0 0 T 0
+ +
P1 (h ) P2 (h ) 0 0 0 0 I
T −T ΞT −1 = − Ω − ΩT . (24)
P2T (h ) +
P3 (h ) +
where Θ is defined in (12), which together with (17) implies
that, for any h, h+ ∈ ρ, (12) holds.
With the support of (10), (22), and (23), it can be verified that The proof is completed.
The condition in (17) cannot be directly employed for filter
T T ΩT A1 (h)T design. One way to facilitate Theorem 2 for the construction of a

ΩT1 A(h) + ēΩT4 Bf (h)C(h) ΩT4 Af (h)Ω−1
3 Ω4
fuzzy filter is to convert (17) into a finite set of LMI constraints.
= To this end, one must further restrict the choice of the fuzzy-
ΩT4 Ω−T T T T −1
3 Ω2 A(h) + ēΩ4 Bf (h)C(h) Ω4 Af (h)Ω3 Ω4 basis-dependent Lyapunov functions. The following theorem
T
R A(h) + ēB̄f (h)C(h) Āf (h) gives a possible way to achieve this.
= Theorem 3: Consider the fuzzy system in (3).
S T A(h) + ēB̄f (h)C(h) Āf (h) For a given
Q1i Q2i
f T T ΩT A2 (h)T positive constant γ, if there exist matrices Qi = >
QT2i Q3i
T 0, and R, S, W, Āf i , B̄f i , L̄f i , for all i, j, l ∈ {1, . . . , r}
Ω4 Bf (h)C(h) 0 B̄f (h)C(h) 0
=f =f satisfying
ΩT4 Bf (h)C(h) 0 B̄f (h)C(h) 0
T T ΩT B1 (h)
ψ11 ψ12
<0 (27)
ΩT1 B(h) + ēΩT4 Bf (h)D(h) ∗ ψ22
=
ΩT4 Ω−T T T
3 Ω2 B(h) + ēΩ4 Bf (h)D(h) where
T
R B(h) + ēB̄f (h)D(h)
=
S T B(h) + ēB̄f (h)D(h) Q1l Q2l R + RT S+W
Ξ̄ = − (28)
f T T ΩT B2 (h) QT2l Q3l W T + ST W + WT
 
T Ξ̄ 0 0
Ω4 Bf (h)D(h) B̄f (h)D(h)
=f = f  
ΩT4 Bf (h)D(h) B̄f (h)D(h) ψ11 =  ∗ Ξ̄ 0 ,
∗ ∗ −I
L̄(h)T  
−Q1i −Q2i
= [ L(h) −Lf (h)Ω−1
3 Ω4 ] = [ L(h) −L̄f (h) ] . (25) 0
ψ22 =  −QT2i −Q3i 
0 −γ I 2
Letting
 RT A + ēB̄ C Ā RT B + ēB̄ D 
i fi j fi i fi j

Q1 (h) Q2 (h)  S T Ai + ēB̄f i Cj Āf i T
S Bi + ēB̄f i Dj 
Ω(h) = Ω, P (h) = T −T T −1 (26)  
QT2 (h) Q3 (h)  
ψ12 = B̄ C
fi j 0 B̄f i Dj 
 f f 
 B̄f i Cj 0 B̄f i Dj 
one can readily obtain from (24)–(26) that (17) is equivalent to
[ Li −L̄f i ] 0
 T 
T 0 0 0 0 (29)
 0 T
0
 T 0 0  then there exists a fuzzy filter in (6) such that the filter er-
 
 0 0 I 0 0 ror system in (8) is stochastically stable with a prescribed H∞
 
  norm bound γ. Moreover, if the earlier condition is satisfied, the
 0 0 0 TT 0
matrices for the filter in (6) are given by
0 0 0 0 I
 
Θ 0 0 ΩT (h+ )A1 (h) ΩT (h+ )B1 (h) Af (h) Bf (h)
r
Ω−T 0 Āf i B̄f i
4
∗
 Θ 0 f ΩT (h+ )A2 (h) f ΩT (h+ )B2 (h) 
 Lf (h) 0
= hi
0 I L̄f i 0
  i=1
× 
∗ ∗ −I L̄(h) 0 
Ω−1

  4 Ω3 0
∗ ∗ ∗ −P (h) 0  × (30)
0 I
∗ ∗ ∗ ∗ −γ I 2
where Ω3 and Ω4 can be obtained by the decomposition on W .
Proof: Suppose that there exist matrices R, S, W, Āf i , B̄f i ,

Q1i Q2i
L̄f i , and Qi = > 0, for all i, j, l ∈ {1, . . . , r}
QT2i Q3i
satisfying (17). Then, we use these matrices and the fuzzy basis
function h ∈ ρ to define the following functions:
r r
Q1l Q2l
Q(h) = hl , Āf (h) = hi Āf i
l=1
QT2l Q3l i=1

r
r
B̄f (h) = hi B̄f i , L̄f (h) = hi L̄f i
i=1 i=1
Fig. 2. Tunnel diode circuit.
which together with (4) imply that
r r
ϕ11 ∗ r
ψ11 ∗ where
= hi hj h+
ϕ21 ϕ22 l
ψ21 ψ22
i=1 j =1 l=1 Q1 Q2 R + RT S+W
Ξ̃ = −
and (17) is clearly verified, where h+ QT2 W T + ST W + W T
Q3
l is hl (θk +1 ), as is defined    
in (12), and ϕ11 , ϕ21 , ϕ22 , ψ11 , ψ21 , and ψ22 are defined as in Ξ̃ 0 0 −Q1 −Q2
0
(17), (19), (27), and (29). ψ̃11 =  ∗ Ξ̃ 0  , ψ̃22 =  −QT2 −Q3 
Corollary 2: Consider the fuzzy system in (3). When ē = 1, ∗ ∗ −I 0 −γ I
2
a given positive constant γ, if there exist matrices Qi =

for
Q1i Q2i and ψ12 is defined in (29), then there exists a fuzzy filter in the
> 0, and R, S, W, Āf i , B̄f i , and L̄f i, for all i, j,
QT2i Q3i form of (6) such that the filter error system in (8) is stochastically
l ∈ {1, . . . , r} satisfying stable with a prescribed H∞ norm bound γ. Moreover, if the
earlier condition is satisfied, the matrices for the filter in (6) are
Π11 Π12 given by (30).
<0 (31)
∗ Π22 Remark 3: Theorem 3 is obtained by restricting the fuzzy-
basis-dependent Lyapunov functions. The expression of fuzzy-
where basis-dependent Lyapunov functions adopted here is consistent
 
−Q1i −Q2i with the compact presentation of system matrices in (3), which is
Ξ̄ 0 0
Π11 = , Π22 =  −QT2i −Q3i  adopted by mostc of the literature. This fuzzy-basis-dependent
0 −I Lyapunov approach has been recognized to be less conservative.
−γ 2 I0
 T T  However, in this basis-dependent framework, the restriction on
R Ai + B̄f i Cj Āf i R Bi + B̄f i Dj the Lyapunov function still introduces some overdesign. How to
Π12 =  S T Ai + B̄f i Cj Āf i S T Bi + B̄f i Dj  further reduce this conservatism still needs further investigation.
[ Li −L̄f i ] 0 Remark 4: The number of inequalities in Theorem 3 will
(32) increase with the number of fuzzy rules of the model, thus;
a computational problem might arise for high-order nonlinear
and Ξ̄ is defined in (28), then there exists a fuzzy filter in the systems. One effective way to solve this problem is to try to
form of (6), and the filter error system in (8) is stochastically reduce the number of fuzzy rules when modeling the nonlinear
stable with a prescribed H∞ norm bound γ. Moreover, if the system based on fuzzy logic, which can be found in [28].
previous condition is satisfied, the matrices for the filter in (6)
are given by (30).
V. ILLUSTRATIVE EXAMPLE
Remark 2: Theorem 3 is obtained by using the basis-
dependent Lyapunov function. It is clear that when Qi = Q In this section, we use an example to illustrate the effective-
for any i ∈ {1, . . . , r}, (13) becomes the quadratic Lyapunov ness of the theoretical results developed before.
function that has been widely used in the literature. Then, Consider a tunnel diode circuit shown in Fig. 2, whose fuzzy
the following corollary based on the quadratic approach is modeling was done in [2], where x1 (t) = vC (t), x2 (t) = iL (t),
obtained. w(t) is the disturbance noise input, y(t) is the measurement
Corollary 3: Consider the fuzzy system in (3). output, and z(t) is the controlled output. With a sampling time
For a given
Q1 Q2 T = 0.02, the discrete-time model is obtained as
positive constant γ, if there exist matrices Q = >
QT2 Q3
0, R, S, W, Āf i , B̄f i , and L̄f i , for all i, j ∈ {1, . . . , r} satisfying xk +1 = A(h)xk + B(h)wk
yk = C(h)xk + D(h)wk
ψ̃11 ψ12
<0 (33)
∗ ψ̃22 zk = L(h)xk (34)
where

0.9887 0.9024 0.0093
A1 = , B1 =
−0.0180 0.8100 0.0181

0.90337 0.8617 0.0091
A2 = , B2 =
−0.0172 0.8103 0.0181
C1 = [ 1 0 ] , C2 = [ 1 0 ]
D1 = 1, D2 = 1, L1 = [ 1 0], L2 = [ 1 0 ] .
To show the effectiveness of the obtained results, we assume the

membership function to be
 (1)

 x +3
−3 ≤ xk ≤ 0
(1)
 k
 ,

 3


 0,
(1)
xk < −3
h1 =
 3 − x(1)
 Fig. 3. Estimation error when w k ≡ 0.

 k
0 ≤ xk ≤ 3
(1)

 ,

 3
 (1)
0, xk > 3
h 2 = 1 − h1 . (35)
The purpose here is to design a filter in the form of (5) such that
the system in (34) is stochastically stable with a guaranteed H∞
norm bound γ.
Suppose ē = 0.8. By solving LMI (27), the minimum H∞
performance γ ∗ = 0.1463 is obtained, and the filter matrices
are obtained:

4.9722 14.2575 −0.3451
Āf 1 = , B̄f 1 =
9.1080 69.4648 −1.4883

4.6054 13.9920 −0.2005
Āf 2 = , B̄f 2 = (36)
8.4237 67.9560 −0.8318
L̄f 1 = [ −1.0000 −0.0002 ] (37)
L̄f 2 = [ −0.9955 0.0152 ] (38)
Fig. 4. Data packet dropout.
5.5335 11.5195
W = . (39)
11.5162 71.9180
To illustrate the performance of the designed filter, we assume
By (30), we have the initial conditions and the external disturbance w(k) to be

0.9524 0.8488 −0.0289  30 ≤ k ≤ 50
Af 1 = , Bf 1 =  2,
−0.0259 0.8300 −0.0161 w(k) = −2, 70 ≤ k ≤ 100 (40)


0.8827 0.8423 −0.0182 0, elsewhere.
Af 2 = , Bf 2 =
−0.0242 0.8100 −0.0086
Lf 1 = [ −1.0000 −0.0002 ] In the simulation, the data packet dropouts are generated ran-
domly according to ē = 0.8, which is shown in Fig. 4. Fig. 5
Lf 2 = [ −0.9955 0.0152 ] . shows the response of signal z̄(k). Fig. 6 shows the simu-
lation results of zk and ẑk . By calculation, we obtain that
First, we assume that wk ≡ 0 and the that initial condition ||z̄||22 = 1.0881 and ||w||22 = 208, which yields γ = 0.0723 (be-
x0 = [ 0.2 −0.8 ] , x̂0 = [ 0 0 ] . Fig. 3 shows that the esti- low the minimum γ ∗ = 0.1463), showing the effectiveness of
mation error response converges to zero. the filter design.
control of fuzzy systems with simultaneous consideration of

packet dropout and signal delay deserves further investigation.
REFERENCES
[1] J. Abedor, K. Nagpal, and K. Poolla, “A linear matrix inequality approach
to peak-to-peak gain minimization,” Int. J. Robust Nonlinear Control,
vol. 6, pp. 899–927, 1996.
[2] W. Assawinchaichote and S. K. Nguang, “H ∞ filtering for fuzzy dynamic
systems with D stability constraints,” IEEE Trans. Circuits Syst. I, vol. 50,
no. 11, pp. 1503–1508, Nov. 2003.
[3] W. Assawinchaichote and S. K. Nguang, “H ∞ filtering for fuzzy singularly
perturbed systems with pole placement constraints: An LMI approach,”
IEEE Trans. Signal Process., vol. 52, no. 6, pp. 1659–1667, Jun. 2004.
[4] X. Ban, X. Gao, X. Huang, and H. Yin, “Stability analysis of the sim-
plest Takagi–Sugeno fuzzy control system using popov criterion,” Int. J.
Innovative Comput., Inform. Control, vol. 3, no. 5, pp. 1087–1096, 2007.
[5] S. G. Cao, N. W. Rees, and G. Feng, “Analysis and desgign of a class of
continuous time fuzzy control systems,” Int. J. Control, vol. 64, no. 64,
pp. 1069–1087, 1996.
[6] Y. Y. Cao and P. M. Frank, “Robust H ∞ disturbance attenuation for a
class of uncertain discrete-time fuzzy systems,” IEEE Trans. Fuzzy Syst.,
vol. 8, no. 4, pp. 406–415, Aug. 2000.
Fig. 5. Estimation error. [7] W. Chang, C. Ku, and P. Huang, “Passive fuzzy control with relaxed con-
ditions for discrete affine T–S fuzzy systems,” Int. J. Innovative Comput.
Inform. Control, vol. 3, no. 4, pp. 853–871, 2007.
[8] G. Feng, “Robust H ∞ filtering of fuzzy dynamic systems,” IEEE Trans.
Aerosp. Electron. Syst., vol. 41, no. 2, pp. 658–671, Apr. 2005.
[9] G. Feng, “A survey on analysis and design of model-based fuzzy control
systems,” IEEE Trans. Fuzzy Syst., vol. 14, no. 5, pp. 676–697, Oct. 2006.
[10] H. Gao and C. Wang, “Delay-dependent robust H ∞ and L 2 –L ∞ filtering
for a class of uncertain nonlinear time-delay systems,” IEEE Trans.
Automat. Control, vol. 48, no. 9, pp. 1661–1666, Sep. 2003.
[11] H. Gao and C. Wang, “Robust L 2 – L ∞ filtering for uncertain systems
with multiple time-varying state delays,” IEEE Trans. Circuits Syst. I,
vol. 50, no. 4, pp. 594–599, Apr. 2003.
[12] H. Gao and C. Wang, “A delay-dependent approach to robust H ∞ filtering
for uncertain discrete-time state-delayed systems,” IEEE Trans. Signal
Process., vol. 52, no. 6, pp. 1631–1640, Jun. 2004.
[13] H. Gao, Z. Wang, and C. Wang, “Improved H ∞ control of discrete-time
fuzzy systems: A cone complementarity linearization approach,” Inform.
Sci., vol. 175, no. 1–2, pp. 57–77, 2005.
[14] Y. He, Q. G. Wang, and C. Lin, “An improved H ∞ filter design for systems
with time-varying interval delay,” IEEE Trans. Circuits Syst. II, vol. 53,
no. 11, pp. 1235–1239, Nov. 2006.
[15] J. Lam and S. S. Zhou, “Dynamic output feedback H ∞ control of discrete-
time fuzzy systems: A fuzzy-basis-dependent Lyapunov function ap-
proach,” Int. J. Syst. Sci., vol. 38, no. 1, pp. 25–37, 2007.
[16] C. Lin, Q. G. Wang, T. H. Lee, and Y. He, “Fuzzy weighting-dependent
Fig. 6. Estimation signals. approach to H ∞ filter design for time-delay fuzzy systems,” IEEE Trans.
Signal Process., vol. 55, no. 6, pp. 2746–2751, Jun. 2007.
[17] H. Liu, F. Sun, K. He, and Z. Sun, “Design of reduced-order H ∞ filter for
VI. CONCLUDING REMARKS Markovian jumping systems with time delay,” IEEE Trans. Circuits Syst.
II, vol. 51, no. 11, pp. 607–612, Nov. 2004.
In this paper, the problem of H∞ fuzzy filtering of non- [18] H. Liu, F. Sun, and Y. N. Hu, “H ∞ control for fuzzy singularly perturbed
linear systems under unreliable communication links has been systems,” Fuzzy Sets Syst., vol. 155, pp. 272–291, 2005.
[19] H. Liu, F. Sun, and Z. Sun, “Reduced-order filtering with energy-to-peak
investigated. The T–S fuzzy system is utilized to model the performance for discrete-time Markovian jumping systems,” IMA J. Math.
nonlinear plant, and the communication link failure is modeled Control Inform., vol. 21, no. 2, pp. 143–158, 2004.
via a stochastic variable satisfying the Bernoulli random binary [20] H. Liu, F. Sun, and Z. Sun, “Stability analysis and synthesis of fuzzy
singularly perturbed systems,” IEEE Trans. Fuzzy Syst., vol. 13, no. 2,
distribution. The basis-dependent Lyapunov function has been pp. 273–284, Apr. 2005.
used to design an H∞ filter such that the filter error system [21] Y. G. Niu and D. W. C. Ho, “Robust observer design for lto stochastic
is stochastically stable and preserves a guaranteed H∞ perfor- time-delay systems via sliding mode control,” Syst. Control Lett., vol. 55,
no. 10, pp. 781–793, 2006.
mance. Some slack matrices have been introduced to facilitate [22] P. Seiler and R. Sengupta, “An H ∞ approach to networked control,” IEEE
the H∞ filter design. An example has been given to illustrate Trans. Autom. Control, vol. 50, no. 3, pp. 356–364, Mar. 2005.
the effectiveness of the proposed approach. [23] P. Shi, “Filtering on sampled-data systems with parametric uncertainty,”
IEEE Trans. Autom. Control, vol. 43, no. 7, pp. 1022–1027, Jul. 1998.
It should be noted that in practical networked control systems, [24] P. Shi, “Filtering for interconnected nonlinear sampled-data systems with
in addition to data missing, the phenomenon of transmission parametric uncertainties,” J. Vib. Control, vol. 5, no. 4, pp. 591–618,
delay often occurs. It can also degrade the performance of the 1999.
[25] P. Shi, E. K. Boukas, and R. K. Agarwal, “Kalman filtering for continuous-
systems and even cause system instability. In this paper, we time uncertain systems with Markovian jumping parameters,” IEEE
have only considered data missing, but the study of networked Trans. Autom. Control, vol. 44, no. 8, pp. 1592–1597, Aug. 1999.
[26] B. Sinopoli, L. Schenato, and M. Franceschetti, “Kalman filtering with and Robotics Systems, Circuits, System and Signal Processing, etc. He was an
intermittent observations,” IEEE Trans. Autom. Control, vol. 49, no. 9, outstanding reviewer for the Automica in 2007. His current research interests
pp. 1453–1464, Sep. 2004. include network-based control, robust control, and time-delay systems and their
[27] K. Tanaka and M. Sugeno, “Stability analysis and design of fuzzy sys- industrial applications.
tems,” Fuzzy Sets Syst., vol. 45, no. 2, pp. 135–156, 1992. Prof. Gao is an Associate Editor of the IEEE TRANSACTIONS ON SYSTEMS,
[28] K. Tanaka and H. O. Wang, Fuzzy Control Systems Design and Analysis: MAN AND CYBERNETICS PART B: CYBERNETICS and the IEEE TRANSACTIONS
A Linear Matrix Inequality Approach. New York: Wiley, 2001. ON INDUSTRIAL ELECTRONICS. He was an outstanding reviewer for the IEEE
[29] G. C. Walsh, H. Ye, and L. Bushnell, “Stability analysis of networked TRANSACTIONS ON AUTOMATIC CONTROL in 2008, and an appreciated reviewer
control systems,” IEEE Trans. Control Syst. Technol., vol. 10, no. 3, for the IEEE TRANSACTIONS ON SIGNAL PROCESSING in 2006. He was the re-
pp. 438–446, May 2002. cipient of the University of Alberta Dorothy J. Killam Memorial Postdoctoral
[30] H. O. Wang, K. Tanaka, and M. F. Griffin, “An approach to fuzzy control Fellow Prize in 2005, the National Outstanding Youth Science Fund in 2008,
of nonlinear systems: Stability and design issues,” IEEE Trans. Fuzzy and the National Outstanding Doctoral Thesis Award in 2007. He was the core-
Syst., vol. 4, no. 2, pp. 14–23, Feb. 1996. cipient of the National Natural Science Award of China in 2008.
[31] Z. Wang, D. W. C. Ho, and X. Liu, “Variance-constrained filtering for
uncertain stochastic systems with missing measurements,” IEEE Trans.
Autom. Control, vol. 48, no. 7, pp. 1254–1258, Jul. 2003.
[32] Z. Wang, D. W. C. Ho, and X. Liu, “Variance-constrained control for Yan Zhao received the B.S. degree in chemical engi-
uncertain stochastic systems with missing measurement,” IEEE Trans. neering and equipment control and the M.S. degree
Syst., Man Cybern.—Part A, vol. 35, no. 5, pp. 746–753, Sep. 2005. in mechanical engineering from the Inner Mongolia
[33] Z. Wang, B. Huang, and P. Huo, “Sampled-data filtering with error covari- University of Technology, Hohhot, China, in 2002
ance assignment,” IEEE Trans. Signal Process., vol. 49, no. 3, pp. 666– and 2005, respectively. She is currently working to-
670, Mar. 2001. ward the Ph.D. degree in control science and engi-
[34] Z. Wang, F. Yang, D. W. C. Ho, and X. Liu, “Robust H ∞ filtering for neering with Harbin Institute of Technology, Harbin,
stochastic time-delay systems with missing measurements,” IEEE Trans. China.
Signal Process., vol. 54, no. 7, pp. 2579–2587, Jul. 2006. Her current research interests include fuzzy con-
[35] Z. Wang, J. H. Zhu, and H. Unbehauen, “Robust filter design with time- trol systems, robust control, and networked control
varying parameter uncertainty and error variance constraints,” Int. J. systems.
Control, vol. 72, no. 1, pp. 30–38, 1999.
[36] H. Wu, “Reliable LQ fuzzy control for continuous-time nonlinear systems
with actuator faults,” IEEE Trans. Syst., Man Cybern.—Part B, vol. 34, James Lam (S’86–M’87–SM’99) received the B.Sc.
no. 4, pp. 1743–1752, Aug. 2004. degree (with first class) in mechanical engineer-
[37] H. Wu and K. Y. Cai, “Mode-independent robust stability for uncertain ing from the University of Manchester, Manchester,
Markovian jump nonlinear systems in fuzzy control,” IEEE Trans. Syst., U.K., in 1983, and the M.Phil. and Ph.D. degrees
Man Cybern.—Part B, vol. 36, no. 3, pp. 509–519, Jun. 2006. in control engineering from the University of Cam-
[38] S. Xu, T. Chen, and J. Lam, “Robust H ∞ filtering for uncertain Markovian bridge, Cambridge, U.K., in 1985 and 1988, respec-
jump systems with mode-dependent time-delays,” IEEE Trans. Autom. tively. He received the Ashbury Scholarship, the A.H.
Control, vol. 48, no. 5, pp. 900–907, May 2003. Gibson Prize, and the H. Wright Baker Prize for his
[39] S. Xu and J. Lam, “Exponential H ∞ filter design for uncertain Takagi– academic performance.
Sugeno fuzzy systems with time delay,” Eng. Appl. Artif. Intell., vol. 17, He is currently a Professor in the Department of
pp. 645–659, 2004. Mechanical Engineering, University of Hong Kong,
[40] M. Yu, L. Wang, and T. Chu, “Sampled-data stabilization of networked Hong Kong. He is an Associate Editor of the Asian Journal of Control, Interna-
control systems with nonlinearity,” Proc. Inst. Elec. Eng. D, Control tional Journal of Systems Science, Journal of Sound and Vibration, International
Theory Appl., vol. 152, no. 6, pp. 609–614, 2005. Journal of Applied Mathematics and Computer Science, Journal of the Franklin
[41] D. Yue and Q.-L. Han, “Robust H ∞ filter design of uncertain descrip- Institute, Dynamics of Continuous, Discrete and Impulsive Systems (Series B:
tor systems with discrete and distributed delays,” IEEE Trans. Signal Applications and Algorithms), and Automatica. He is also a member of the
Process., vol. 52, no. 11, pp. 3200–3212, Nov. 2004. Editorial Board of the Institute of Engineering and Technology (IET) Control
[42] D. Yue, Q.-L. Han, and J. Lam, “Network-based robust H ∞ control of Theory and Applications, Open Electrical and Electronic Engineering Journal,
systems with uncertainty,” Automatica, vol. 41, no. 6, pp. 999–1007, Research Letters in Signal Processing, International Journal of Systems, Control
2005. and Communications, and Journal of Electrical and Computer Engineering. His
[43] W. Zhang, M. Branicky, and S. Phillips, “Stability of networked control current research interests include reduced-order modeling, delay systems, de-
systems,” IEEE Control Syst. Mag., vol. 21, no. 1, pp. 84–99, Feb. 2001. scriptor systems, stochastic systems, multidimensional systems, robust control,
[44] S. Zhou, G. Feng, and J. Lam, “Fuzzy robust H ∞ control for discrete- and filtering. He was an Editor-in-Chief of the Institute of Electrical Engineers
time fuzzy systems via basis-dependent Lyapunov functions,” Inform. (IEE) Proceedings Control Theory and Applications.
Sci., vol. 174, pp. 197–217, 2004. Prof. Lam is a Chartered Mathematician and a Chartered Scientist. He is a
[45] S. Zhou, J. Lam, and A. K. Xue, “H ∞ filtering of discrete-time fuzzy Fellow of the Institute of Mathematics and Its Applications, and the IET. He is
systems via basis-dependent Lyapunov function approach,” Fuzzy Sets a Scholar and a Fellow of the Croucher Foundation. He is an Associate Editor
Syst., vol. 158, pp. 180–193, 2007. of the IEEE TRANSACTIONS ON SIGNAL PROCESSING.
Ke Chen received the B.S. degree in mathematics in

Huijun Gao (M’06) was born in Heilongjiang, 2002 and the M.S. degree in bioinformatics in 2005
China, in 1976. He received the M.S. degree in elec- from Nankai University, Tianjin, China. He is cur-
trical engineering from Shenyang University of Tech- rently working toward the Ph.D. degree in electrical
nology, Shenyang, China, in 2001, and the Ph.D. de- and computer engineering with the University of Al-
gree in control science and engineering from Harbin berta, Edmonton, AB, Canada.
Institute of Technology, Harbin, China, in 2005. His current research interests include the applica-
From November 2003 to August 2004, he was tion of mathematical models in biological sciences.
a Research Associate in the Department of Mechan-
ical Engineering, University of Hong Kong, Hong
Kong. In November 2004, he joined Harbin Insti-
tute of Technology, where he is currently a Professor.
From October 2005 to October 2007, he was a Postdoctoral Researcher in the
Department of Electrical and Computer Engineering, University of Alberta,
Edmonton, AB, Canada. He is an Associate Editor of the Journal of Intelligent
An Interval Type-2 Fuzzy Rough Set Model

for Attribute Reduction
Haoyang Wu, Yuyuan Wu, and Jinping Luo
Abstract—Rough set theory is a very useful tool for describing have difficulties in handling such values mentioned earlier [4].
and modeling vagueness in ill-defined environments. Traditional For example, two close real values may only differ as a result
rough set theory is restricted to crisp environments. However, of noise; however, in traditional rough set theory, one of them
nowadays, it has been extended to fuzzy environments, resulting in
the development of the so-called fuzzy rough sets. Type-2 fuzzy sets may be considered different from the other in a different order
possess many advantages over type-1 fuzzy sets, but for the gen- of magnitude. A possible solution is to discretize the data set
eral type-2 fuzzy sets, the computational complexity is severe. On beforehand and obtain a new crisp data set, but this method also
the other hand, set-theoretic and arithmetic computations for the generates information loss [4].
interval type-2 fuzzy sets are very simple. Motivated by the afore- To overcome this shortcoming, some methods that can ex-
mentioned accomplishments, in this paper, the concept of fuzzy
rough sets is generalized to interval type-2 fuzzy environments. tend the traditional rough set theory to deal with real-valued
Subsequently, a method of attribute reduction within the interval domains have been proposed [3], [17], [18]. Among these
type-2 fuzzy rough set framework is proposed. Lastly, the proper- methods, hybridization of fuzzy and rough sets can lead one
ties of the interval type-2 fuzzy rough sets are presented. to consider not only rough approximations of fuzzy sets, but
Index Terms—Fuzzy rough sets, interval type-2 fuzzy sets, rough also approximations of sets by means of similarity relations
approximation, rough sets, type-2 fuzzy sets. or fuzzy partitions. Dubois and Prade [3] claimed that these
hybrid notions came up in a natural way when a linguistic cat-
I. INTRODUCTION egory, denoting a set of objects, had to be approximated in
terms of already existing labels, or when the indiscernibility
OUGH set theory was proposed by Pawlak in 1982 [1]. It
R is a very useful tool to detect the data dependencies and re-
duce the number of attributes contained in a data set. Since then,
relation between the objects no longer obeyed ideal laws of
equivalence and was a matter of degree. According to Jensen
and Shen [6], fuzzy rough sets offer a high degree of flexi-
the rough set theory has attracted a lot of researchers around the
bility, providing robust solutions and advanced tools for data
world, and has been applied to many domains [3]–[14]. Its suc-
analysis.
cess is partially owing to the following properties: 1) only the
On the other hand, there exist some disadvantages in using
facts hidden in the data are analyzed; 2) no additional infor-
fuzzy sets for reasoning from the data, and the hybridization of
mation about the data is required, such as thresholds or expert
fuzzy and rough sets may help to deal with them. According to
knowledge; and 3) given a data set with discretized attribute
Liu et al. [11], when a fuzzy implication inference rule and a
values, it is possible to find a subset (termed as reduct) within
premise are given, only one fixed value that acts as the final rea-
the original attributes that are the most informative [4]. Fuzzy
soning consequence can be obtained. However, in many cases,
set theory was proposed by Zadeh in 1965 [2]. It extended the
there exist some special constraints that must be satisfied. Liu
classical notion of set and permitted the gradual assessment of
et al. [11] studied the fuzzy reasoning based on fuzzy rough sets.
the membership of elements in a set. To date, fuzzy set the-
According to their research, the fuzzy reasoning consequence
ory has shown enormous promise in many disciplines, such as
can be expressed by a fuzzy interval, and a suitable value can
artificial intelligence, control systems and so on [15], [16].
be selected from the fuzzy interval as a final reasoning conse-
It is generally accepted that theories of fuzzy and rough sets
quence. Thus, it is possible to satisfy some special constraints.
are related but distinct, and they are complementary [10]. Tra-
So far, many studies have been carried out in the field of
ditional rough set theory is based on an equivalent relation.
fuzzy rough sets. For example, Dubois and Prade [3] combined
However, the values of attributes may be both symbolic and
fuzzy and rough sets by proposing a definition for rough fuzzy
real-valued. In other words, the traditional rough set theory will
sets and fuzzy rough sets. Shen and Jensen [8] proposed an ap-
proach that integrated a fuzzy rule induction algorithm with a
Manuscript received September 6, 2007; revised June 29, 2008, September fuzzy rough method for feature selection. Tsang et al. [9] es-
17, 2008, and December 10, 2008; accepted December 17, 2008. First published tablished solid mathematical foundations for attribute reduction
January 27, 2009; current version published April 1, 2009.
H. Wu and Y. Wu are with the State Key Laboratory of Multiphase with fuzzy rough sets. Furthermore, Yeung et al. [10] inves-
Flow in Power Engineering, School of Energy and Power Engineering, Xi’an tigated the generalization of fuzzy rough sets, while Wu and
Jiaotong University, Xi’an 710049, China (e-mail: hywch@mail.xjtu.edu.cn; Zhang [12] proposed constructive and axiomatic approaches of
yywu@mail.xjtu.edu.cn).
J. Luo is with the Kunlun Technology Industry Corporation, Hangzhou fuzzy approximation operators. Lingras and Jensen [13] pre-
310012, China (e-mail: jinpluo@hotmail.com). sented a review on the fuzzy and rough hybridization in su-
Color versions of one or more of the figures in this paper are available online pervised learning, information retrieval, feature selection, and
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TFUZZ.2009.2013458 neural and evolutionary computing.
1063-6706/$25.00 © 2009 IEEE
To date, most of the research works in the field of fuzzy can also be expressed as
where 0 ≤ µÃ (x, u) ≤ 1. A
rough sets are restricted to ordinary (type-1) fuzzy environ-
ments. Recently, Wu et al. [14] proposed a measure of rough- A≡ µÃ (x, u)/(x, u)
ness of a type-2 fuzzy set based on similarity relation. Jensen x∈U u ∈J x
and Shen [7] proposed an interval-valued approach for the fuzzy
rough feature selection, which could not only handle the missing ≡ fx (u)/u x, Jx ⊆ [0, 1] (2)
x∈U u ∈J x
values effectively, but also handle the uncertainty that could not
be modeled by a type-1 approach. According to Mendel [20], where fx (u) = µÃ (x, u).
there exist at least four sources of uncertainties in type-1 fuzzy The class of all type-2 fuzzy sets of the universe U is denoted
logic systems (T1 FLS), which are as follows: 1) meanings of as F̃ (U ) in the following sections.
the words that are used in the antecedents and consequents of Definition 2: At each value of x, say x = x , the 2-D plane
rules can be uncertain (words mean different things to different whose axes are u and µÃ (x , u) is called the vertical slice of
people); 2) consequents may have a histogram of values associ- µÃ (x, u). A secondary membership function is a vertical slice
ated with them, especially when knowledge is extracted from a of µÃ (x, u). It is µÃ (x = x , u) for x ∈ U and ∀u ∈ Jx ⊆ I,
group of experts, all of whom do not collectively agree; 3) mea- i.e.
surements that activate a T1 FLS may be noisy and therefore
uncertain; and 4) the data that are used to tune the parameters µÃ (x = x , u) ≡ µÃ (x ) ≡ fx (u)/u, Jx ⊆ I (3)
u ∈J x
of a T1 FLS may also be noisy. All these uncertainties lead to
uncertain fuzzy-set membership functions. where 0 ≤ fx (u) ≤ 1. The amplitude of a secondary member-
Ordinary type-1 fuzzy sets cannot model such uncertainties ship function is called a secondary grade. In formula (2), fx (u)
directly, because they are characterized by crisp membership and µÃ (x, u) are all secondary grades.
functions. However, type-2 fuzzy sets that are characterized by Definition 3: The domain of a secondary membership func-
type-1 fuzzy membership functions are capable of modeling tion is called the primary membership of x. In formula (3), Jx
the four uncertainties, because their membership functions are is the primary membership of x . Uncertainty in the primary
themselves fuzzy. In the past decades, many researchers inves- memberships of a type-2 fuzzy set A consists of a bounded re-
tigated type-2 fuzzy sets and their properties, e.g., Dubois and gion that is called the footprint of uncertainty (F OU ), which is
Prade [19], Mendel [20], Karnik and Mendel [21], Mendel and the union of all primary memberships, i.e.
John [22], and so on. According to Liang and Mendel [23] and
F OU (A) ≡ Jx . (4)
Mendel et al. [24], most people only use interval type-2 fuzzy
sets in practical fields because of the computational complexity x∈U
in using general type-2 fuzzy sets. In the following sections, F OU (A) is also denoted as DA.

Motivated by the aforementioned accomplishments, in this
Let DA(x) ≡ Jx ∀x ∈ U . Therefore, formula (2) can be re-
paper, the concept of fuzzy rough sets proposed by Dubois and expressed as
Prade [3] is generalized to interval type-2 (IT2) fuzzy envi-
ronments. Section II of this paper recalls the preliminaries of =
A µÃ (x, u)/(x, u). (5)
type-2 fuzzy sets and fuzzy rough sets, while Section III gives
(x,u )∈D A
the definitions of IT2 fuzzy rough sets. Section IV proposes
For any A ∈ F̃ (U ), a lower and an upper membership func-
a method of attribution reduction based on the concept of IT2
fuzzy rough sets. Lastly, Section V draws the conclusions. Fur- tion are the two type-1 membership functions that are the bounds
for DA. The lower membership function (denoted as DA) is as-
thermore, the properties of IT2 fuzzy rough sets are given in the

sociated with the lower bound of DA, and the upper membership
Appendix.
function (denoted as DA) is associated with the upper bound of
DA.
Definition 4: If all the secondary grades of a type-2 fuzzy set
II. PRELIMINARIES are equal to 1, i.e. µ (x, u) = 1 ∀x ∈ U, ∀u ∈ Jx ⊆ I, then
A Ã
A. Type-2 Fuzzy Sets A is defined as an interval type-2 fuzzy set. In the following
In this section, the preliminaries of type-2 fuzzy sets given by sections, the class of all interval type-2 (IT2) fuzzy sets of the
Mendel [20], Karnik and Mendel [21], and Mendel and John [22] universe U is denoted as F̃IT2 (U ).
are briefly recalled, while some notations are modified. Fig. 1(a) illustrates two IT2 fuzzy sets Ñ and Z̃. The shaded
1) Definitions of Type-2 Fuzzy Sets: regions in Fig. 1(a) are F OU . Fig. 1(b) illustrates an example of
Definition 1: Let U be a finite and nonempty set, which secondary membership function µÑ (−0.3), whose correspond-
is referred to as the universe, I = [0, 1]. A type-2 fuzzy set, ing primary membership is [0.4, 0.6].
denoted as A, is characterized by a type-2 membership function For any X̃ ∈ F̃IT2 (U ) and x ∈ U, µX̃ (x) is an interval type-1
µÃ (x, u) : U × I → I, where x ∈ U and u ∈ Jx ⊆ I, i.e. (IT1) fuzzy set on I. Therefore, DX̃(x) can be expressed as a
range [lX̃ (x), rX̃ (x)] ⊆ I and DX̃ = x∈U DX̃(x). Accord-
ing to formula in [20, B-9], type-2 fuzzy set inclusion is defined
≡ {((x, u), µ (x, u)) | x ∈ U, u ∈ Jx ⊆ I}
A (1) as follows.
Ã
WU et al.: INTERVAL TYPE-2 FUZZY ROUGH SET MODEL FOR ATTRIBUTE REDUCTION 303
where ★ is the minimum or algebraic product t-norm

for fuzzy intersection, ∨ is the maximum operation, is
called the join operation, µA∪B(x), µÃ (x), and µB̃ (x) are
the secondary membership functions, and all are type-1
fuzzy sets.
b) Intersection of two type-2 fuzzy sets A∩B

µA∩B(x) ≡ [fx (u)★gx (w)]/(u ∧ w)
u ∈J xu w ∈J xw
≡ µÃ (x) µB̃ (x), x∈U (7)

where ∧ is the minimum operation, is called the meet oper-
ation, µA∩B(x), µÃ (x), and µB̃ (x) are the secondary member-
ship functions, and all are type-1 fuzzy sets.
3) Operations of IT2 Fuzzy Sets: Let U be a
nonempty universe, X̃1 , . . . , X̃n ∈ F̃IT2 (U ). For any
x ∈ U, µX̃ 1 (x), . . . , µX̃ n (x) are vertical slices and all are
IT1 fuzzy sets. For simplicity, let µX̃ 1 (x), . . . , µX̃ n (x) be
denoted as F1 , . . . , Fn . Correspondingly, their domains are
[l1 , r1 ], . . . , [ln , rn ] ⊆ I.
a) Union of IT2 fuzzy sets X̃1 ∪ . . . ∪ X̃n : The join ni=1 Fi
of n IT1 fuzzy sets F1 , . . . , Fn is an IT1 fuzzy set with do-
main [(l1 ∨ . . . ∨ ln ), (r1 ∨ . . . ∨ rn ], where ∨ is the max-
imum operation
µX̃ 1 ∪...∪X̃ n (x) ≡ F1 . . . Fn
Fig. 1. IT2 fuzzy sets corresponding to the information system given in Table I.
(a) Two IT2 fuzzy sets. (b) Secondary membership function µ Ñ (−0.3).
= 1/q, x ∈ U. (8)
q ∈[(l 1 ∨...∨l n ),(r 1 ∨...∨r n )]
Definition 5: For X̃, Ỹ ∈ F̃ (U ), X̃ ⊆ Ỹ ⇔ µX̃ (x) ≤ µỸ (x) b) Intersection of IT2 fuzzy set X̃1 ∩ . . . ∩ X̃n : The meet
∀x ∈ U . ni=1 Fi of n IT1 fuzzy sets F1 , . . . , Fn is an IT1 fuzzy
Proposition 1: For X̃, Ỹ ∈ F̃IT2 (U ), X̃ ⊆ Ỹ ⇔ DX̃ ⊆ set with domain [(l1 ★ . . . ★ln ), (r1 ★ . . . ★rn )], where ★
DỸ ⇔ µD X̃ (x) ≥ µD Ỹ (x) and µD X̃ (x) ≤ µD Ỹ (x), x ∈ U . denotes either the minimum or algebraic product t-norm.
Proof: For X̃, Ỹ ∈ F̃IT2 (U ) and x ∈ U, µX̃ (x) and µỸ (x) In the following sections, ★ is specified as the minimum
are IT1 fuzzy sets on I. Thus, µX̃ (x) ≤ µỸ (x) is equiv- t-norm
alent to DX̃(x) ⊆ DỸ (x). Therefore, X̃ ⊆ Ỹ ⇔ µX̃ (x) ≤
µỸ (x), x ∈ U ⇔ DX̃(x) ⊆ DỸ (x), x ∈ U ⇔ DX̃ ⊆ DỸ . µX̃ 1 ∩...∩X̃ n (x) ≡ F1 . . . Fn
On the other hand, DX̃ ⊆ DỸ ⇔ µD X̃ (x) ≥ µD Ỹ (x) and
µD X̃ (x) ≤ µD Ỹ (x), x ∈ U . = 1/q, x ∈ U. (9)
q ∈[(l 1 ∧...∧l n ),(r 1 ∧...∧r n )]
Therefore, X̃ ⊆ Ỹ ⇔ µD X̃ (x) ≥ µD Ỹ (x) and µD X̃ (x) ≤
µD Ỹ (x), x ∈ U . B. Fuzzy Rough Sets
2) Operations of Type-2 Fuzzy Sets: Let U be a nonempty
B ∈ F̃ (U ) The concepts of vagueness (for fuzzy sets) and indiscerni-
universe, A,
bility (for rough sets) are related but distinct concepts, both

of which are caused by uncertainties in knowledge or data.

A= µÃ (x)/x = fx (u)/u x, Jxu ⊆ I Fuzzy rough sets encapsulate these two concepts, and there-
x∈U x∈U u ∈J xu
fore, are adequate to deal with the uncertainties. Dubois and

B= µB̃ (x)/x = gx (w)/w x, Jx ⊆I. Prade [3] pioneered the concept of fuzzy rough sets. They
w
x∈U x∈U w ∈J xw constructed a pair of lower and upper approximation oper-
ators of fuzzy sets with respect to a fuzzy similarity rela-
The union and intersection for type-2 fuzzy sets are defined tion by using the t-norm Min and its dual conorm Max.
as follows: Here, the preliminaries of fuzzy rough sets proposed by
a) Union of two type-2 fuzzy sets A ∪B
Dubois and Prade [3] are briefly recalled with some modified
notations.
µA∪B(x) ≡ [fx (u)★gx (w)]/(u ∨ w) Definition 6: Let U be a nonempty universe, R be a fuzzy
u ∈J xu w ∈J xw
binary relation on U , and F(U ) be the fuzzy power set of U .
≡ µÃ (x) µB̃ (x), x ∈ U (6) For any A ∈ F(U ), a fuzzy rough set is a pair (R∗ (A), R∗ (A))
such that for every x ∈ U µR̃ ((x, y), t) denoted as µR̃ (x , y )

R∗ (A)(x) ≡ inf max{1 − R(x, y), µA (y)} µR̃ (x , y ) ≡ f(x ,y ) (t)/t. (14)
y ∈U
t∈J ( x , y )
= ∧ [(1 − R(x, y)) ∨ µA (y)] (10)
y ∈U The F OU of R̃ is denoted as
∗
R (A)(x) ≡ sup min{R(x, y), µA (y)} DR̃(x, y) ≡ J(x,y )
y ∈U

= ∨ [R(x, y) ∧ µA (y)]. (11) DR̃ ≡ F OU (R̃) ≡ DR̃(x, y). (15)
y ∈U
(x,y )∈U ×W
According to Dubois and Prade [3], a fuzzy rough set repre- Definition 8: Let U and W be two nonempty universes. An
sents the distortion of a fuzzy set A owing to the indiscernibility IT2 fuzzy set R̃ ∈ F̃IT2 (U × W ) is defined as an IT2 fuzzy
relation on U , as if looking at A with blurring glasses. For any relation from U to W
A, B ∈ F(U ) and x, y ∈ U , the approximation operators R∗
and R∗ have the following properties: R̃ ≡ 1/((x, y), t)
((x,y ),t)∈D R̃
(FR1) R∗ Ac = (R∗ A)c , R∗ Ac = (R∗ A)c
(FR2) R∗ (A ∩ B) = R∗ A∩R∗ B, R∗ (A∪B) = R∗ A ∪ R∗ B where
(FR3) R∗ A ⊆ A, A ⊆ R∗ A DR̃ ≡{((x, y), t) | (x, y) ∈ U ×W, t ∈ I, µR̃ ((x, y), t)=1}.
(FR4) R∗ (U − {y})(x) = R∗ (U − {x})(y),
R∗ x1 (y) = R∗ y1 (x) (16)
(FR5) R∗ U = U, R∗ ∅ = ∅ B
Definition 9: Let A, B
∈ F̃IT2 (W ), define A ⊆
if DA
(FR6) R∗ A ⊆ R∗ (R∗ A), R∗ (R∗ A) ⊆ R∗ A.
DB and DA ⊆ DB. If A B and B A, then A = B.
Here, x1 (y) = 1 for y = x and x1 (y) = 0 otherwise.
III. INTERVAL TYPE-2 FUZZY ROUGH SETS B. Definitions of IT2 Fuzzy Rough Sets
As mentioned in Section I, although fuzzy rough sets have Definition 10: Let R̃ ∈ F̃IT2 (U × W ), define a mapping f˜ :
been investigated fruitfully, so far, most of the studies in this field U → F̃IT2 (W ) by
are based on the concept of type-1 fuzzy sets, i.e., membership
µf˜(x) (y, v) ≡ µR̃ ((x, y), v), x ∈ U, y ∈ W, v ∈ I. (17)
functions are crisp. When problems to be solved contain uncer-
tain membership functions, type-2 fuzzy sets should be used; Hence, Df˜(x) = {(y, v) ∈ W × I|µf˜(x) (y, v) = 1}. The
hence, conventional fuzzy rough sets are infeasible. In this sec-
lower and upper membership functions of Df˜(x) are denoted
tion, the concept of fuzzy rough sets proposed by Dubois and
as Df˜(x), Df˜(x) ∈ F(W ), respectively.
Prade [3] will be generalized to interval type-2 fuzzy environ-
Since R̃ ∈ F̃IT2 (U × W ), for any x ∈ U and y ∈
ments. The properties of IT2 fuzzy rough sets are provided in
W, µR̃ (x, y) is an IT1 fuzzy set on I. Hence, DR̃(x, y) can
the Appendix.
be expressed as a range [lR̃ (x, y), rR̃ (x, y)] ⊆ I. Therefore, for
any x ∈ U, y ∈ W
A. Definitions of IT2 Fuzzy Relation
Definition 7: Let U and W be two nonempty universes. A µD f˜(x) (y) = lR̃ (x, y) µD f˜(x) (y) = rR̃ (x, y). (18)
type-2 fuzzy set R̃ ∈ F̃ (U × W ) is defined as a type-2 fuzzy ∈ F̃IT2 (W ), define
binary relation from U to W Definition 11: Let R̃ ∈ F̃IT2 (U × W ), A
lower and upper IT2 fuzzy rough approximation operators f̃ , f˜ :
R̃ ≡ {(((x, y), t), µR̃ ((x, y), t))|(x, y) ∈ U × W, F̃IT2 (W ) → F̃IT2 (U ) by
t ∈ J(x,y ) ⊆ I} (12)
˜
f (A) ≡ 1/(x, u) (19)
where 0 ≤ µR̃ ((x, y), t) ≤ 1. R̃ can also be expressed as )(x)
x∈U u ∈D f˜( A

≡
f˜(A) 1/(x, u) (20)
R̃ ≡ µR̃ ((x, y), t)/((x, y), t) )(x)
x∈U u ∈D f˜( A
(x,y )∈U ×W t∈J ( x , y )

where for any x ∈ U, Df˜(A)(x) ≡ [µf˜(D A) (x), µf˜(D A) (x)],
≡ f(x,y ) (t)/t (x, y) (13)
Df˜(A)(x) ≡ [µ ˜ (x), µ ˜ (x)], and
(x,y )∈U ×W t∈J ( x , y ) )
f (D A )
f (D A
where f(x,y ) (t) = µR̃ ((x, y), t). µf˜(D A) (x) ≡ inf max{1 − µD f˜(x) (y), µD A(y)}
y ∈W
At each value of (x, y), say (x, y) = (x , y ), the 2-D plane,
= ∧ [(1 − µD f˜(x) (y)) ∨ µD A(y)] (21)
whose axes are t and µR̃ ((x , y ), t), is a vertical slice of y ∈W
µf˜(D A) (x) ≡ inf max{1 − µD f˜(x) (y), µD A(y)} Df˜(a2 ) = 1.0/b1 + 0.6/b2 + 0.6/b3 + 0.8/b4
y ∈W
= ∧ [(1 − µD f˜(x) (y)) ∨ µD A(y)] (22) Df˜(a3 ) = 0.9/b1 + 0.5/b2 + 0.7/b3 + 0.8/b4 .
y ∈W
DA
DA, ∈ F(W ) are described as follows:
µ˜ (x) ≡ sup min{µD f˜(x) (y), µD A(y)}
f (D A ) y ∈W = 0.8/b1 + 0.4/b2 + 0.5/b3 + 0.7/b4
DA
= ∨ [µD f˜(x) (y) ∧ µD A(y)] (23) = 0.9/b1 + 0.5/b2 + 0.6/b3 + 0.8/b4 .
y ∈W DA
µ˜ ) (x) ≡ sup min{µD f˜(x) (y), µD A(y)}
For every x ∈ U, Df˜(A)(x),
Df˜(A)(x) can be calculated by
f (D A y ∈W
using Definition 11. For simplicity, at this juncture, only x = a1
= ∨ [µD f˜(x) (y) ∧ µD A(y)]. (24) will be considered
y ∈W
µf˜(D A) (a1 ) = ∧ [(1 − µD f˜(a 1 ) (y)) ∨ µD A(y)] = 0.5
f˜(A))
The pair (f˜(A), is defined as an interval type-2 fuzzy y ∈W
rough set. µf˜(D A) (a1 ) = ∧ [(1 − µD f˜(a 1 ) (y)) ∨ µD A(y)] = 0.6
Obviously, Df̃ (A) Df̃ (A)
= f̃ (DA), = f̃ (DA), and Df˜ y ∈W
= f˜(DA),
(A) Df˜(A) = f˜(DA). µ˜ ) (a1 ) = ∨ [µD f˜(a 1 ) (y) ∧ µD A(y)] = 0.8
f (D A y ∈W
Definition 12: Let U and W be two nonempty universes and
R̃ ∈ F̃IT2 (U × W ), then: µ˜ ) (a1 ) = ∨ [µD f˜(a 1 ) (y) ∧ µD A(y)] = 0.9.
f (D A y ∈W
1) if for any x ∈ U , there always exists y ∈ W such that
µR̃ ((x, y), 1) = 1, and µR̃ ((x, y), u) = 0 for 0 ≤ u < 1, Hence, Df˜(A)(a 1 ) = [0.5, 0.6], Df˜(A)(a
1 ) = [0.8, 0.9].
then R̃ is defined as a serial IT2 fuzzy relation from U to For all x ∈ U , the results are
W; 1 ) = [0.5, 0.6]
Df˜(A)(a 2 ) = {0.6}
Df˜(A)(a
2) if U = W , then R̃ is defined as an IT2 fuzzy relation on
U. 3 ) = [0.5, 0.6]
Df˜(A)(a
Definition 13: Let U be a nonempty universe and R̃ be an IT2
fuzzy relation on U , then: 1 ) = Df˜(A)(a
Df˜(A)(a 2 ) = Df˜(A)(a
3 ) = [0.8, 0.9].
1) if for any x ∈ U, µR̃ ((x, x), 1) = 1, and µR̃ ((x, x), u) =
This example is for the case of IT2 fuzzy rough sets. As a
0 for 0 ≤ u < 1, then R̃ is defined as a reflexive IT2 fuzzy comparison, the aforementioned IT2 fuzzy sets are reduced to
relation on U . type-1 fuzzy sets as follows.
2) if for any x, y ∈ U, J(x,y ) = J(y ,x) , and µR̃ ((x, y), u) = Let U = {a1 , a2 , a3 }, W = {b1 , b2 , b3 , b4 }, R ∈ F(U ×
µR̃ ((y, x), u) for all u ∈ J(x,y ) , then R̃ is defined as a W ), A ∈ F(W ). Let µR (ai , bj ) = R(i, j), i = 1 ∼ 3, j =
symmetric IT2 fuzzy relation on U . 1 ∼ 4. R and A are given as follows:
3) if for any x, y ∈ U, lR̃ (x, y) ≥ ∨z ∈U [lR̃ (x, z) ∧ lR̃ (z, y)]  
and rR̃ (x, y) ≥ ∨z ∈U [rR̃ (x, z) ∧ rR̃ (z, y)], then R̃ is de- 0.9 0.3 0.5 0.6
fined as a transitive IT2 fuzzy relation on U . R =  0.8 0.4 0.4 0.7 
0.8 0.4 0.5 0.6
C. Example A = 0.8/b1 + 0.4/b2 + 0.5/b3 + 0.7/b4 .
Let U = {a1 , a2 , a3 }, W = {b1 , b2 , b3 , b4 }, R̃ ∈ F̃IT2 (U × For every x ∈ U, R∗ (A)(x), R∗ (A)(x) can be calculated

W ), A ∈ F̃IT2 (W ). Let DR̃(ai , bj ) = [lR̃ (i, j), rR̃ (i, j)], DA by using Definition 6. For simplicity, here only x = a1 is
(bj ) = [lÃ (j), rÃ (j)], i = 1 ∼ 3, j = 1 ∼ 4. lR̃ , rR̃ , lÃ , rÃ are considered
given as follows:
    R∗ (A)(a1 ) = ∧ [(1 − R(a1 , y)) ∨ µA (y)] = 0.5
y ∈W
0.9 0.3 0.5 0.6 1.0 0.5 0.7 0.8
∗
lR̃ =  0.8 0.4 0.4 0.7  rR̃ =  1.0 0.6 0.6 0.8  R (A)(a1 ) = ∨ [R(a1 , y) ∧ µA (y)] = 0.8.
y ∈W
0.8 0.4 0.5 0.6 0.9 0.5 0.7 0.8
For all x ∈ U , the results are
lÃ = [ 0.8 0.4 0.5 0.7 ] rÃ = [ 0.9 0.5 0.6 0.8 ] .
R∗ (A) = 0.5/a1 + 0.6/a2 + 0.5/a3
Hence, for every x ∈ U, Df˜(x), Df˜(x) ∈ F(W ) are described
as follows: R∗ (A) = 0.8/a1 + 0.8/a2 + 0.8/a3 .
Df˜(a1 ) = 0.9/b1 + 0.3/b2 + 0.5/b3 + 0.6/b4 IV. IT2 FUZZY-ROUGH ATTRIBUTE REDUCTION
Df˜(a2 ) = 0.8/b1 + 0.4/b2 + 0.4/b3 + 0.7/b4 According to Jensen and Shen [4], rough set theory has a
significant advantage in finding a reduct within the original
Df˜(a3 ) = 0.8/b1 + 0.4/b2 + 0.5/b3 + 0.6/b4
attributes that are the most informative. However, traditional at-
Df˜(a1 ) = 1.0/b1 + 0.5/b2 + 0.7/b3 + 0.8/b4 tribute reduction methods can only be operated effectively with
the data sets containing discrete values. It is inadequate to deal The fuzzy-rough dependency function is defined by
with the case where values of attributes are real-valued. As most
|µPOS P (Q) (x)| µPOS P (Q) (x)
of the data sets contain real-valued attributes, it is necessary to γP (Q) = = x∈U . (30)
perform a discretization step beforehand. This is typically im- |U | |U |
plemented by standard fuzzification techniques. Nevertheless, For the case of multiple attributes, let P contain multiple at-
the membership degrees of attribute values to fuzzy sets are not tributes, e.g., P = {a, b}. In the crisp case, U/P contains sets
exploited in the attribute reduction process. of objects grouped together that are indiscernible according to
To overcome this shortcoming, Jensen and Shen [4] proposed both attributes a and b. In the fuzzy case, objects may belong
a fuzzy-rough attribute reduction algorithm. According to their to many fuzzy equivalence classes. Hence, the Cartesian prod-
research, the fuzzy-rough reduction is more powerful than the uct of U/IND({a}) and U/IND({b}) should be considered in
conventional rough-set-based approach. However, the fuzzy- determining U/P .
rough attribute reduction is based on the concept of the type-1 In general, U/P = {a ∈ P : U/IND({a})} and each set
fuzzy sets, and it cannot deal with the type-2 fuzzy cases. In this in U/P denotes a fuzzy equivalence class. For example,
section, the fuzzy-rough attribute reduction will be extended to if P = {a, b}, U/IND({a}) = {Na , Za } and U/IND({b}) =
IT2 fuzzy environments. The results of fuzzy-rough attribute {Nb , Zb }, then U/P = {Na ∩ Nb , Na ∩ Zb , Za ∩ Nb , Za ∩
reduction reported by Jensen and Shen [4] are recalled at the Zb }. The extent to which an object belongs to a fuzzy equiva-
beginning. Subsequently, the definitions of IT2 fuzzy-rough at- lence class is calculated by using the conjunction of constituent
tribute reduction are introduced. Lastly, an example is given to fuzzy equivalence classes, e.g., Fi , i = 1, . . . , n
demonstrate IT2 fuzzy-rough attribute reduction.
µF 1 ∩...∩F n (x) = min(µF 1 (x), . . . , µF n (x)). (31)
A. Fuzzy-Rough Attribute Reduction
Jensen and Shen [4] noted that just as crisp equivalence B. IT2 Fuzzy-Rough Attribute Reduction
classes were central to rough sets, fuzzy equivalence classes Similar to the fuzzy-rough attribute reduction, IT2 fuzzy
were central to the fuzzy rough set approach. For typical appli- equivalence classes will be central to the IT2 fuzzy-rough at-
cations, this means that decision and conditional values may all tribute reduction under IT2 fuzzy environments. The family of
be fuzzy. According to Dubois and Prade [3], the family of nor- IT2 fuzzy sets produced by an IT2 fuzzy partitioning of the
mal fuzzy sets produced by a fuzzy partitioning of the universe universe of discourse can play the role of IT2 fuzzy equivalence
of discourse can play the role of fuzzy equivalence classes. classes. Fig. 1(a) illustrates two IT2 fuzzy equivalence classes,
Let P be a subset of the attribute set A and U/P = i.e., Ñ (denoting the word “Negative”) and Z̃ (denoting the
{F1 , . . . , Fk } be the fuzzy partition of U , then the fuzzy P - word “Zero”).
lower and P -upper approximations are defined by Dubois and Definition 14: Let P be a subset of attribute set A, U/P =
Prade [3] as {F̃1 , . . . , F̃k } be the IT2 fuzzy partition of U , and X̃ ∈ F̃IT2 (U )
µP X (Fi ) = inf max{1 − µF i (x), µX (x)} (25) be an IT2 fuzzy concept to be approximated. The IT2 fuzzy
x∈U P -lower and P -upper approximations are defined as
µP X (Fi ) = sup min{µF i (x), µX (x)} (26)
x∈U µP X̃ (F̃i ) ≡ 1/u
u ∈D P X̃ ( F̃ i )
where i = 1, . . . , k, X ∈ F(U ) is a fuzzy concept to be ap-
proximated, and Fi is a fuzzy equivalence class that belongs to
µP X̃ (F̃i ) ≡ 1/u, i = 1, . . . , k (32)
U/P . For the case of attribute reduction, the fuzzy P -lower and u ∈D P X̃ ( F̃ i )
P -upper approximations are redefined as
where
µP X (x) = sup min µF (x), inf max{1 − µF (y),
F ∈U /P y ∈U DP X̃(F̃i ) ≡ inf max{1 − µD F̃ i (x), µD X̃ (x)},
x∈U

µX (y)} (27) inf max{1 − µD F̃ i (x), µD X̃ (x)} (33)
x∈U

µP X (x) = sup min µF (x), sup min{µF (y), µX (y)} DP X̃(F̃i ) ≡ sup min{µD F̃ i (x), µD X̃ (x)}
F ∈U /P y ∈U
x∈U
(28)
sup min{µD F̃ i (x), µD X̃ (x)} . (34)
where x ∈ U . x∈U
Fuzzy-rough attribute reduction builds on the notion of fuzzy For the case of attribute reduction, the IT2 fuzzy P -lower and
lower approximation to enable reduction of data sets containing P -upper approximations are redefined as
real-valued attributes. Let P and Q be two subsets of the attribute
set A, then the membership of an object x ∈ U that belongs to µP X̃ (x) ≡ 1/u (35)
fuzzy positive region is defined by u ∈D P X̃ (x)

µPOS P (Q) (x) = sup µP X (x). (29) µP X̃ (x) ≡ 1/u, i = 1, . . . , k (36)
X ∈U /Q u ∈D P X̃ (x)
where TABLE I
EXAMPLE OF INFORMATION SYSTEM

DP X̃(x) ≡ sup min{µD F̃ (x), inf max{1 − µD F̃ (y),
y ∈U
F̃ ∈U /P
µD X̃ (y)}}, sup min{µD F̃ (x),

F̃ ∈U /P

inf max{1 − µD F̃ (y)µD X̃ (y)}} , x∈U QuickReduct algorithm, an IT2 fuzzy-rough QuickReduct al-
y ∈U
gorithm is proposed as follows:
DP X̃(x) ≡ sup min{µD F̃ (x), sup min{µD F̃ (y), 1. R ← { }, γ b est ← 0, γ b est ← 0, γ prev ← 0, γ prev ← 0
F̃ ∈U /P y ∈U
µD X̃ (y)}}, sup min{µD F̃ (x), 2. do

F̃ ∈U /P 3. T ←R

sup min{µD F̃ (y), µD X̃ (y)}} , x ∈ U. 4. γ prev ← γ b est , γ prev ← γ b est
y ∈U
5. for each θ ∈ (C − R)
Just as the fuzzy-rough attribute reduction builds on the no- 6. if γ R ∪{θ } (Q) > γ T (Q)
tion of fuzzy lower approximation, the IT2 fuzzy-rough attribute
reduction also builds on the notion of IT2 fuzzy lower ap- 7. T ← R ∪ {θ}
proximation to enable reduction of data sets under IT2 fuzzy 8. γ b est ← γ T (Q)
environments.
Definition 15: Let P and Q be two subsets of the attribute 9. R←T
set A. The membership of an object x ∈ U that belongs to IT2 10. for each θ ∈ (C − R)
fuzzy positive region is defined by
11. if γ R ∪{θ } (Q) > γ T (Q)
µ̃POS P (Q) (x) ≡ X̃ ∈U /Q µP X̃ (x), x ∈ U. (37) 12. T ← R ∪ {θ}

13. γ b est ← γ T (Q)
The lower and upper membership functions of µ̃POS P (Q) (x) 14. R←T
are denoted as µ̃POS (Q) (x) and µ̃POS P (Q) (x). The IT2 fuzzy- 15. until γ b est ≤ γ prev or γ b est ≤ γ prev
P
rough dependency functions are defined by
Here, C is a set of conditional attributes and Q is a set of decision
|µ̃POS attributes. The IT2 fuzzy-rough dependency functions γ and γ
(Q )
(x)| |µ̃POS P (Q ) (x)|
γ P (Q) ≡ P
γ P (Q) ≡ . (38) are employed to choose the attributes that can be added to the
|U | |U | current reduct candidate. The algorithm terminates when the
addition of any remaining attribute does not increase the IT2
For the case of multiple attributes, the Cartesian product of fuzzy-rough dependency degrees.
IT2 fuzzy equivalence classes should be considered. In general, D. IT2 Fuzzy-Rough Attribute Reduct Example
U/P = {a ∈ P : U/IND({a})}, where each set in U/P de-
notes an IT2 fuzzy equivalence class. The extent to which an Following the example by Jensen and Shen [4], an informa-
object belongs to an IT2 fuzzy equivalence class is calculated tion system IS = (U, C ∪ Q) defined in Table I is considered,
by using the meet of constituent IT2 fuzzy equivalence classes, where U = {u1 , . . . , u6 } is a set of finite objects, A = C ∪ Q
e.g., F̃i , i = 1, . . . , n is a set of attributes such that for every a ∈ A, a : U → Va , Va
being the value set of attribute a, C = {c1 , c2 , c3 } is a set of con-
ditional attributes, and Q = {d} is a set of decision attributes.
µF̃ 1 ∩...∩F̃ n (x) ≡ ni=1 µF̃ i (x). (39) For each object ui ∈ U (i = 1, . . . , 6) and each conditional
attribute cj ∈ C (j = 1, 2, 3), cj (ui ) is a real value, which can be
viewed as an objective measured data. For the decision attribute,
C. IT2 Fuzzy-Rough Attribute Reduct Computation
the corresponding result for each object is a range belonging to
In the conventional fuzzy-rough QuickReduct algorithm in I. Here, a range, instead of a crisp value, is used to describe
[4], the dependency function γ is employed to choose the at- the extent to which one likes or dislikes a certain object. For
tributes that can be added to the current reduct candidate. The al- example, range [0, 0.5) stands for fuzzy judgment “No,” range
gorithm terminates when the addition of any remaining attribute (0.5,1] stands for fuzzy judgment “Yes,” where 0 means totally
does not increase the dependency. Similar to the fuzzy-rough “No” and 1 means totally “Yes.” Therefore, with respect to
the decision attribute set Q = {d}, there are two IT2 fuzzy Thus, µD {c 1 }Jñ o (u1 ) = 0.4 ∨ 0 = 0.4.
equivalence classes, i.e. Hence
min{µD Ñ c (u1 ), inf max{1 − µD Ñ c (y), µD Jñ o (y)}}

U/Q = {Jño , J˜yes }. 1 y ∈U 1
Here = min{0.6, inf{0.4 ∨ 0.1, 0.4 ∨ 0, 0.6 ∨ 0.2, 1 ∨ 0,

DJño = 0/u1 + 0/u2 + 0.1/u3 + 0/u4 + 0/u5 + 0.3/u6 1 ∨ 0, 1 ∨ 0.4}}
DJño = 0.1/u1 + 0/u2 + 0.2/u3 + 0/u4 + 0/u5 + 0.4/u6 = min{0.6, 0.4} = 0.4,
DJ˜yes = 0/u1 + 0.5/u2 + 0/u3 + 0.7/u4 + 0.6/u5 + 0/u6 min{µD Z̃ c (u1 ), inf max{1 − µD Z̃ c (y), µD Jñ o (y)}}
1 y ∈U 1
DJ˜yes = 0/u1 + 0.6/u2 + 0/u3 + 0.8/u4 + 0.7/u5 + 0/u6 . = min{0, inf{1 ∨ 0.1, 1 ∨ 0, 0.75 ∨ 0.2, 0.75 ∨ 0,
Using the IT2 fuzzy sets defined in Fig. 1(a) (for 0.5 ∨ 0, 0.5 ∨ 0.4}}
all conditional attributes), the following IT2 fuzzy equiva-
lence classes are obtained: U/{c1 } = {Ñc 1 , Z̃c 1 }, U/{c2 } = = min{0, 0.5} = 0.
{Ñc 2 , Z̃c 2 }, U/{c3 } = {Ñc 3 , Z̃c 3 }. For simplicity, only the
lower membership functions of Ñc 1 and Z̃c 1 are listed Thus, µD {c 1 }Jñ o (u1 ) = 0.4 ∨ 0 = 0.4.
Therefore, D{c1 }Jño (u1 ) = [µ D {c 1 }Jñ o (u1 ), µD {c 1 }Jñ o
DÑc 1 = 0.6/u1 + 0.6/u2 + 0.4/u3 + 0/u4
(u1 )]= {0.4}.
+ 0/u5 + 0/u6 Calculating the IT2 fuzzy {c1 }-lower approximation of Jño
DZ̃c 1 = 0/u1 + 0/u2 + 0.25/u3 + 0.25/u4 for every object gives
+ 0.5/u5 + 0.5/u6 . D{c1 }Jño (u1 ) = {0.4} D{c1 }Jño (u2 ) = {0.4}
The first step of the IT2 fuzzy-rough attribute reduction is
D{c1 }Jño (u3 ) = {0.4} D{c1 }Jño (u4 ) = {0.25}
to calculate IT2 fuzzy lower approximations of {c1 }, {c2 }, and
{c3 }. For simplicity, only {c1 } will be considered, i.e., using D{c1 }Jño (u5 ) = {0.5} D{c1 }Jño (u6 ) = {0.5}.
{c1 } to approximate Q.
For the first IT2 fuzzy equivalence class Jño , µ{c 1 }Jñ o (x) = The corresponding values for J˜yes can also be calculated as

u ∈D {c 1 }Jñ o (x) 1/u, x ∈ U , where
follows:
D{c1 }Jño (x) = [µD {c 1 }Jñ o (x), µD {c 1 }Jñ o (x)], x∈U D{c1 }J˜yes (u1 ) = {0.4} D{c1 }J˜yes (u2 ) = {0.4}
µD {c 1 }Jñ o (x) = sup min{µD F̃ (x), D{c1 }J˜yes (u3 ) = {0.4} D{c1 }J˜yes (u4 ) = {0.25}
F̃ ∈U /{c 1 }
D{c1 }J˜yes (u5 ) = {0.5} D{c1 }J˜yes (u6 ) = {0.5}.
inf max{1 − µD F̃ (y), µD Jñ o (x)}}
y ∈U
Using these values, the membership of an object x ∈ U
µD {c 1 }Jñ o (x) = sup min{µD F̃ (x), belonging to IT2 fuzzy positive region can be calculated by
F̃ ∈U /{c 1 }
using
inf max{1 − µD F̃ (y), µD Jñ o (x)}}.
y ∈U
µ̃POS {c 1 } (Q) (x) = X̃ ∈U /Q µ{c 1 }X̃ (x), x ∈ U.
Considering the IT2 fuzzy equivalence classes of {c1 } (i.e.,
Ñc 1 and Z̃c 1 ), for object u1 , the following hold: This results in
min{µD Ñ c (u1 ), inf max{1 − µD Ñ c (y), µD Jñ o (y)}}
1 y ∈U 1
µ̃POS (u1 ) = 0.4 µ̃POS {c (Q) (u1 ) = 0.4
{c 1 } (Q) 1}
= min{0.6, inf{0.4 ∨ 0, 0.4 ∨ 0, 0.6 ∨ 0.1, 1 ∨ 0,
µ̃POS (u2 ) = 0.4 µ̃POS {c (Q) (u2 ) = 0.4
1 ∨ 0, 1 ∨ 0.3}} {c 1 } (Q) 1}
= min{0.6, 0.4} = 0.4, µ̃POS (u3 ) = 0.4 µ̃POS {c (Q) (u3 ) = 0.4
{c 1 } (Q) 1}
min{µD Z̃ c (u1 ), inf max{1 − µD Z̃ c (y), µD Jñ o (y)}} µ̃POS (u4 ) = 0.25 µ̃POS {c (Q) (u4 ) = 0.25
1 y ∈U 1 {c 1 } (Q) 1}
= min{0, inf{1 ∨ 0, 1 ∨ 0, 0.75 ∨ 0.1, 0.75 ∨ 0, µ̃POS (u5 ) = 0.5 µ̃POS {c (Q) (u5 ) = 0.5
{c 1 } (Q) 1}
0.5 ∨ 0, 0.5 ∨ 0.3}}

µ̃POS (u6 ) = 0.5 µ̃POS {c (Q) (u6 ) = 0.5.
{c 1 } (Q) 1}
= min{0, 0.5} = 0.
Therefore, γ {c } (Q) = γ {c 1 } (Q) = 2.45/6. Calculations for complexity of calculating the Cartesian product of fuzzy equiv-
1
{c2 } and {c3 } gives alence classes becomes unacceptable for large-feature subsets.
In some situations, the fuzzy lower approximation might not be
γ {c } (Q) = 2.2/6 γ {c 2 } (Q) = 2.4/6
2 a subset of the fuzzy upper approximation. To overcome these
γ {c } (Q) = 0.4/6 γ {c 3 } (Q) = 0.4/6. shortcomings, Jensen and Shen [6] proposed three techniques
3
based on fuzzy lower approximation, fuzzy boundary region,
It can be observed that attribute c1 will cause the greatest and fuzzy discernibility matrix. In the future research, it might
increase in IT2 fuzzy-rough dependency degrees. Hence, this be useful to combine the model of IT2 fuzzy rough sets with
attribute is chosen and added to the potential reduct set. The these three techniques as well. In this way, the Cartesian product
process iterates and produces two IT2 fuzzy-rough dependency of the IT2 fuzzy equivalence classes might no longer be used,
degrees as follows: and the explosive growth in the number of the considered IT2
γ {c (Q) = 2.45/6 γ {c 1 ,c 2 } (Q) = 2.45/6 fuzzy equivalence classes might be avoided.
1 ,c 2 }
On the other hand, there are at least two approaches for the
γ {c (Q) = 2.55/6 γ {c 1 ,c 3 } (Q) = 2.65/6. development of the fuzzy rough set theory—constructive and
1 ,c 3 }
axiomatic [10]. Since this paper focuses on the constructive
Since a larger increase of IT2 fuzzy-rough dependency de-
approach, in the future, it might be useful to investigate the IT2
grees is produced by adding attribute c3 to the reduct candi-
fuzzy rough sets using the axiomatic approach, i.e., using the
date set, the new reduct candidate set becomes {c1 , c3 }. Lastly,
lower and upper approximation operators as primitive notions.
adding c2 to the potential reduct set gives
γ {c (Q) = 2.45/6 γ {c 1 ,c 2 ,c 3 } (Q) = 2.45/6.
1 ,c 2 ,c 3 } APPENDIX
As the IT2 fuzzy-rough dependency degrees will not be in- ∈ F̃IT2 (W ), the following hold:
Proposition 2: For any A, B
creased by adding c2 to the potential reduct set, the algorithm
stops and outputs the IT2 fuzzy-rough reduct {c1 , c3 }. ∪ B)
1) D(A ⊆ DA
∪ DB;

However, for the case of fuzzy-rough attribute reduction
∩ B)
2) D(A ⊇ DA
∩ DB;

in [4], the reduct is {c1 , c2 }, and γ{c 1 ,c 2 }
= 3.4/6. This ex-
ample shows that IT2 fuzzy-rough attribute reduction generates A
3) A ∪ B;

different results when compared with conventional fuzzy-rough
attribute reduction, where: ∩B
4) A A.

1) the dependency degrees for the case of the IT2 fuzzy-
rough attribute reduction are decreased from γ = 3.4/6 Proof:

1) For any y ∈ W, DA(y)
= [lÃ (y), rÃ (y)], DB(y) =
to γ = 2.55/6, γ = 2.65/6;
[lB̃ (y), rB̃ (y)]. Thus, ∪ B)(y)
D(A = [lÃ (y) ∨ lB̃ (y),
2) the reduct set is changed from {c1 , c2 } to {c1 , c3 };
3) lower IT2 fuzzy-rough dependency degrees may be pro- rÃ (y) ∨ rB̃ (y)]
duced by using larger conditional attribute sets. For
∪ B)(y)
∀u ∈ D(A
example γ {c ,c ,c } (Q) < γ {c ,c } (Q), γ {c 1 ,c 2 ,c 3 } (Q) <
1 2 3 1 3
γ {c 1 ,c 3 } (Q). ⇒ u ≥ lÃ (y) ∨ lB̃ (y), u ≤ rÃ (y) ∨ rB̃ (y)
⇒ u ≥ lÃ (y) ∨ lB̃ (y), u ≤ rÃ (y), or
V. CONCLUSION
u ≥ lÃ (y) ∨ lB̃ (y), u ≤ rB̃ (y)
This paper has presented a detailed investigation on the prop-
erties of IT2 fuzzy rough sets, including proposed definitions ⇒ u ≥ lÃ (y), u ≤ rÃ (y), or u ≥ lB̃ (y), u ≤ rB̃ (y)
for IT2 fuzzy rough sets as well as a method of attribute re-

⇒ u ∈ DA(y),
or u ∈ DB(y)
duction within the IT2 fuzzy rough set framework. This study
differs from [7] in two aspects: 1) the starting point here is a ∪ DB(y).
⇒ u ∈ DA(y)
general IT2 fuzzy binary relation, whereas the study by Jensen
and Shen [7] was based on an interval-valued fuzzy tolerance Thus, D(A ∪ B)(y)
⊆ DA(y) ∪ DB(y) ∀y ∈ W . There-
relation and 2) the definition of IT2 fuzzy rough sets in this fore, D(A ∪ B) ⊆ DA ∪ DB.
study is different from that of the interval-valued fuzzy rough 2) The proof procedure is similar to 1).
sets in the work by Jensen and Shen [7]. 3) For any y ∈ W, DA(y) ∪ B)
= [lÃ (y), rÃ (y)], D(A
According to Klawonnm [25], transitivity may not be a de- (y) = [lÃ (y) ∨ lB̃ (y), rÃ (y) ∨ rB̃ (y)].
sirable property for a fuzzy similarity relation. Because the def- Thus, µD A(y) = lÃ (y) ≤ lÃ (y) ∨ lB̃ (y) = µD ( A∪B) (y).
inition of IT2 fuzzy rough sets is based on a general IT2 fuzzy Therefore
binary relation (see Definitions 7–11) instead of a fuzzy simi-
larity relation, Klawonnm’s result has no impact on this paper. DA ∪ B).
⊆ D(A (40)
Note that a considerable reference to [4] has been made in
this paper. Recently, Jensen and Shen [6] have reported that Similarly we can obtain DA ⊆ D(A
∪ B).
According to
there are several problems with respect to [4]. For example, the
Definition 9, there exists A A ∪ B.
4) For any y ∈ W, DA(y)=[l Proof:

Ã (y), rÃ (y)], D(A ∩ B)(y)=
∩ f̃ (B),
1) Let Ỹ = f̃ (A) then according to formula (9) and
[lÃ (y) ∧ lB̃ (y), rÃ (y) ∧ rB̃ (y)].
Thus, µD A(y) = lÃ (y) ≥ lÃ (y) ∧ lB̃ (y) = µD ( A∩B) (y). Definition 11, for any x ∈ U
Therefore DỸ (x) = [µf̃ (D A) (x) ∧ µf̃ (D B) (x), µf̃ (D A) (x)
DA ∩ B).
⊇ D(A (41) ∧ µf̃ (D B) (x)]
Similarly we can obtain DA ⊇ D(A ∩ B).
According to
⇒ DỸ (x) = µf̃ (D A) (x) ∧ µf̃ (D B) (x)
Definition 9, there exists A ∩ B A.
Proposition 3: Let R̃ ∈ F̃IT2 (U × W ), and f̃ and f˜ be the ∩ B)(x)
Df̃ (A = [µf̃ (D ( A∩B)) (x), µf̃ (D ( A∩B)) (x)]
lower and upper IT2 fuzzy rough approximation operators, for
B
any A, ∈ F̃IT2 (W ), the following hold: ∩ B)(x)

⇒ Df̃ (A = µf̃ (D ( A∩B)) (x).
⊆B
1) A ⇒ f̃ (A)
⊆ f̃ (B);

According to formula (41), there exists: DA ∩ B);
⊇ D(A
⊆B
2) A ⇒ f˜(A)
⊆ f˜(B).
hence, µD A(y) ≥ µD ( A∩B) (y), y ∈ W . Therefore
Proof: µf̃ (D A) (x) = ∧ [(1 − µD f˜(x) (y)) ∨ µD A(y)]

⊆B ⇒ µ (y) ≥ y ∈W
1) According to Proposition 1, A DA
µD B(y), µD A(y) ≤ µD B(y), y ∈ W . Therefore, for any ≥ ∧ [(1 − µD f˜(x) (y)) ∨ µD ( A∩B) (y)]
y ∈W
x∈U
= µf̃ (D ( A∩B)) (x).

∀u ∈ Df̃ (A)(x)
Similarly, the following holds:
⇒ u ∈ [µf̃ (D A) (x), µf̃ (D A) (x)]
µf̃ (D B) (x) ≥ µf̃ (D ( A∩B)) (x).
⇒ u ≥ µf̃ (D A) (x), u ≤ µf̃ (D A) (x)
Thus,
⇒ u ≥ ∧ [(1 − µD f˜(x) (y)) ∨ µD A(y)] µf̃ (D A) (x) ∧ µf̃ (D B) (x) ≥ µf̃ (D ( A∩B)) (x)
y ∈W
u ≤ ∧ [(1 − µD f˜(x) (y)) ∨ µD A(y)] ∩ B)(x)

⇒ DỸ (x) ≥ Df̃ (A ∀x ∈ U
y ∈W
∩ f̃ (B)]
⇒ D[f̃ (A) ⊇ Df̃ (A
∩ B).

⇒ u ≥ ∧ [(1 − µD f˜(x) (y)) ∨ µD B(y)]
y ∈W
Similarly, we can derive D[f̃ (A) ∩ f̃ (B)]
⊇ Df̃ (A ∩ B).

u ≤ ∧ [(1 − µD f˜(x) (y)) ∨ µD B(y)] Therefore, according to Definition 9, the following holds:
y ∈W
∩ B)
f̃ (A f̃ (A) ∩ f̃ (B).

⇒ u ≥ µf̃ (D B) (x) u ≤ µf̃ (D B) (x) 2) The proof procedure is similar to 1).
∪ f̃ (B),
3) Let Ỹ = f̃ (A) then according to formula (8) and

⇒ u ∈ Df̃ (B)(x). Definition 11, for any x ∈ U ,
Thus, DỸ (x) = [µf̃ (D A) (x) ∨ µf̃ (D B) (x), µf̃ (D A) (x)

Df̃ (A)(x)
⊆ Df̃ (B)(x) ∀x ∈ U ∨ µf̃ (D B) (x)]

=
⇒Df̃ (A)
Df̃ (A)(x) ⊆
Df̃ (B)(x)
= Df̃ (B)
x∈U x∈U
⇒ DỸ (x) = µf̃ (D A) (x) ∨ µf̃ (D B) (x)
⊆ f̃ (B)
⇒f̃ (A) (by Proposition 1). ∪ B)(x)
Df̃ (A = [µf̃ (D ( A∪B)) (x), µf̃ (D ( A∪B)) (x)]
2) The proof procedure is similar to 1). ∪ B)(x)

⇒ Df̃ (A = µf̃ (D ( A∪B)) (x).
˜
Proposition 4: Let R̃ ∈ F̃IT2 (U × W ), f̃ and f be lower and
upper IT2 fuzzy rough approximation operators, for any A, B∈ ⊆ D(A
According to formula (40), there exists DA ∪ B).

F̃IT2 (W ), the following hold: Hence, µD A(y) ≤ µD ( A∪B) (y), y ∈ W .
∩ B)
f̃ (A)
∩ f̃ (B);
Therefore
1) f̃ (A
µf̃ (D A) (x) = ∧ [(1 − µD f˜(x) (y)) ∨ µD A(y)]
∩ B)
2) f˜(A f˜(A)
∩ f˜(B);
y ∈W
∪ f̃ (B)
3) f̃ (A) f̃ (A
∪ B);
≤ ∧ [(1 − µD f˜(x) (y)) ∨ µD ( A∪B) (y)]
y ∈W
∪ f˜(B)
4) f˜(A) f˜(A
∪ B).
= µf̃ (D ( A∪B)) (x).
Similarly, the following holds: where

µf̃ (D B) (x) ≤ µf̃ (D ( A∪B)) (x). 1, y = x, v = 1
µ1̃ U \{x } (y, v) =
0, otherwise
Thus,
µf̃ (D A) (x) ∨ µf̃ (D B) (x) ≤ µf̃ (D ( A∪B)) (x) 1, y = x, v = 1
µ1̃ x (y, v) =
0, otherwise.
∪ B)(x)
⇒ DỸ (x) ≤ Df̃ (A ∀x ∈ U Proof:
1) According to the definitions of µ1̃ U \{x } (y, v) and µ1̃ x (y, v),
∪ f̃ (B)]
⇒ D[f̃ (A) ⊆ Df̃ (A
∪ B).

the following hold:
Similarly, we can derive D[f̃ (A) ∪ f̃ (B)]
⊆ Df̃ (A ∪ B).

1, y = x
Therefore, according to Definition 9, the following holds: µD 1̃ U \{x } (y) = µD 1̃ U \{x } (y) =
0, y = x
∪ f̃ (B)
f̃ (A) f̃ (A
∪ B).

1, y = x
µD 1̃ x (y) = µD 1̃ x (y) =
4) The proof procedure is similar to 3). 0, y = x.
Proposition 5: Let R̃ ∈ F̃IT2 (U × U ), and f̃ and f˜ be the Since R̃ is symmetric, then for any x, z ∈ U, J(x,z ) = J(z ,x) ,
lower and upper IT2 fuzzy rough approximation operators. If R̃ and µR̃ ((x, z), u) = µR̃ ((z, x), u) for all u ∈ J(x,z ) . Thus,
∈ F̃IT2 (U ), the following hold:
is reflexive, then for any A µf˜(x) (z, u) = µR̃ ((x, z), u) = µR̃ ((z, x), u) = µf˜(z ) (x, u), so
A;
µD f˜(x) (z) = µD f˜(z ) (x), µD f˜(x) (z) = µD f˜(z ) (x).
1) f˜(A)
2) According to Definition 11, for any x ∈ U
f˜(A).
2) A
Proof: Since R̃ is reflexive, then for any x ∈ U , there always Df̃ (1̃U \{z } )(x) = [µf̃ (D 1̃ U \{z } ) (x), µf̃ (D 1̃ U \{z } ) (x)],
hold that µR̃ ((x, x), 1) = 1, and µR̃ ((x, x), u) = 0 for 0 ≤ u <
1, i.e., µf˜(x) (x, 1) = 1, and µf˜(x) (x, u) = 0, for 0 ≤ u < 1. where
Thus, according to formula (18),
µf̃ (D 1̃ U \{z } ) (x) = ∧ [(1 − µD f˜(x) (y)) ∨ µD 1̃ U \{z } (y)]
y ∈U
µD f˜(x) (x) = µD f˜(x) (x) = 1.
= ∧ [(1 − µD f˜(x) (y)) ∨ µD 1̃ U \{z } (y)]
1) According to Definition 11, for any x ∈ U y ∈U,y = z
µD f˜( A) (x) = µf˜(D A) (x) ∧ [(1 − µD f˜(x) (z)) ∨ µD 1̃ U \{z } (z)]
= ∧ [(1 − µD f˜(x) (y)) ∨ 1]
= ∧ [(1 − µD f˜(x) (y)) ∨ µD A(y)] y ∈U,y = z
y ∈U
∧ [(1 − µD f˜(x) (z)) ∨ 0]
= ∧ [(1 − µD f˜(x) (y)) ∨ µD A(y)]
y ∈U,y = x
= 1 − µD f˜(x) (z).
∧ [(1 − µD f˜(x) (x)) ∨ µD A(x)]
Similarly, µf̃ (D 1̃ U \{z } ) (x) = 1 − µD f˜(x) (z). Hence, Df̃
= ∧ [(1 − µD f˜(x) (y)) ∨ µD A(y)]
y ∈U,y = x (1̃U \{z } )(x) = {1 − µD f˜(x) (z)}.
∧ [0 ∨ µD A(x)] Similarly, we can obtain
= ∧ [(1 − µD f˜(x) (y)) ∨ µD A(y)] ∧ µD A(x) Df̃ (1̃U \{x} )(z) = {1 − µD f˜(z ) (x)}.
y ∈U,y = x
According to 1),
≤ µD A(x).
µD f˜(x) (z) = µD f˜(z ) (x)
Thus, Df˜(A) ⊆ DA. Similarly we can derive Df˜(A) ⊆
˜ and thus
DA. Therefore, according to Definition 9, there exists f (A)

A. Df̃ (1̃U \{z } )(x) = Df̃ (1̃U \{x} )(z)∀x, z ∈ U.
2) The proof procedure is similar to 1).
Therefore,
Proposition 6: Let R̃ ∈ F̃IT2 (U × U ), and f̃ and f˜ be the
lower and upper IT2 fuzzy rough approximation operators. If R̃ µf̃ ( 1̃ U \{z } ) (x) = µf̃ ( 1̃ U \{x } ) (z)∀x, z ∈ U.
is symmetric, then for any x, z ∈ U , the following hold:
3) According to Definition 11, for any x ∈ U,
1) µD f˜(x) (z) = µD f˜(z ) (x), µD f˜(x) (z) = µD f˜(z ) (x);
2) µf˜( 1̃ U \{z } ) (x) = µf˜( 1̃ U \{x } ) (z); Df˜(1̃z )(x) = [µ ˜ (x), µ ˜ (x)]
f (D 1̃ z ) f (D 1̃ z )
3) µ ˜ (x) = µ ˜ (z)
f ( 1̃ z ) f ( 1̃ x )
where For any x ∈ U

µ˜ (x) = ∨ [µD f˜(x) (y) ∧ µD 1̃ z (y)] Df̃ (1̃W \{z } )(x) = [µf̃ (D 1̃ W \{z } ) (x), µf̃ (D 1̃ W \{z } ) (x)],
f (D 1̃ z ) y ∈U
= ∧ [µD f˜(x) (y) ∧ µD 1̃ z (y)] µf̃ (D 1̃ W \{z } ) (x) = ∧ [(1 − µD f˜(x) (y)) ∨ µD 1̃ W \{z } (y)]
y ∈W
y ∈U,y = z
∨ [µD f˜(x) (z) ∧ µD 1̃ z (z)] = ∧ [(1 − µD f˜(x) (y)) ∨ µD 1̃ W \{z } (y)]

y ∈W,y = z
= ∧ [µD f˜(x) (y) ∧ 0] ∨ [µD f˜(x) (z) ∧ 1] ∧ [(1 − µD f˜(x) (z)) ∨ µD 1̃ W \{z } (z)]
y ∈U,y = z
= µD f˜(x) (z). = ∧ [(1 − µD f˜(x) (y)) ∨ 1]

y ∈W,y = z
Similarly, ∧ [(1 − µD f˜(x) (z)) ∨ 0]

µ˜ (x) = µD f˜(x) (z). = 1 − µD f˜(x) (z) = 1 − lR̃ (x, z),
f (D 1̃ z )
Hence Similarly,
µf̃ (D 1̃ W \{z } ) (x) = 1 − µD f˜(x) (z) = 1 − lR̃ (x, z).
Df˜(1̃z )(x) = {µD f˜(x) (z)}. (42)
Therefore,
Similarly, we can obtain
Df̃ (1̃W \{z } )(x) = {1 − lR̃ (x, z)}.
Df˜(1̃x )(z) = {µD f˜(z ) (x)}.
3) For any x ∈ U
According to 1),
Df˜(1̃X )(x) = [µ ˜ (x), µ ˜ (x)]
f (D 1̃ X ) f (D 1̃ X )
µD f˜(x) (z) = µD f˜(z ) (x)
µ˜ (x) = ∨ [µD f˜(x) (y) ∧ µD 1̃ X (y)]
and thus f (D 1̃ X ) y ∈W
= ∨ µD f˜(x) (y) = ∨ rR̃ (x, y).

Df˜(1̃z )(x) = Df˜(1̃x )(z)∀x, z ∈ U. y ∈X y ∈X
Therefore, Similarly,
µ˜ (x) = ∨ rR̃ (x, y).
µ˜ (x) = µ ˜ (z)∀x, z ∈ U. f (D 1̃ X ) y ∈X
f ( 1̃ z ) f ( 1̃ x )
Therefore,
Proposition 7: Let R̃ ∈ F̃IT2 (U × W ). Then, for any x ∈ Df˜(1̃X )(x) = { ∨ rR̃ (x, y)}.
U, z ∈ W, X ⊆ W , the following hold: y ∈X
4) For any x ∈ U
1) Df˜(1̃z )(x) = {rR̃ (x, z)}; Df˜(1̃X )(x) = [µf˜(D 1̃ X ) (x), µf˜(D 1̃ X ) (x)]
2) Df̃ (1̃W \{z } )(x) = {1 − lR̃ (x, z)};
µf˜(D 1̃ X ) (x) = ∧ [(1 − µD f˜(x) (y)) ∨ µD 1̃ X (y)]
y ∈W
3) Df˜(1̃X )(x) = { ∨ rR̃ (x, y)};
y ∈X = ∧ [1 − µD f˜(x) (y)]
y∈
/X
4) Df˜(1̃X )(x) = { ∧ [1 − lR̃ (x, y)]}
y∈
/X = ∧ [1 − lR̃ (x, y)].
y∈
/X
where Similarly,

1, y ∈ X, v = 1 µf˜(D 1̃ X ) (x) = ∧ [1 − lR̃ (x, y)].
µ1̃ X (y, v) =
0, otherwise. y∈
/X
Proof: Therefore,
1) According to formula (42) and (18), for any x ∈ U, Df˜(1̃X )(x) = { ∧ [1 − lR̃ (x, y)]}.
y∈
/X
Df˜(1̃z )(x) = {µD f˜(x) (z)} = {rR̃ (x, z)}.
2) Proposition 8: Let R̃ ∈ F̃IT2 (U × W ). If R̃ is serial, the
following hold:
1, y ∈ X, v = 1
µ1̃ X (y, v) =
0, otherwise 1) f˜(1̃W ) = 1̃U ;

1, y ∈ X 2) f˜(0̃W ) = 0̃U ;
⇒ µD 1̃ X (y) = µD 1̃ X (y) =
0, y ∈
/ X.
f˜(A)
3) f˜(A)
where ∧ [(1 − µD f˜(x) (y )) ∨ µD A(y )]

1, y ∈ W, v = 0 = ∧ [(1 − µD f˜(x) (y)) ∨ µD A(y)]
µ0̃ W (y, v) = y ∈W,y = y
0, otherwise
∧ [0 ∨ µD A(y )]

1, y ∈ U, v = 0
µ0̃ U (y, v) = = ∧ [(1 − µD f˜(x) (y)) ∨ µD A(y)] ∧ µD A(y )
0, otherwise. y ∈W,y = y
Proof: ≤ µD A(y )
1) For any x ∈ U, R̃ is serial
µ ) (x) = µ ˜ ) (x)
D f˜( A f (D A

⇒ ∃y ∈ W, µR̃ ((x, y ), 1) = 1, and
= ∨ [µD f˜(x) (y) ∧ µD A(y)]
y ∈W
µR̃ ((x, y ), u) = 0, 0≤u<1
= ∨ [µD f˜(x) (y) ∧ µD A(y)]
⇒ ∃y ∈ W, lR̃ (x, y ) = rR̃ (x, y ) = 1 y ∈W,y = y
⇒ ∨ rR̃ (x, y) = 1. ∨ [µD f˜(x) (y ) ∧ µD A(y )]

y ∈W
According to Proposition 7, = ∨ [µD f˜(x) (y) ∧ µD A(y)] ∨ [1 ∧ µD A(y )]

y ∈W,y = y

Df˜(1̃W )(x) = ∨ rR̃ (x, y) = {1} ∀x ∈ U. = ∨ [µD f˜(x) (y) ∧ µD A(y)] ∨ µD A(y )
y ∈W y ∈W,y = y
≥ µD A(y ).
Therefore, f˜(1̃W ) = 1̃U .
2) Hence,
µD f˜( A) (x) ≤ µD A(y ) ≤ µ ˜ (x).
1, y ∈ W, v = 0 D f (A )
µ0̃ W (y, v) =
0, otherwise
Thus,
⇒ µD 0̃ W (y) = µD 0̃ W (y) = 0.
For any x ∈ U ⊆ Df˜(A).
Df˜(A)
µf˜(D 0̃ W ) (x) = ∧ [(1 − µD f˜(x) (y)) ∨ µD 0̃ W (y)] ⊆ Df˜(A). Therefore, ac-

y ∈W Similarly, we can derive Df˜(A)
= ∧ [1 − µD f˜(x) (y)] cording to Definition 9, f˜(A) f˜(A).

y ∈W
Proposition 9: Let R̃ ∈ F̃IT2 (U × U ), and f̃ and f˜ be the
= ∧ [1 − lR̃ (x, y)]. lower and upper type-2 fuzzy rough approximation operators. If
y ∈W
∈ F̃IT2 (U ), the following hold:
R̃ is transitive, then for any A
Since R̃ is serial, then ∃y ∈ W such that lR̃ (x, y ) = 1;
hence, ∧ [1 − lR̃ (x, y)] = 0, i.e., µf˜(D 0̃ W ) (x) = 0 ∀x ∈ U . f̃ (f̃ (A));
1) f̃ (A)
y ∈W
Similarly, we can obtain f˜(A).

2) f˜(f˜(A))
µf˜(D 0̃ W ) (x) = 0 ∀x ∈ U.
Proof: Since R̃ is transitive, then for any x, y, z ∈ U
Thus,
lR̃ (x, z) ≥ ∨ [lR̃ (x, y) ∧ lR̃ (y, z)]
y ∈U
Df˜(0̃W )(x) = [µf̃ (D 0̃ W ) (x), µf̃ (D 0̃ W ) (x)] = {0} ∀x ∈ U.
rR̃ (x, z) ≥ ∨ [rR̃ (x, z) ∧ rR̃ (y, z)].
Therefore, y ∈U
f˜(0̃W ) = 0̃U . According to formula (18)
3) R̃ is serial ⇒ ∃y ∈ W such that lR̃ (x, y ) = rR̃ (x, y ) =

1 ⇒ µD f˜(x) (y ) = µD f˜(x) (y ) = 1. µD f˜(x) (y) = lR̃ (x, y) µD f˜(x) (y) = rR̃ (x, y).
For any x ∈ U
Thus, for any x, z ∈ U ,
µD f˜( A) (x) = µf˜(D A) (x)
µD f˜(x) (z) ≥ ∨ [µD f˜(x) (y) ∧ µD f˜(y ) (z)] (43)
y ∈U
= ∧ [(1 − µD f˜(x) (y)) ∨ µD A(y)]
y ∈W
µD f˜(x) (z) ≥ ∨ [µD f˜(x) (y) ∧ µD f˜(y ) (z)]. (44)
y ∈U
= ∧
[(1 − µD f˜(x) (y)) ∨ µD A(y)]
y ∈W,y = y
1) For any x ∈ U , Thus,

µD f˜( f˜( A)) (x) µ ) (x) ≥ µ )) (x) ∀x ∈ U,
D f˜( A D f˜( f˜( A
= µf˜(D f˜( A)) (x) (by Definition 11) i.e.,

⊇ Df˜(f˜(A)).
Df˜(A)
= ∧ [(1 − µD f˜(x) (y)) ∨ µD f˜( A) (y)]
y ∈U
Similarly, we can derive
= ∧ [(1 − µD f˜(x) (y)) ∨ µf˜(D A) (y)] (by Definition 11)
y ∈U
⊇ Df˜(f˜(A)).
Df˜(A)
= ∧ {(1 − µD f˜(x) (y)) ∨ ∧ [(1 − µD f˜(y ) (z)) ∨ µD A(z)]}
y ∈U z ∈U
Therefore, according to Definition 9,
= ∧ ∧ {(1 − µD f˜(x) (y)) ∨ (1 − µD f˜(y ) (z)) ∨ µD A(z)}
y ∈U z ∈U f˜(A).
f˜(f˜(A))
= ∧ { ∧ [(1 − µD f˜(x) (y)) ∨ (1 − µD f˜(y ) (z))] ∨ µD A(z)}.
z ∈U y ∈U
On the other hand Proposition 10: Let R̃ ∈ F̃IT2 (U × U ), and f̃ and f˜ be the
lower and upper IT2 fuzzy rough approximation operators. If
µD f˜( A) (x) = µf˜(D A) (x) = ∧ [(1 − µD f˜(x) (z)) ∨ µD A(z)]. ∈ F̃IT2 (U ), the
R̃ is reflexive and transitive, then for any A
z ∈U
following hold:
According to formula (43), there exists
1 − µD f˜(x) (z)) ≤ ∧ [(1 − µD f˜(x) (y)) = f̃ (f̃ (A));

y ∈U 1) f̃ (A)
∨ (1 − µD f˜(y ) (z))] ∀z ∈ U. = f˜(A).
2) f˜(f˜(A))
Thus,
Proof: 1) Since R̃ is reflexive, then according to
µD f˜( A) (x) ≤ µD f˜( f˜( A)) (x) ∀x ∈ U f̃ (A).
Proposition 5, f̃ (f̃ (A)) Since R̃ is transitive, then ac-

cording to Proposition 9, f̃ (A) f̃ (f̃ (A)).
i.e.,
Therefore, according to Definition 9, the following holds
Df˜(A)
⊆ Df˜(f˜(A)). = f̃ (f̃ (A)).
f̃ (A)
2) The proof procedure is similar to 1).
Similarly we can derive
It can be seen that except (FR1), properties at the end of
⊆ Df̃ (f˜(A)).
Df˜(A) Section II have their IT2 fuzzy counterparts: (FR2)–(FR6) cor-
Therefore, according to Definition 9, respond to Propositions 4, 5, 6, 8 and 9, respectively.
f˜(f˜(A)).
f˜(A)
ACKNOWLEDGMENT
2) For any x ∈ U The authors wish to thank the anonymous referees for their
µ helpful comments. H. Wu is very grateful to F. Chen (Alice),
)) (x)
D f˜( f˜( A L. Chen, S. Chen, Z. Li, Wan-wan (Apple), Dou-dou (Lily), and
= µ˜ (x) (by Definition 11) Miao-miao (Cindy) for their great support.
))
f (D f˜( A
= ∨ [µD f˜(x) (y) ∧ µ ) (y)] REFERENCES

y ∈U D f˜( A
[1] Z. Pawlak, “Rough sets,” Int. J. Comput. Inform. Sci., vol. 11, pp. 341–
= ∨ [µD f˜(x) (y) ∧ µ ˜ ) (y)] (by Definition 11) 356, 1982.
y ∈U f (D A
[2] L. A. Zadeh, “Fuzzy sets,” Inform. Control., vol. 8, pp. 338–353, 1965.
= ∨ {µD f˜(x) (y) ∧ ∨ [µD f˜(y ) (z) ∧ µD A(z)]} [3] D. Dubois and H. Prade, “Rough fuzzy sets and fuzzy rough sets,” Int. J.
y ∈U z ∈U Gen. Syst., vol. 17, pp. 191–208, 1990.
[4] R. Jensen and Q. Shen, “Fuzzy-rough attribute reduction with application
= ∨ ∨ {µD f˜(x) (y) ∧ µD f˜(y ) (z) ∧ µD A(z)} to web categorization,” Fuzzy Sets Syst., vol. 141, pp. 469–485, 2004.
y ∈U z ∈U
[5] R. Jensen and Q. Shen, “Fuzzy-rough sets assisted attribute selection,”
= ∨ { ∨ [µD f˜(x) (y) ∧ µD f˜(y ) (z)] ∧ µD A(z)}. IEEE Trans. Fuzzy Syst., vol. 15, no. 1, pp. 73–89, Feb. 2007.
z ∈U y ∈U [6] R. Jensen and Q. Shen, “New approaches to fuzzy-rough feature selec-
tion,” IEEE Trans. Fuzzy Syst., to be published.
On the other hand [7] R. Jensen and Q. Shen, “Interval-valued fuzzy-rough feature selection and
application for handling missing values in datasets,” in Proc. 8th Annu.
U.K. Workshop Comput. Intell. (UKCI), 2008.
µ ) (x) = µ˜ ) (x) = ∨ [µD f˜(x) (z) ∧ µD A(z)]. [8] Q. Shen and R. Jensen, “Selecting informative features with fuzzy-rough
D f˜( A f (D A z ∈U sets and its application for complex systems monitoring,” Pattern Recog-
nit., vol. 37, pp. 1351–1363, 2004.
According to formula (44), [9] E. C. C. Tsang, D. Chen, D. S. Yeung, X. Z. Wang, and J. W. T. Lee,
“Attributes reduction using fuzzy rough sets,” IEEE Trans. Fuzzy Syst.,
µD f˜(x) (z)) ≥ ∨ [µD f˜(x) (y) ∧ µD f˜(y ) (z))] ∀z ∈ U. vol. 16, no. 5, pp. 1130–1141, Oct. 2008.
y ∈U
[10] D. S. Yeung, D. Chen, E. C. C. Tsang, J. W. T. Lee, and X. Z. Wang, “On Haoyang Wu received the B.S. degree in computer
the generalization of fuzzy rough sets,” IEEE Trans. Fuzzy Syst., vol. 13, science and the Ph.D. degree in electronic engineer-
no. 3, pp. 343–361, Jun. 2005. ing from Xi’an Jiaotong University, Xi’an, China, in
[11] M. Liu, D. Chen, C. Wu, and H. Li, “Fuzzy reasoning based on a new 1994 and 1999, respectively.
fuzzy rough set and its application to scheduling problems,” Comput. From 2000 to 2002, he was a Postdoctoral Re-
Math. Appl., vol. 51, pp. 1507–1518, 2006. searcher at Tsinghua University, Beijing, China. Cur-
[12] W. Z. Wu and W. X. Zhang, “Constructive and axiomatic approaches of rently he is a Visiting Researcher at Xi’an Jiaotong
fuzzy approximation operators,” Inf. Sci., vol. 159, pp. 233–254, 2004. University. His current research interests include ar-
[13] P. Lingras and R. Jensen, “Survey of rough and fuzzy hybridization,” in tificial intelligence, data mining, game theory, and
Proc. 16th Int. Conf. Fuzzy Syst. (FUZZ IEEE 2007), pp. 125–130. quantum computation.
[14] H. Wu, Y. Wu, H. Liu, and H. Zhang, “Roughness of type-2 fuzzy sets
based on similarity relations,” Int. J. Uncertainty, Fuzziness Knowl. Based
Syst., vol. 15, pp. 513–517, 2007.
[15] L. H. Tsoukalas and R. E. Uhrig, Fuzzy and Neural Approaches in Engi- Yuyuan Wu received the B.S. degree in power en-
neering. New York: Wiley, 1996. gineering from Xi’an Jiaotong University, Xi’ian,
[16] L. X. Wang, Adaptive Fuzzy Systems and Control: Design and Stability China, in 1966, and the Ph.D. degree in cryogenics
Analysis. Upper Saddle River, NJ: Prentice-Hall, 1994. from the University of Southampton, Southampton,
[17] A. Skowron and J. Stepaniuk, “Tolerance approximation spaces,” Fun- U.K., in 1984.
dam. Inform., vol. 27, pp. 245–253, 1996. Currently he is a Full Professor at Xi’an Jiaotong
[18] W. Ziarko, “Variable precision rough set model,” J. Comput. Syst. Sci., University. His current research interests include air
vol. 46, pp. 39–59, 1993. conditioning and cryogenics.
[19] D. Dubois and H. Prade, Fuzzy Sets and Systems: Theory and Applications.
New York: Academic, 1980.
[20] J. M. Mendel, Uncertain Rule-Based Fuzzy Logic Systems: Introduction
and New Directions. Upper Saddle River, NJ: Prentice-Hall, 2001.
[21] N. N. Karnik and J .M. Mendel, “Operations on type-2 fuzzy sets,” Fuzzy
Jinping Luo received the B.S. degree in computer
Sets Syst., vol. 122, pp. 327–348, 2001.
science from Airforce Engineering College, Xi’an,
[22] J. M. Mendel and R. I. John, “Type-2 fuzzy sets made simple,” IEEE
Shaanxi, China, in 1990, and the Ph.D. degree in
Trans. Fuzzy Syst., vol. 10, no. 2, pp. 117–127, Apr. 2002.
Computer Engineering from the National University
[23] Q. Liang and J. M. Mendel, “Interval type-2 fuzzy logic systems: Theory
of Defense Technology, Changsha, Hunan, China, in
and design,” IEEE Trans. Fuzzy Syst., vol. 8, no. 5, pp. 535–550, Oct.
2000.
2000.
Currently he is with Kunlun Technology Indus-
[24] J. M. Mendel, R. I. John, and F. Liu, “Interval type-2 fuzzy logic systems
try Corporation (a software company), Hangzhou,
made simple,” IEEE Trans. Fuzzy Syst., vol. 14, no. 6, pp. 808–821, Dec.
China. His current research interests include dis-
2006.
tributed computing and E-government.
[25] F. Klawonnm, “Should fuzzy equality and similarity satisfy transitivity?
Comments on the paper by M. De Cock and E. Kerre,” Fuzzy Sets Syst.,
vol. 133, pp. 175–180, 2003.
Theory of Extended Fuzzy Discrete-Event Systems

for Handling Ranges of Knowledge Uncertainties
and Subjectivity
Xinyu Du, Student Member, IEEE, Hao Ying, Senior Member, IEEE, and Feng Lin, Senior Member, IEEE
Abstract—In 2001, we originated a theory of fuzzy discrete- velopment of DES theory is a recent endeavor. Only after the
event systems (FDESs) that generalized the conventional/crisp proliferation of complex systems such as computer systems and
discrete-event systems (DESs). Vagueness and imprecision con- networks, have DESs been systematically studied. DES theory
cerning states and event transitions of DESs were represented by
membership grades and computed via fuzzy logic. Our application addresses the issues of modeling, control, and optimization of
of the FDES theory to computerized human immunodeficiency DES [1]–[3]. It has been applied to practical systems such as au-
virus/acquired immune deficiency syndrome treatment regimen tomated manufacturing systems, database concurrency control,
selection, although preliminarily successful, suggests that a more feature interactions in telecommunication networks, protocol
comprehensive FDES theory is needed to address two general is- verification and synthesis in communication networks, and pro-
sues critically important not only to biomedical applications, but
also to real-world problems in other industries. First, domain ex- tocol conversion and gateway synthesis in computer networks.
perts should have means other than point estimates and type-1 Although seemingly sufficient for many application domains,
fuzzy sets mandated in the current framework to describe uncer- crisp DES theory is not adequate for some important fields such
tainties, subjectivity, and imprecision in their (complex) knowledge as biomedicine in which the state and state transition of a system
and experience. Second, when a group of experts with distinct opin- (e.g., a person’s health status) are always somewhat uncertain
ions is involved, they should not be forced to reach consensus for
the sake of system development. This is because collective consen- and vague even in a deterministic sense. Subjective human ob-
sus may not be achievable, which is often the case in medicine, servation, judgment, and interpretation (e.g., by a physician or a
where individual experts’ opinions should be equally respected patient) invariably play a significant role in describing the status
since the underlying ground truth is unknown most of the time. The of state, usually not crisp.
theory of extended FDES presented in this paper addresses both the Definitions of fuzzy state, fuzzy event, and fuzzy automata
problems and contains the FDES theory as a special case. Experts
are now allowed to use interval numbers and type-1 and type-2 were proposed and studied as early as in the late 1960s [4], [5].
fuzzy sets to intuitively and quantitatively express their diverse However, only a few studies followed in the next 30 years or
knowledge and experience, which will then be processed by the new so [6]–[8]. A comprehensive theory of fuzzy DESs (FDESs)
theory to form fuzzy state vectors and fuzzy event transition ma- was yet to be established, so were many fundamentally impor-
trices. Accordingly, we have established mathematical operations tant concepts, methods, and theorems in the traditional DES,
that cover the computations of fuzzy states, fuzzy event transitions,
and parallel composition. Numerical examples are provided. such as controllability, observability, and optimal control. To
effectively represent deterministic uncertainties and vagueness
Index Terms—Automata, discrete-event systems (DESs), fuzzy as well as human subjective observation and judgment encoun-
logic, type-2 fuzzy systems.
tered in many real-world problems especially those in medicine,
I. INTRODUCTION we recently originated a more comprehensive theory of FDES
[9], [10]. We introduced fuzzy states and fuzzy event transition
HERE exist countless (complex) systems that cannot be
T effectively described, at a higher level, by differential
equations, but can be by traces (or sequences) of events that
and generalized conventional/crisp DES to FDES. The largely
graph-based framework of DES was unsuitable for the expan-
sion, and we thus reformulated it using state vectors and event
record significant qualitative changes in the state of the sys-
transition matrices that could be extended to fuzzy vectors and
tem. These states are logical or symbolic rather than numeri-
matrices by allowing their elements to take values in [0, 1].
cal. These systems can be described as discrete-event systems
We also extended optimal control and observability of DES to
(DESs) whose behavior consists of sequence of occurrences of
FDES. The new observability allows one to determine whether
distinct events. These events, for instance, can describe send-
or not the system output observed is sufficient for decision mak-
ing or receiving messages in computer networks, or processing
ing. The FDES theory is consistent with the DES theory, both
a part in a manufacturing plant. Comprehensive study and de-
at conceptual and computation levels, in that the former con-
tains the latter as a special case when the membership grades
Manuscript received September 7, 2007; revised April 18, 2008 and October are either 0 or 1. Other researchers have expanded different as-
2, 2008; accepted December 4, 2008. First published December 22, 2008; current
version published April 1, 2009.
pects of the FDES theory [11]–[20]. Other independent studies
The authors are with the Department of Electrical and Computer Engineering, include [21] and [22]. More recently, we further developed the
Wayne State University, Detroit, MI 48202 USA (e-mail: as7682@wayne.edu; FDES theory so that it possessed self-learning capability [23].
hao.ying@wayne.edu; flin@indigo.eng.wayne.edu).
As the first application, we have successfully applied the
at http://ieeexplore.ieee.org. FDES theory to develop a novel system that prescribes drugs
Digital Object Identifier 10.1109/TFUZZ.2008.2011279 for human immunodeficiency virus (HIV)/acquired immune
1063-6706/$25.00 © 2009 IEEE
DU et al.: THEORY OF EXTENDED FDESs FOR HANDLING RANGES OF KNOWLEDGE UNCERTAINTIES AND SUBJECTIVITY 317
deficiency syndrome (AIDS) patients who are about to receive

their first combination antiretroviral treatment (the only effec-
tive long-term treatment to date) [24], [25]. The United Nations
estimates that 38 million people worldwide are infected with
HIV and that more than 22 million have died. HIV/AIDS treat-
ment is, unfortunately, among one of the most complex treat-
ments for any disease. Although the U.S. Department of Human
Health and Services HIV/AIDS treatment guidelines cover the
first-round combination antiretroviral therapy, they do not pro-
vide individualized treatment advice. In this regard, computer
software that can utilize clinical information to recommend reg-
imens (a regimen is a combination of different drugs) that will
be sufficiently potent, well tolerated, and taken on schedule
by an HIV-infected person would be a great advance. Accord-
ingly, the objective of our project was to build such a system
whose regimen choice for any given patient will match expert
AIDS physician’s selection to produce the (anticipated) optimal
Fig. 1. Type-1 fuzzy sets “medium” and “high” for potency of regimens of
treatment outcome. Preliminary retrospective evaluation of our interest as defined by a single AIDS expert or as a consensus of a group of
prototype system using patients treated in our institution’s AIDS experts. Also shown is how a potency of 85% for a particular regimen leads to
Clinical Center demonstrated encouraging results when the sys- the two memberships of the two fuzzy sets.
tem operated in either self-learning mode or nonlearning mode.
Our first-hand application experience indicates that a more timate, based on their experience and knowledge, a percentage
comprehensive FDES theory is needed to address two general number to each of the four factors for all the treatment regimens
issues critically important not only to biomedical applications, under consideration (e.g., potency is 85% for regimen X, ad-
but also to applications in other industries. First, extracting herence is 80% for regimen Y, and adverse events is 20% for
application-specific knowledge and experience from domain regimen Z). This turned out to be a difficult task for the physi-
experts is a crucial but challenging task for both the system cians. While the percentages were subjective summarization of
developer and the domain expert. It is usually not easy for the the clinical possibilities for the parameters and represented the
expert to explicitly express his/her ideas qualitatively. Doing the experts’ best point estimates, there were uncertainties as the
same quantitatively is obviously even harder. From the system true values did not exist in the literature, and consequently,
developer’s standpoint, how to accurately translate the domain the ground truth was (and still is) unknown. It is obvious that
expertise into an FDES-usable form is a major technical hur- allowing the experts to use interval numbers (e.g., potency is
dle. The expert should have means other than point estimates [82%, 88%] for regimen X) or fuzzy numbers (e.g., potency
and type-1 fuzzy sets mandated in the current framework to is around 85%), impossible in the current FDES theory, would
describe uncertainties, subjectivity, and imprecision in his/her make the knowledge acquisition process much easier and more
knowledge and experience. Let us use our HIV/AIDS applica- intuitive for the experts with more realistic knowledge represen-
tion as a concrete example to illustrate this issue. tation. A fuzzy set can be considered as a fuzzy number if a set
A physician must consider many factors when selecting an an- of conditions described later is met.
tiretroviral treatment regimen for a patient. In our FDES scheme, The second issue is that when a group of experts with dis-
the primary four clinical parameters germane to this decision are tinct opinions is involved, they should not be forced to reach
as follows: 1) anticipated potency of the regimen; 2) anticipated consensus for the sake of system development. This is because
adherence for the patient under the regimen; 3) prognosis for collective consensus may not be achievable, which is often the
adverse events under the regimen; and 4) expected future drug case in medicine, where individual experts’ opinions should be
options due to the potential for development of resistance to equally respected since the underlying ground truth is unknown
the regimen. We used the literature-cited percentage of patients most of the time. This can also be true for industrial applications.
who achieved plasma HIV ribonucleic acid (RNA) less than Continue the aforementioned HIV/AIDS treatment example. If
400 copies/mL after 48 weeks of treatment as our measure of physician A thinks potency of regimen X to be 83%, but physi-
potency. Adherence is a very complicated issue involving many cian B judges it to be 87%, under the present FDES framework,
factors impacting the patient’s ability to comply with the pre- they are forced to compromise and/or reach a consensus (e.g.,
scribed regimen. We defined adherence as expected percentage potency is 85%). The adoption of interval values or fuzzy num-
of drug doses prescribed by the physician that were actually bers (e.g., potency is [82%, 88%] or around 85%) would better
taken by the patient weekly for each regimen. The parameter represent the diverse estimates.
adverse events was defined as undesirable side effects and tox- A similar scenario occurs when the expert physicians defined
icities. Finally, the factor of future drug options meant drugs fuzzy sets to linguistically describe different levels of the four
available after the current treatment was no longer viable. In the clinical parameters. For instance, two fuzzy sets “high” and
knowledge acquisition phase, one of the tasks of the two AIDS “medium,” as shown in Fig. 1, were chosen to characterize the
expert physicians (i.e., domain experts) in our team was to es- possibilities of the system changing from the initial state (i.e.,
untreated mode) to “high” or “medium” state as far as potency is the system moving from state m to state n. δ̃ : Q̃ ◦ Σ̃ → Q̃ is
concerned. The fuzzy sets represented the consensus of the two a transition mapping that describes what event can occur at the
experts. The knowledge acquisition could not have been accom- current time and what the resulting new state is. Here, the sym-
plished had the experts failed to reach consensuses (different bol ◦ represents some fuzzy operation specified by δ̃. It includes,
experts may insist to use different fuzzy sets to reflect their own but not limited to, the max–product or max–min fuzzy infer-
experiences and knowledge). Clearly, a much better way is that ence operation in the fuzzy set theory [27]. Which operation to
experts will not be forced to reach consensus to begin with. use is application-dependent and is up to the system developer.
An effective approach is to use a type-2 fuzzy set [26], which Max–product or max–min is an operation in which the opera-
utilizes a lower primary membership function and an upper tors max( ), product( ), and min( ) are used. Product( ) means
primary membership function to bound an infinite number of multiplication of the arguments while max( ) [or min( )] op-
type-1 membership functions to characterize imprecision and eration picks up the largest (or smallest) argument among the
uncertainties. In addition, a secondary membership function arguments involved. If the current fuzzy state vector is q̃k and
can be defined over the membership grades at each value of the event σ̃ j occurs, then the new fuzzy state vector can be
potency. computed: q̃k +1 = q̃k ◦ σ̃ j .
These issues motivated us to expand the FDES theory so that We also extended the parallel composition of DES to FDES
they would be effectively dealt with. It turns out that all the ele- [10]. In the interest of brevity, they will not be presented here.
ments in fuzzy state vectors and fuzzy event transition matrices In the FDES framework, all the memberships in the fuzzy
must be fuzzy numbers, as opposed to crisp numbers under the state vectors and fuzzy event transition matrices are crisp num-
current framework. The operations of the fuzzy numbers are bers in [0, 1]. While this type of membership representation is
based on interval number operations. To differentiate the new adequate for many applications, there are situations where it may
FDES from the existing ones, we call them the extended FDES not be powerful enough to handle complex uncertainties (recall
(EFDES). We have also extended the theory to cover the parallel the issues mentioned in Section I and we will demonstrate more
composition of EFDES. Like the transition from DES to FDES specific cases later). To address this problem, we propose to
where the latter contains the former as a special case, FDES is use type-1 fuzzy sets to characterize the memberships, leading
a special case of EFDES because a crisp number is a special to the EFDES theory presented next. We will use the notations
case of a fuzzy number. That means the theory developed in this established in this section and expand them to present the new
paper can be applied to FDES as well. theory in a parallel fashion.
II. THEORY OF EFDESS B. EFDESs Theory
We first point out that multiple FDESs or EFDES running 1) Conceptual and Notational Extensions From the FDES
concurrently are required when the problem of interest is not Theory: Like FDES, a general EFDES is also modeled by a
very simple. This is the case for our HIV/AIDS application, for fuzzy automaton
example. To make the presentation better in Section II-A and
Section II-B1–II-B3, we will focus only on a single FDES or G = (Q, Σ, δ, q0 ) (2)
EFDES system when describing the theories.
Before developing the EFDES theory, we will first provide a where Q is the set of all fuzzy state vectors. Note that the
brief overview of the FDES theory. More detailed information notations for (2) and those for (1) are similar with the only
on the theory is available in [10]. difference being the symbol ∼ not adopted in (2). This will be the
case for the remaining notations for the EFDES in this section—
all the notations for the EFDES will not use this symbol to
A. Outline of the FDESs Theory distinguish them from their FDES counterparts. The state vector
A general FDES is modeled by a fuzzy automation G̃ qk at time instance k is mathematically represented as
G̃ = (Q̃, Σ̃, δ̃, q̃0 ) (1) qk = [ k V1 k

V2 ··· k
VN ].
where Q̃ is the set of all possible fuzzy state vectors and The variable q0 is used to signify the initial state vector. Here,
q̃0 ∈ Q̃ is the initial state vector. The state vector q̃k = N is the dimension of the state vector. Unlike the FDES, the
[k v1 , k v2 , . . . , k vN ] ∈ Q̃ with N being the total number of state variable k Vi is generally a fuzzy set, but can be a number
states and k = 1, 2, . . . , indicates time instance. Subscripts and in [0, 1] or an interval as special cases (in the rest of the paper, an
superscripts are positive or nonnegative integers unless pointed interval or an interval number will mean a special type-1 fuzzy
out otherwise. The state variable k vi ∈ [0, 1] represents the set whose membership function is always 1 over the interval and
possibility (i.e., membership) of the system being in state i. Σ̃ 0 outside the interval). The universe of discourse of k Vi (i.e., the
is the set of all possible events. The jth event is represented by x-axis) is [0, 100%] representing all the possible likelihoods that
an event transition matrix σ̃ j = [j am n ]N ×N where the element the EFDES can be in state i. The vertical axis of k Vi is [0, 1] to
j
am n ∈ [0, 1] describes the likelihood (i.e., membership) of represent the degree of truth. A linguistic label may be assigned
to each state variable for easier representation and interpretation k +1

When fuzzy set arguments are involved, Vn in qk +1 can
(e.g., the system is in the state “high” with possibility k Vi ). be calculated by
The set of all events is Σ and the event σ j ∈ Σ is formally
represented by a fuzzy event transition matrix
k +1
Vn = maxS k Vi × jAin
1≤i≤N
  = maxS (k V1 × jA1n , k V2 × jA2n , . . . , k VN × jAN n )

A11 j
··· j
A1N
 .. .. ..  (3)
σj =  . . . , 1≤j≤M
j
AN 1 ··· j
AN N where × means productS ( ). Alternatively, if the maxS –minS
operation is preferred, k +1 Vn is obtained by
where M is the total number of events. jAm n , 1 ≤ m, n ≤ N , k +1

Vn = maxS minS (k Vi , jAin )
is a fuzzy set characterizing the behavior of the EFDES moving 1≤i≤N
from state m to state n when the jth event takes place. The
= maxS (minS (k V1 , jA1n ),
universe of discourse of jAm n is [0, 100%] to represent all the
possibilities for the system’s transition between the two states. minS (k V2 , jA2n ), . . . , minS (k VN , jAN n )). (4)
The vertical axis ranges from 0 to 1 to show the degree of
truth. Each of state m and state n has a linguistic label. In the To carry out the operations of productS ( ), maxS ( ), and
matrix row direction and the column direction, they should be minS ( ), the arithmetic based on Zadeh’s extension principle
consistent with the linguistic labels of the corresponding state may be employed [28]. Generally speaking, however, using
vectors. Note that in our previous effort [10], all the elements a computer program to automatically produce a closed-form
in the event transition matrices are real numbers in [0, 1] (i.e., mathematical solution can be very challenging, if not impossi-
membership grades), not fuzzy sets. We should also mention ble, if the fuzzy sets are allowed to be of arbitrary functions. This
that an event itself is always crisp-–it either takes place or does is primarily due to the nature of the method in which the mem-
not happen. But unlike the conventional DES where the system bership functions of the fuzzy sets have to be compared with
can be only in one state for any given time, an FDES or EFDES each other to determine which function is the largest or smallest
is allowed to be partially in more than one state at the same time. and in which parts of the universes of discourse. Alternatively,
This is realized by fuzzy event transition matrices. one may discretize the universes of discourse of the fuzzy sets
The parameter δ in (2) is a transition mapping that describes so that an approximate numerical solution can be generated by
how to generate a new state vector from the current fuzzy state a computer program in an automation fashion.
vector and a fuzzy event transition matrix. Like the FDES theory, While permitting k Vi and jAm n to be arbitrary fuzzy sets
may be desirable from the standpoint of a theory and its com-
the mapping is mathematically described by
prehensiveness, it may not be reasonable for practical applica-
tions. Take the theory of fuzzy control and fuzzy modeling as
δ :Q◦Σ→Q an example. Theoretically speaking, arbitrary fuzzy sets are al-
lowed, but only certain types of fuzzy sets (i.e., fuzzy numbers)
and the new fuzzy state vector can be computed by qk +1 = are practically sensible, and hence have been heavily used in al-
qk ◦ σ j . Unlike the FDES theory though, the arguments of the most all the studies in the literature. Fuzzy numbers are a class
operation ◦ are no longer restricted to numbers. Instead, they of general fuzzy sets that satisfy the following four require-
can be fuzzy sets. ments [27]: 1) it is a normal (continuous) type-1 fuzzy set (i.e.,
Through the aforementioned extensions, it should be obvious its height must be 1); 2) its support set is bounded; 3) its α-cut
that the FDES theory is contained in the EFDES theory as a set is a closed interval; and 4) the support set of its strong 0-cut
special case because a crisp number is a special case of a fuzzy set is bounded. The restrictions induced by the requirements
set. are quite mild, but the benefit is substantial—eliminating many
2) Computation of New Fuzzy States: In qk +1 = qk ◦ σ j , of the general fuzzy sets that do not make practical sense in
the arguments of the operation ◦ can be fuzzy sets or their spe- the first place. Obviously, there are an infinite number of types
cial cases—interval numbers and numbers (an interval should be of fuzzy numbers, all of which meet these four requirements.
expressed as a fuzzy set whose membership is constant 1 over the Using fuzzy numbers for k Vi and jAm n should be adequate
interval while a number should be expressed as a singleton fuzzy not only because they are very diverse and flexible to represent
set). Different kinds of fuzzy sets may be used in a mixture fash- various knowledge and uncertainties, but also because they have
ion. The most common ◦ operations are the max–product oper- clearer and more intuitive meanings than fuzzy sets of arbitrary
ation and the max–min operation. To avoid possible confusion shapes do. Widely used fuzzy number types include the triangu-
regarding the type of the arguments, we use maxS –productS lar and the trapezoidal. Any type of fuzzy numbers is allowed in
[i.e., maxS ( ) and productS ( )] and maxS – minS when the ar- the EFDES theory and a specific choice depends on the system
guments are fuzzy sets, and maxI –productI and maxI – minI designer.
when the arguments are interval numbers. When the arguments An even more important advantage of using fuzzy numbers
are numbers, just max–product and max–min will be utilized, for k Vi and jAm n as opposed to general fuzzy sets is that it
which are the conventions in the literature. is much easier to obtain closed-form mathematical results for
the operations (3) and (4) using an arithmetic procedure that is minS (k Vi , jAin )
based on the α-cut set approach [28]. Better yet, the procedure
can be automatically executed by a computer program using a = α ⊗[min(kv i (α), jain (α)), min(k v i (α), j ain (α))]
software package capable of symbolic mathematical operations α ∈[0, 1]
(e.g., Mathematica and Maple software on the market). We point (10)

out that even though the complexity of their execution is signifi-
cantly different, the extension principle-based arithmetic and the where × in (8) is the standard multiplication operation for num-
α-cut-set-based approach generate the same result when fuzzy bers. Therefore, (3) becomes
numbers are involved. We now outline the latter method. k +1
Vn
According to the first decomposition theorem [27] in fuzzy
set theory, a continuous fuzzy number X can be represented as = α ⊗ max (kv i (α) × jain(α)), max (k v i (α) × j ain(α))
1 ≤i ≤N 1 ≤i ≤N
the union of its α-cut sets α ∈[0, 1]

X= α ⊗ [x(α), x(α)] (11)
α ∈[0, 1]
and (4) can be written as
where [x(α), x(α)] is an α-cut set of X, which is an interval (i.e., k +1
Vn
an interval number) containing all the elements in the universe
of discourse of X whose membership grades are greater than or = α ⊗ max (min(kv i , jain )), max (min(k v i , j ain )) .
1≤i≤N 1≤i≤N
equal to the specified membership grade α. Here, x(α) and x(α) α ∈[0, 1]
are the lower

and upper terminal points whose values depend (12)
on α. And is the Zadeh fuzzy union operation [i.e., max( )]
and α ⊗ [x(α), x(α)] means that the membership grade for We point out that because fuzzy numbers can be represented
[x(α), x(α)] is α. as the union of their α-cut sets and this representation can be
Based on the α-cut representation, the operations of fuzzy realized by a symbolic mathematical software package automat-
numbers × [i.e., products( )], maxS ( ), and minS ( ) in (3) and ically and quickly with the result stored in the computer prior
(4) can be expressed as [27] to the computations in (5)–(12), the computation burden of this
approach is low.
k
Vi × jAin 3) Construction of Fuzzy Event Transition Matrices: Fuzzy

= α ⊗ ([kv i (α), k v i (α)] × [jain (α), j ain (α)]) event transition matrices play a central role in the theory of
α ∈[0, 1] EFDES. They are the core of EFDES and capture and represent
much of expert knowledge/experience that may be subjective,
(5) vague, and imprecise in practical applications. Furthermore,
k j they are used to compute future states from the present one for a
maxS ( Vi , Ain )
particular system variable of interest, say x (e.g., potency). The
= α ⊗ maxI ([kv i (α), k v i (α)], [jain (α), j ain (α)]) process for generating a fuzzy event transition matrix in EFDES
α ∈[0, 1] is more complex than that in FDES partially because fuzzy sets,
j
(6) Am n , are now needed to be produced to represent the fuzzy
probabilities in the end of the process instead of crisp numbers in
minS (k Vi , jAin ) the case of FDES. To construct a fuzzy event transition matrix is
to determine all the fuzzy sets, each of which is an element of the
= α ⊗ minI ([kv i (α), k v i (α)], [jain (α), j ain (α)])
matrix jAm n . The elements should capture expert knowledge
α ∈[0, 1]
on state transitions and represent the possibility of the system
(7) moving from one state to another when a system event happens
(e.g., patient is treated by a regimen).
where k v i (α), k v i (α), jain (α), and j ain (α) are numbers in
For an N × N event transition matrix, there are N different
[0, 1], and × in (5) is the multiplication operation for interval
numbers. According to the interval number operations [27], the states. The element jAm n indicates the fuzzy possibility for the
previous expressions lead to the following: system variable to transfer from state m (supposedly “medium”)
to state n (supposedly “high”). The system developer may use
k
Vi × jAin real (measurement) data to compute the fuzzy possibility. How-

= α ⊗ [kv i (α) × jain (α), k v i (α) × j ain (α)] (8) ever, this can be difficult because the states themselves are only
α ∈[0, 1]
vaguely defined (e.g., “medium” and “high”); thus, quantifying
the possibility can be challenging. Additionally, the real data
maxS (k Vi , jAin ) may not be available in many real-world applications. Alterna-
tively, the developer may determine the fuzzy possibilities in a
= α ⊗[max(kv i (α), jain (α)), max(k v i (α), j ain (α))]
more subjective manner. One possible approach is to directly
α ∈[0, 1]
write out the fuzzy sets (or in special cases, membership grades
(9) and interval numbers) in an ad hoc fashion. Nevertheless, it can
be hard to manually generate N × N fuzzy sets systematically by T that is calculated according to [30]
and consistently without contradictions if N is large and/or more
than one expert’s knowledge is to be captured but experts with T = {ϕT (x, y)|ϕT (x, y)
conflicting or opposite opinions disagree with one another with- = sup (ϕD (x, y D ) ∩ ϕF (x, y F )), ∀x ∈ X
out consensus (a norm in medicine). To resolve the difficulty, we y =y D ∩y F
propose the following systematic method for determining jAm n .
y = yD ∩ yF ∀y D ∈ µD
x ⊆ [0, 1], ∀ y F ∈ µFx ⊆ [0, 1]}
With the help of a group of domain experts, the system devel-
oper first defines a fuzzy set, designated as D, over the universe (15)
of discourse of x, denoted by X, that represents the experts’
knowledge concerning the possibility of the system transiting where ∩ can be any fuzzy AND operator such as the Zadeh AND
from state m to state n for different values of x (say, potency) operator [29] and sup( ) is the operation to pick up the largest
after the jth event occurs. (For instance, D can be a fuzzy set element (i.e., the max( ) operator). The primary membership
describing the possibility of general patient population moving grade y (=y D ∩ y F ) and the secondary membership grade
from “medium” potency state to “high” potency state after they (=ϕD (x, y D ) ∩ ϕF (x, y F )) are computed. If more than one
start regimen ABC (i.e., the jth event) whose potency ranges value of ϕD (x, y D ) ∩ ϕF (x, y F ) is produced for the same value
from, say, 70% to 95%). This fuzzy set can be either type-1 or of y, the largest value will be selected as the result by the sup( )
type-2, depending on the extent of experts’ disagreement. It does operation.
not need to be defined over the entire universe of discourse; only A fuzzy set type reducer is then needed to convert T to a
part of the universe of interest would suffice (e.g., 60%–100% type-1 fuzzy set, which is jAm n . Note that jAm n represents the
in the aforementioned example). Then, the developer needs to relationship between y and z, instead of x and y needed for other
assign a value, designated as F, which can be a crisp number, type-2 applications in the literature. There exist several type
an interval, a type-1 fuzzy set, or an type-2 fuzzy set, to x (i.e., reducers, and we modify the so-called height type-reduction
the potency value of regimen ABC) that is associated with the method for our purpose. Consequently, the membership function
particular event. for jAm n , denoted by jAm n (y), defined over the universe of
Without loss of generality, assume that D is a type-2 fuzzy discourse [0, 1] is [29]
set represented by a 3-D membership function z = ϕD (x, y), j
Am n (y) = sup ϕT (x, y). (16)
where x ∈ X and y is a variable representing the membership x
grade for x. The relationship between x and y is characterized
Since F can be a crisp number, an interval, a type-1 fuzzy set, or
by two membership functions (instead of one in the case of
a type-2 fuzzy set, and D can be a type-1 fuzzy set or a type-2
type-1 fuzzy sets)—an upper membership function and a lower
fuzzy set, there exist eight combinations. Formula (15) covers
membership function. For a given value of x = x , the upper
the most general case—both F and D are type-2 fuzzy sets. The
and lower membership functions will intersect with the vertical
formulas for the rest of the seven situations can be derived from
line of x = x , and respectively produce an upper point and
(15) as a special case. The results are summarized in Table I.
a lower terminal point of an interval of membership grades.
Note that situation No. 8 is the FDES.
The interval is denoted by µD D
x . Obviously, µx is a subinterval There lacks a method in the current literature to produce a so-
of [0, 1] (i.e., µx ⊆ [0, 1]). A type-1 fuzzy set can be defined
D
lution for (15) when both D and F are general type-2 fuzzy sets.
over µD x , which mathematically is z = ϕD (x , y). Therefore, If they are interval type-2 fuzzy sets, numerical solutions can
z represents the membership of y. The collection of all these
be obtained via algorithm for situations Nos. 1–5 [29]. The re-
type-1 fuzzy sets for the entire universe of discourse of x is
maining three situations are readily computable as they involve
called a secondary membership function [29], which is z =
type-1 fuzzy sets only. For better illustration, in Section III, we
ϕD (x, y). All the intervals µD x due to different values of x form will use the HIV/AIDS example to demonstrate how to carry
so-called primary memberships µD x (obviously, µx ⊆ [0, 1]).
D
out (15) and (16) for situations Nos. 4 and 6.
The collection of all the left and right terminal points of µD x 4) Parallel Composition of EFDESs: Like the DES and
forms the upper and lower membership functions, respectively.
FDES theories, to model a (complex) EFDES, a convenient
Note that the two functions along with the area bounded by them
way is often to model each subsystem first, and then synthe-
are the primary memberships. Formally, D can be represented by
size them by parallel composition. Since the memberships of
all the states and events have been extended from numbers to
D = {ϕD (x, y)|∀x ∈ X, ∀y ∈ µD
x }. (13)
fuzzy sets, we need to modify the operations of the parallel
composition accordingly. Suppose two EFDES are given
In the interest of generality, we suppose that F is a type-2 fuzzy
set (crisp number, interval number, and type-1 fuzzy set are all Gm 1 = (Qm 1 , Σm 1 , δm 1 , q0 ) Gm 2 = (Qm 2 , Σm 2 , δm 2 , q0 ).
its special cases)
Then, through parallel composition , the combined EFDES
F = {ϕF (x, y)|∀x ∈ X, ∀y ∈ µFx }. (14) Gm = (Qm , Σm , δm , q∗0 ) is
Gm = Gm 1 ||Gm 2
We then apply a fuzzy intersection operator on D and F.
Generally speaking, the result is a type-2 fuzzy set. Denote it = (Qm 1 × Qm 2 , Σm 1 ∪ Σm 2 , δm 1 × δm 2 , (q0 , q0 ))
TABLE I where m, m1 , and m2 are positive integers. We now describe

FORMULAS FOR COMPUTING THE FUZZY NUMBERS jA m n IN FUZZY EVENT
TRANSITION MATRICES
how the parallel composition of EFDES is executed.
There are two steps: 1) compute the state vectors for Gm
and 2) compute the fuzzy event transition matrices for Gm . For
the first step, assume that qh and qi are fuzzy state vectors in
Qm 1 and Qm 2 , respectively, and h Vj (1 ≤ j ≤ N ) and i Vk
(1 ≤ k ≤ N ) are elements of qh and qi , respectively. Here, h
and i are nonnegative integers, and j, k, N , and N are positive
integers, where N and N are the dimensions of fuzzy state vec-
tors in Gm 1 and Gm 2 , respectively. The dimension of the state
vectors in Gm is one row and N × N column. The composi-
tion of the jth element in qh (i.e., h Vj ) with the kth element in
qi (i.e., i Vk ) yields the element in the composite state vector
located in the column corresponding to (j, k). The value of that
new element is a fuzzy set calculated by
minS (h Vj , i V k ). (17)
In the second step, we determine the event transition matrices
of Gm . Designate M ∗ , M , and M as the total number of events
for Gm , Gm 1 , and Gm 2 , respectively. Suppose that β 1 Ak 1 k 2 is
an element of the event matrix σ β 1 (1 ≤ β1 ≤ M ) in Gm 1 and
β2
Aj 1 j 2 is an element of the event matrix σ β 2 (1 ≤ β2 ≤ M )
in Gm 2 . After the parallel composition, the dimension of σ β 1
and σ β 2 will no longer be the same as that in Gm 1 and Gm 2 ,
respectively. The dimension of all the events in Gm should be
(N × N ) × (N × N ). Let σ ∗β 1 signify the representation of
σ β 1 in Gm (discussion concerning σ ∗β 2 would be similar and is
omitted for brevity). Suppose that β 1 A∗k j (1 ≤ k, j ≤ N × N )
is an element in σ ∗β 1 that represents a concurrent transition from
state k1 to state k2 (1 ≤ k1 , k2 ≤ N ) in Gm 1 and from state j1
to state j2 (1 ≤ j1 , j2 ≤ N ) in Gm 2 . It can be determined as
follows.
The first scenario is that Gm 2 does not have the event σ β 1 .
Then
β1
β1 ∗
Ak 1 k 2 , if j1 = j2
Ak j = (18)
Z, if j1 = j2 .
For those β 1 A∗k j corresponding to the unchanged state in Gm 2
(i.e., j1 = j2 ), the behavior of the event transition from k to j
in Gm should be the same as that from k1 to k2 in Gm 1 . Hence,
β1 ∗
Ak j = β 1 Ak 1 k 2 . For those β 1 A∗k j representing state change
in Gm 2 (i.e., when j1 = j2 ), β 1 A∗k j = Z, where Z represents a
type-1 singleton fuzzy number 0 whose membership is 1 at 0 in
the universe of discourse and is 0 elsewhere (by fuzzy number
notation, this fuzzy number can be expressed as 1/0). This is
because Gm 2 does not have the event σ β 1 ; hence, any state
change in Gm 2 should not appear in Gm , and thus Z.
The second scenario is that Gm 2 also has the event σ β 1 . Then
β1
A∗k j = minS (β 1 Ak 1 k 2 , β 2 A j 1 j 2 ).
This is because σ β 1 now takes place in both Gm 1 and Gm 2 to
the extents of β 1 Ak 1 k 2 and β 2 Aj 1 j 2 , respectively. Therefore,
the fuzzy AND operation should be applied (e.g., the Zadeh AND
operator used here).
We point out that state explosion problem due to parallel com-
position is universal to all DESs and EFDES is no exception.
Many methods have been proposed in the DES literature to ad-

dress this problem. They include modular control, decentralized
control (e.g., [31]), hierarchical control [32], online control with
limited and variable look-ahead policies [33], and various tech-
niques for model abstractions. These methods can also be used
in EFDES.
III. ILLUSTRATIVE EXAMPLES

To make the aforementioned theoretical development easier to
understand, we now provide three detailed numerical examples
that are related to medical applications.
Example 1: Construct fuzzy event transition matrices of
EFDES for HIV/AIDS treatment. Continue the HIV/AIDS treat-
ment application mentioned in Section I. We will consider two
of the eight scenarios in Table I (more specifically, situations
Nos. 4 and 6). Fig. 2. Same two fuzzy sets “medium” and “high,” as shown in Fig. 1. Also
shown is type-1 fuzzy number P defining potency of the particular regimen
There were three states for regimen’s potency—Initial, being “around 85%” and fuzzy number P M resulted from fuzzification using
Medium, and High. When the FDES theory was applied to a the Zadeh fuzzy intersection operation.
treatment-naive patient, the state vector was
Initial Medium High We now discuss how to produce the fuzzy numbers 1A12 and
1
A13 under different conditions.
q̃0 = [ 1 0 0 ]
which means that the patient was not in Medium or High potency A. First Scenario
state but in the Initial state because he/she had never received The first scenario is that a group of experts all agree to use
treatment before. For clearer illustration, we list the state names the same type-1 fuzzy sets [corresponding to the fuzzy set D in
on the top of the vector. Using the EFDES theory presented (13)] “Medium” and “High” for potency (the fuzzy sets describe
before, the three crisp membership values become three fuzzy the possibilities of the system changing from the initial state to
numbers as follows: “Medium” or “High” state). But they fail to reach a consensus
Initial Medium High on the exact potency value of the regimen. Hence, the regimen’s
potency should not be a crisp number but a fuzzy number. This
q0 = [ 0 V1 0
V2 0
V3 ] (19) is situation No. 6 in Table I. Fig. 2 shows such a hypothetical
where let 0 V1 = 1/1 and 0 V2 = 0 V3 = Z. Z is the singleton example where a symmetrical triangular fuzzy number P cen-
fuzzy number 0 defined in Section II-B4 while 1/1 represents tered on 85% is used to represent diverse opinions of the experts
a type-1 singleton fuzzy number 1. One may use other fuzzy on the potency value [corresponding to the fuzzy set F in (14)].
numbers 0 and 1 (e.g., the trapezoidal function), which will The membership function of P is
likely make the example complicated in computation.  x − 0.82

 , 0.82 < x < 0.85
Now, we need to create the fuzzy event transition matrix for 

 0.03
potency, which is 3 × 3. The process for the FDES is illustrated
µP (x) = 0.88 − x
in Fig. 1 where the regimen’s potency is assumed to be 85%. 
 , 0.85 ≤ x < 0.88

 0.03
Using the two membership grades in Fig. 1 (i.e., the intersections 
of the vertical line of 85% potency and the two fuzzy sets) can 0, elsewhere in [0, 1]
construct the fuzzy event transition matrix for the FDES as where x is the potency value. Using Table I, one can obtain
1
A12 and 1A13 . By applying an intersection operation (min( )
Initial Medium High
  is used here), two fuzzy numbers, PM and PH , are resulted
0 0.4296 1 Initial [corresponding to the fuzzy set T computed by (15)]. PM is
  produced by intersecting P with “Medium” point by point along
σ̃ 1 =  0 0 0  Medium (20)
the potency axis. The membership function of the resulting PM ,
0 0 0 High.
shown as a bold curve in Fig. 2, is
The corresponding fuzzy event transition matrix σ 1 for the 
 x − 0.82
EFDES is 
 , 0.82 < x < 0.8354

 0.03


Initial Medium High 
 − 12 ( x −00 . .17 2 )2
  e , 0.8354 ≤ x < 0.8703
µP M (x) =
Z 1
A12 1
A13 Initial 
 0.88 − x
  
 , 0.8703 ≤ x < 0.88
σ1 =  Z Z Z  Medium (21) 


 0.03

Z Z Z High. 0, elsewhere in [0, 1].
The minimal and maximal membership values of µP M (x) are

0 and 0.5138, respectively (Fig. 2). Therefore, according to
Table I, the fuzzy number 1A12 is an interval number [0, 0.5138].
In our case, PH happens to be identical to P (thus, it is not marked
in Fig. 2). Subsequently, the minimal and maximal membership
values after the intersection are 0 and 1, respectively, which
means that 1A13 is an interval number [0, 1]. Hence, the fuzzy
event transition matrix σ 1 in (21) is
 
Z [0, 0.5138] [0, 1]
 
σ1 =  Z Z Z . (22)
Z Z Z
Compared to the FDES event transition matrix σ̃ 1 in (20), the

nonzero elements in (22) are interval numbers instead of crisp
numbers. The practical implication is that the uncertainty on
the exact potency value has been conveniently expressed by
the domain expert (i.e., the physician), and has been effectively
processed by the EFDES approach.
B. Second Scenario
This scenario is the opposite of the first scenario—the experts
agree to use the exact same (crisp) value to represent the regi-
men’s potency, but they have different definitions for the term
“Medium” (and/or “High”). If the inclusion of all the differ-
ent opinions of the experts is desired, “Medium” and “High”
should be defined as type-2 fuzzy sets [29]. This is situation
no. 4 in Table I. Fig. 3 shows an example. The primary mem-
berships of a type-2 fuzzy set consists of an upper membership
function and a lower membership function to bound a range of
opinions technically called footprint of uncertainty [Fig. 3(a)].
The primary memberships for “Medium” and “High” are math-
ematically defined in Table I, and the two bounded regions (i.e.,
the footprints of uncertainty) are shown as two gray areas with
distinct intensities in Fig. 3(a). Fig. 3. Define “medium” and “high” potency as type-2 fuzzy sets to cover
According to Table I, we can obtain 1A12 and 1A13 as fol- diverse opinions of a group of AIDS experts. (a) Primary memberships with
the footprints of uncertainty of the type-2 fuzzy sets “medium” and “high.” (b)
lows. Draw a vertical line at a potency value [85% in Fig. 3(a)]. Secondary membership function for the fuzzy set “medium” at potency of 85%
The line will intersect with the upper and lower membership is supposedly triangular (solid line) or if the type-2 fuzzy set is assumed to be
functions of the two type-2 fuzzy sets, resulting in two intervals an interval type-2 fuzzy set (dotted line).
of the membership grades—one for “Medium” and the other for
“High.” They are actually two results after applying the Zadeh
fuzzy intersection operation to “Medium” and the 85% potency for “Medium” at 85% potency is of the triangular type
line, and “High” and the same line, respectively. The intervals 
 2(y − yM ) y M + yM
then serve as the universes of discourse for secondary member- 
 , yM < y <

 y − y 2
ship functions of the respective type-2 fuzzy sets at the potency 
 M
M
value where the vertical line is put. A secondary membership ϕM (y) = 2(yM − y) yM + y M
function describes the likelihood whether each of the member- 
 , ≤ y < yM

 y − y 2
ship grades will take place. Like the primary memberships and 
 M M

footprint of uncertainty, the secondary membership functions 0, elsewhere in [0, 1]
are developed by the experts/developer. For instance, they may
be defined in such a way to reflect the experts’ individual prefer- where y, the horizontal axis of Fig. 3(b), is the membership
ence/bias exhibited when they define the primary memberships. grade for potency in Fig. 3(a). Similarly, the secondary mem-
In Fig. 3(b), we suppose that the secondary membership function bership function for the type-2 fuzzy set “High” is supposed to
TABLE II membership grades for a parameter of interest (e.g., potency) at

MATHEMATICAL DEFINITIONS OF THE LOWER AND UPPER PRIMARY
MEMBERSHIP FUNCTIONS OF THE TYPE-2 FUZZY SETS FOR POTENCY
each parameter value equally or uniformly, we can simplify the
SHOWN IN FIG. 3(A) aforementioned results by letting ϕM (y) = 1 or ϕH (y) = 1. To
make a concrete example, if the secondary membership function
for “Medium” is as shown in Fig. 3(b), then the fuzzy number
1
A12 reduces to an interval
ϕ1 A 1 2 (y) = [0.3247, 0.5461] .
Example 2: Computing new fuzzy states using fuzzy event
transition matrices. Assume that the fuzzy event transition ma-
trix is (23). It can then be used to compute new fuzzy state vector
q1 from the current state, which is the initial state q0 in (19)
 
Z 1A12 1A13
1  
q1 = q0 ◦ σ 1 = Z Z ◦ Z Z Z 
be triangular 1
 Z Z Z
 2(y − yH ) y H + yH

 , yH < y < = [Z 1
A12 1
A13 ] .

 yH − yH 2


ϕH (y) = 2(yH − y) yH + y H The maxS –productS operation is used for ◦ as all the ele-

 , ≤ y < yH ments are type-1 fuzzy numbers.
 yH − yH
 2

 To make this example nontrivial, let us now assume that the

0, elsewhere in [0, 1]. HIV/AIDS patient has failed the first-round treatment and must
Note that yM , yM , yH , and yH are marked in Fig. 3(a) and their receive a second-round treatment. Without loss of generality,
values can be determined using the functions in Table II. Given suppose the potency event transition matrix for a second-round
potency of 85%, yM = 0.3247, yM = 0.5461, yH = 0.9802, and regimen that the patient takes is obtained via one of the two
yH = 1. approaches described earlier
Assigning ϕM (y) and ϕH (y) to the fuzzy numbers 1A12 and  
Z Z Z
1
A13 , respectively, completes the process of producing the ma- σ 2 =  Z 2A22 Z 
trix. The fuzzy number 1A12 is triangular 2 2
 Z A32 A33
 9.0334y − 2.9331, 0.3247 < y < 0.4354

where the type-1 fuzzy numbers A22 , 2A32 , and 2A33 have the 2
ϕ1 A 1 2 (y) = 4.9331 − 9.0334y, 0.4354 ≤ y < 0.5461 following triangular membership functions:


0, elsewhere in [0, 1] 
 10y − 2, 0.2 < y < 0.3
and the fuzzy number 1A13 is also triangular ϕ2 A 2 2 (y) = 4 − 10y, 0.3 ≤ y < 0.4
 
 101.0101y − 99.0101,
 0.9802 < y < 0.9901 0, elsewhere in [0, 1]

ϕ1 A 1 3 (y) = 101.0101 − 101.0101y, 0.9901 ≤ y < 1  10y − 6, 0.6 < y < 0.7


0, elsewhere in [0, 1]. ϕ2 A 3 2 (y) = 8 − 10y, 0.7 ≤ y < 0.8

Consequently, the fuzzy event transition matrix σ 1 is 0, elsewhere in [0, 1]

   10y − 8, 0.8 < y < 0.9
Z ϕ1 A 1 2 ϕ1 A 1 3
σ1 =  Z ϕ2 A 3 3 (y) = 10 − 10y, 0.9 ≤ y < 1 .
Z Z . (23) 
0, elsewhere in [0, 1]
Z Z Z
In contrast to the FDES event transition matrix σ̃ in (20) The new potency state after the patient is treated is
1
where all the elements are crisp numbers, all the elements in (23) q2 = q1 ◦ σ 2
are fuzzy numbers. From practice perspective, the advantage is  
substantial—the EFDES method is capable of simultaneously Z Z Z
capturing and representing a range of opinions contributed by = [Z 1
A12 1
A13 ] ◦  Z 2
A22 Z 
2 2
a group of domain experts that do not agree with one another. Z A32 A33
This is something unachievable under the framework of FDES.
= [Z maxS (Z, 1A12 × 2A22 , 1A13 × 2A32 )
So far, we have considered the case when the type-2 fuzzy
sets are general in the sense that their primary memberships and maxS (Z, Z, 1A13 × 2A33 )]
secondary membership functions can be arbitrary but reasonable
= [Z maxS (1A12 × 2A22 , 1A13 × 2A32 ) A13 × 2A33 ]
1
shapes. If the fuzzy set is an interval type-2 fuzzy set, which
2 2
can be in the case if the experts want to treat the likelihoods of = [Z V2 V3 ]
and G2 heart condition. In G1 , the lung condition is classified

as “Excellent” (E, for short), “Fair” (F), and “Poor” (P), which
is represented by a fuzzy state vector
qk = [ k V1 k
V2 k
V3 ]
k k k
where V1 , V2 , and V3 are fuzzy numbers charactering the
states “Excellent,” “Fair,” and “Poor,” respectively. There are
six events denoted as σ j , j = 1, . . . , 6. Patient’s heart condition
is also classified as “Excellent” (E , for short), “Fair” (F ), and
“Poor” (P ), which is described by a fuzzy state vector in G2
qi = [ i V 1 i
V 2 i
V 3 ].
The three elements are fuzzy numbers concerning “Excellent,”
“Fair,” and “Poor.” Suppose that there are six events in total for
the heart as well; they are labeled as αn , n = 1, . . . , 6. Also,
suppose that G1 and G2 do not share a common event, i.e., the
treatment will not affect heart and lung simultaneously.
Fig. 4. Graphical representations of the triangular fuzzy numbers 2 V 2 and The dimension of the resultant state vector of G1 G2 is 9
2V .
3 (i.e., 3 × 3) that covers the following nine combinations of the
states: (E, E ), (E, F ), (E, P ), (F, E ), (F, F ), (F, P ), (P, E ),
(P, F ), and (P, P ). The value of each composite state can be
where 2 V2 and 2 V3 are type-1 fuzzy numbers whose member- calculated according to (17)
ship functions are (Fig. 4)
√ (qk , qi ) = [(E, E ), (E, F ), (E, P ), (F, E ), (F, F ),
 √ 2162.7 + 1010.1y − 52.5051, 0.5881 < y < 0.6931
ϕ2 V 2 (y) = − 2162.7 + 1010.1y + 54.5051, 0.6931 ≤ y < 0.8 (F, P ), (P, E ), (P, F ), (P, P )]

0, elsewhere in [0, 1] = [minS (k V1 , i V 1 ), minS (k V1 , i V 2 ), minS (k V1 , i V 3 ),
√
 √ 2070.7 + 1010.1y − 53.5051, 0.7842 < y < 0.8911 minS (k V2 ,i V 1 ), minS (k V2 , i V 2 ), minS (k V2 , i V 3 ),
ϕ2 V 3 (y) = − 2070.7 + 1010.1y + 55.5051, 0.8911 ≤ y < 1
 minS (k V3 ,i V 1 ), minS (k V3 , i V 2 ), minS (k V3 , i V 3 )].
0, elsewhere in [0, 1].
Next, we reformulate all the event transition matrices in G1 and
Using the key building blocks illustrated in Examples 1 and 2,
G2 . Without loss of generality, let us focus on event σ 1 only,
we are developing an EFDES for HIV/AIDS treatment that will
which is related to change of lung condition from “Excellent” to
include a system validation against the actual patient treatment
“Fair.” It is represented by the following fuzzy transition matrix:
records. A complete description of the system is beyond the 1 
scope and space limit of this paper. It would require a full-length B11 1B12 1B13
paper as was the case for our FDES HIV/AIDS system [24].  
σ 1 =  1B21 1B22 1B23 .
Example 3: Parallel composition of EFDESs. This example is 1
B31 1B32 1B33
unrelated to the first two examples and HIV/AIDS. However, in
a sense it is realistic in that it is related to two important human Elements in σ ∗1 , which is σ 1 after G1 G2 , can be computed
organs—heart and lung. Suppose an EFDES is composed of two using (18) since G2 does not contain event σ 1 . The result is,
subsystems G1 and G2 . G1 describes patient’s lung condition (∗), as shown at the bottom of this page.
(E, E ), (E, F ), (E, P ), (F, E ), (F, F ), (F, P ), (P, E ), (P, F ), (P, P )

1 
B11 Z Z 1
B23 Z Z 1
B13 Z Z (E, E )
 
 Z 1
B11 Z Z 1
B23 Z Z 1
B13 Z  (E, F )
 
 Z Z 1
B11 Z Z 1
B12 Z Z 1
B13 
  (E, P )
1 
 B21 Z Z 1
B22 Z Z 1
B23 Z Z 
  (F, E )
∗  
σ1 =  Z 1
B21 Z Z 1
B22 Z Z 1
B23 Z  (F, F ) (*)
 
 Z Z 1
B21 Z Z 1
B22 Z Z 1
B23 
  (F, P )
1 
 B31 Z Z 1
B32 Z Z 1
B33 Z Z 
  (P, E )
 
 Z 1
B31 Z Z 1
B32 Z Z 1
B33 Z  (P, F )
Z Z 1
B31 Z Z 1
B32 Z Z 1
B33 (P, P ).
For better illustration, we list all the state combinations on [12] Y. Cao and M. Ying, “Observability and decentralized control of fuzzy
the top and right side of the matrix. We now explain how the discrete-event systems,” IEEE Trans. Fuzzy Syst., vol. 14, no. 2, pp. 202–
216, Apr. 2006.
elements are obtained via (18). Take the situation when lung [13] Y. Cao, M. Ying, and G. Chen, “State-based control of fuzzy discrete-
condition changes from “Excellent” to “Fair” but heart condition event systems,” IEEE Trans. Syst. Man, Cybern. B, Cybern., vol. 37,
remains the same as an example. It involves three elements in no. 2, pp. 410–424, Apr. 2007.
[14] R. Huq, G. K. I. Mann, and R. G. Gosine, “Fuzzy discrete-event system
σ ∗1 : (E, E )→ (F, E ), (E, F )→ (F, F ), and (E, P )→ (F, P ). based behavior modulation in mobile robotics,” in Proc. 2005 IEEE/RSJ
Clearly, the value of all these three elements is 1B21 according Int. Conf. Intell. Robots Syst., Edmonton, AB, Canada, 2005, pp. 2241–
to (18). On the other hand, consider the situation when both 2246.
[15] R. Huq, G. K. I. Mann, and R. G. Gosine, “Distributed fuzzy discrete-
lung and heart conditions change from “Excellent” to “Fair.” event system for robotic sensory information processing,” Expert Syst.,
It involves only one element in σ ∗1 corresponding to (E, E )→ vol. 23, pp. 273–288, 2006.
(F, F ). Given that j1 = j2 and G2 does not contain event σ 1 , [16] R. Huq, G. K. I. Mann, and R. G. Gosine, “Behavior-modulation tech-
nique in mobile robotics using fuzzy discrete-event system,” IEEE Trans.
extent for the behavior of event transition (E, E )→ (F, F ) in Robot., vol. 22, no. 5, pp. 903–916, Oct. 2006.
Gm should be Z. [17] F. Liu and D. W. Qiu, “Decentralized supervisory control of fuzzy discrete-
event systems,” in Proc. Eur. Control Conf. 2007 (ECC’07), Kos, Greece,
2007, pp. 4984–4990.
IV. CONCLUSION [18] F. Liu and D. W. Qiu, “Decentralized supervisory control of fuzzy discrete-
event systems,” Eur. J. Control, vol. 14, pp. 234–243, 2008.
We have extended the theory of FDES to a theory of EFDES [19] D. Qiu, “Supervisory control of fuzzy discrete-event systems: A formal
approach,” IEEE Trans. Syst. Man, Cybern. B, Cybern., vol. 35, no. 1,
that can effectively deal with the two general and critically im- pp. 72–88, Feb. 2005.
portant issues concerning the representation and acquisition of [20] D. Qiu and F. Liu, “Fuzzy discrete-event systems under fuzzy observability
experts’ knowledge and subjectivity. Under the new framework, and a test-algorithm,” IEEE Trans. Fuzzy Syst., to be published.
[21] J. N. Mordeson and D. S. Malik, Fuzzy Automata and Languages: Theory
elements in the fuzzy state vectors and fuzzy event transition and Applications. Boca Raton,, FL, U.K.: Chapman & Hall/CRC, 2002.
matrices can be type-1 fuzzy numbers as opposed to numbers in [22] G. G. Rigatos, “Fuzzy stochastic automata for intelligent vehicle control,”
the original FDES theory. The extended theory not only enables IEEE Trans. Ind. Electron., vol. 50, no. 1, pp. 76–79, Feb. 2003.
[23] H. Ying, F. Lin, R. D. MacArthur, J. A. Cohn, D. Barth-Jones, H. Ye, and
domain experts to more conveniently and realistically express L. R. Crane, “A self-learning fuzzy discrete-event system for HIV/AIDS
their uncertain and imprecise knowledge and subjective judg- treatment regimen selection,” IEEE Trans. Syst. Man, Cybern.B, Cybern.,
ments but also makes their capture and representation possible. vol. 37, no. 4, pp. 966–979, Aug. 2007.
[24] H. Ying, F. Lin, R. D. MacArthur, J. A. Cohn, D. C. Barth-Jones, H. Ye,
We have established mathematical operations that cover the con- and L. R. Crane, “A fuzzy discrete-event system approach to determin-
struction of the event transition matrices and the computation of ing optimal HIV/AIDS treatment regimens,” IEEE Trans. Inf. Technol.
the new fuzzy states and new fuzzy event transitions. Further- Biomed., vol. 10, no. 4, pp. 663–676, Oct. 2006.
[25] F. Lin, H. Ying, R. D. MacArthur, J. A. Cohn, D. Barth-Jones, and
more, we have developed the corresponding parallel composi- L. R. Crane, “Decision making in fuzzy discrete-event systems,” Inf.
tion. Three detailed numerical examples are provided to show Sci., vol. 177, pp. 3749–3763, 2007.
how to use the key components of the extended theory. [26] J. M. Mendel, “Advances in type-2 fuzzy sets and systems,” Inf. Sci.,
vol. 177, pp. 84–110, 2007.
[27] G. J. Klir and B. Yuan, Fuzzy Sets and Fuzzy Logic: Theory and Applica-
tions. Upper Saddle River, NJ: Prentice-Hall, 1995.
REFERENCES [28] M. Grabisch, H. T. Nguyen, and E. A. Walker, Fundamentals of Un-
certainty Calculi With Applications to Fuzzy Inference. Norwell, MA:
[1] C. G. Cassandras and S. Lafortune, Introduction to Discrete-Event Sys-
Kluwer, 1995.
tems. Norwell, MA: Kluwer, 1999.
[29] J. M. Mendel, Uncertain Rule-Based Fuzzy Logic Systems: Introduction
[2] F. Lin and W. M. Wonham, “On observability of discrete-event systems,”
and New Directions. Upper Saddle River, NJ: Prentice-Hall, 2000.
Inf. Sci., vol. 44, pp. 173–198, 1988.
[30] N. N. Karnik and J. M. Mendel, “Operations on type-2 fuzzy sets,” Fuzzy
[3] P. J. Ramadge and W. M. Wonham, “Supervisory control of a class of
Sets Syst., vol. 122, pp. 327–348, 2001.
discrete event processes,” SIAM J. Control Optim., vol. 25, pp. 206–230,
[31] F. Lin and W. M. Wonham, “Decentralized supervisory control of discrete-
1987.
event systems,” Inf. Sci., vol. 44, pp. 199–224, 1988.
[4] E. S. Santos, “General formulation of sequential machines,” Inf. Control,
[32] F. Lin and W. M. Wonham, “Decentralized control and coordination of
vol. 12, pp. 5–10, 1968.
discrete-event systems with partial observation,” IEEE Trans. Autom.
[5] W. G. Wee and K. S. Fu, “A formulation of fuzzy automata and its appli-
Control, vol. 35, no. 12, pp. 1330–1337, Dec. 1990.
cation as a model of learning systems,” IEEE Trans. Syst. Sci. Cybern.,
[33] N. Ben Hadj-Alouane, S. Lafortune, and F. Lin, “Variable lookahead
vol. SSC-5, no. 3, pp. 215–223, Jul. 1969.
supervisory control with state information,” IEEE Trans. Autom. Control,
[6] F. Steimann and K. P. Adlassning, “Clinical monitoring with fuzzy au-
vol. 39, no. 12, pp. 2398–2410, Dec. 1994.
tomata,” Fuzzy Sets Syst., vol. 61, pp. 37–42, 1994.
[7] J. Waissman, R. Sarratr, T. Escobet, J. Aguilar, and B. Dahhou, “Wastewa-
ter treatment process supervision by means of a fuzzy automaton model,”
in Proc. IEEE Int. Symp. Intell. Control, Rio Patras, Greece, 2000, pp. 163–
168.
[8] A. Kandel and S. C. Lee, Fuzzy Switching and Automata: Theory and Xinyu Du (S’05) received the B.S. and M.S.
Applications. New York: Russak, 1979. degrees in automation from Tsinghua University,
[9] F. Lin and H. Ying, “Fuzzy discrete-event systems and their observabil- Beijing, China, in 2001 and 2004, respectively. He
ity,” in Proc. Annu. Conf. North Ame. Fuzzy Inf. Process. Soc.–NAFIPS, is currently working toward the Ph.D. degree in the
Vancouver, BC, Canada, 2001, pp. 1271–1276. Department of Electrical and Computer Engineering,
[10] F. Lin and H. Ying, “Modeling and control of fuzzy discrete-event sys- Wayne State University, Detroit, MI.
tems,” IEEE Trans. Syst. Man, Cybern. B, Cybern., vol. 32, no. 4, pp. 408– His current research interests include the devel-
415, Aug. 2002. opment of novel biomedical acoustic sensor sys-
[11] Y. Cao and M. Ying, “Supervisory control of fuzzy discrete-event systems, type-2 fuzzy systems, and fuzzy discrete event
tems,” IEEE Trans. Syst. Man, Cybern. B, Cybern., vol. 35, no. 2, pp. 366– systems.
371, Apr. 2005.
Hao Ying (S’88–M’90–SM’97) received the B.S. Feng Lin (S’86–M’87–SM’07) received the B.Eng.
and M.S. degrees in electrical engineering from degree in electrical engineering from Shanghai Jiao-
Donghua University (formerly China Textile Univer- Tong University, Shanghai, China, in 1982, and the
sity), Shanghai, China, in 1982 and 1984, respec- M.A.Sc. and Ph.D. degrees in electrical engineering
tively, and the Ph.D. degree in biomedical engineer- from the University of Toronto, Toronto, ON, Canada,
ing from the University of Alabama, Birmingham, in in 1984 and 1988, respectively.
1990. From 1987 to 1988, he was a Postdoctoral Fel-
In 2000, he joined Wayne State University (WSU), low at Harvard University, Cambridge, MA. Since
Detroit, MI, where he is currently a Professor in the 1988, he has been with the Department of Electrical
Department of Electrical and Computer Engineering. and Computer Engineering, Wayne State University,
Between 1992 and 2000, he was a faculty member at Detroit, MI, where he is currently a Professor. His
the University of Texas Medical Branch, Galveston. Between 1998 and 2000, current research interests include discrete-event systems, hybrid systems, ro-
he was an Adjunct Associate Professor of the Biomedical Engineering Program bust control, and image processing. He was a consultant for GM, Ford, Hitachi,
at the University of Texas, Austin. He has authored the book Fuzzy Control and other auto companies. He is the author of a book titled Robust Control
and Modeling: Analytical Foundations and Applications (IEEE Press, 2000). Design: An Optimal Control Approach.
He has authored or coauthored 89 peer-reviewed journal publications and over Prof. Lin received a George Axelby Outstanding Paper Award from the
120 conference papers. He is an Associate Editor, an Area Editor, or a member IEEE Control Systems Society. He was also the recipient of a Research Initi-
of the Editorial Board for 11 international journals. ation Award from the National Science Foundation, an Outstanding Teaching
Prof. Ying is a member of the Fuzzy Systems Technical Committee of the Award from Wayne State University, a Faculty Research Award from ANR
IEEE Computational Intelligence Society and is also the Chair of its Task Force Pipeline Company, and a Research Award from Ford. He was an Associate Ed-
on Competitions. He is an elected board member of the North American Fuzzy itor of the IEEE TRANSACTIONS ON AUTOMATIC CONTROL.
Information Processing Society (NAFIPS). He was the Program Chair for the
NAFIPS Conference in 1994 and 2005 and the Publication Chair for the 2000
IEEE International Conference on Fuzzy Systems. He was a Program Commit-
tee Member for 35 international conferences.
A Hybrid Approach for Design of Stable Adaptive

Fuzzy Controllers Employing Lyapunov Theory and
Particle Swarm Optimization
Kaushik Das Sharma, Amitava Chatterjee, and Anjan Rakshit
Abstract—This paper proposes a new approach for designing cal control law leads the system toward larger integral absolute
stable adaptive fuzzy controllers, which employs a hybridization error (IAE) [2]. A major drawback of many existing designs
of a conventional Lyapunov-theory-based approach and a parti- of adaptive FLCs (AFLCs) is that they are developed for sys-
cle swarm optimization (PSO) based stochastic optimization ap-
proach. The objective is to design a self-adaptive fuzzy controller, tems with unlimited actuation authority [21]. In [5], one of the
optimizing both its structures and free parameters, such that the earliest methods to propose Lyapunov-based learning for stable
designed controller can guarantee desired stability and can simul- AFLCs, the controller had a serious drawback that it needed
taneously provide satisfactory performance. The design methodol- adaptation of controller parameters for every variation in refer-
ogy for the controller simultaneously utilizes the good features of ence signal, which implied that for temporally varying reference
PSO (capable of providing good global search capability, required
to provide a high degree of automation) and Lyapunov-based tun- signals, the controller parameters would never converge. In [6],
ing (providing fast adaptation utilizing a local search method). it was demonstrated that a modified form of stable AFLC can
Three different variants of the hybrid controller are proposed in be proposed where the controller parameters would converge
this paper. These variants are implemented for benchmark simu- even in the presence of a temporally varying reference signal.
lation case studies and real-life experimentation, and their results However, the controllers in [6] were proposed based on control
demonstrate the usefulness of the proposed approach.
laws adapting a parameter vector containing the positions of the
Index Terms—Adaptive fuzzy logic controllers (AFLCs), hybrid fuzzy sets of the controller outputs only. The basic structure of
approaches, Lyapunov theory, particle swarm optimization (PSO). the FLC was assumed to be a static one. A further improvement
of this scheme was proposed in [12], where a GA-based strategy
I. INTRODUCTION was hybridized with a Lyapunov-based strategy to adapt both
controller structures and free parameters, containing center lo-
UZZY logic controllers (FLCs) evolved as a popular do-
F main of control system and control engineering more than
two decades ago [17]. In recent times, more and more attention is
cations of input membership functions (MFs), scaling gains,
learning rate, etc., in addition to the positions of the singletons
describing output fuzzy sets of the FLC. But the problem asso-
being paid to systematic design of FLCs from the point of view
ciated with this hybrid model is that this GA-based Lyapunov
of stable behavior and optimization of performance indexes.
method needs several thousands of fitness evaluations to achieve
Several such stable and optimized adaptive FLCs have been
parameter convergence and guarantee closed-loop stability.
proposed utilizing, e.g., neural network [1], Lyapunov meth-
This paper proposes a systematic design of stable AFLCs
ods [2], [21], genetic algorithm (GA) [3], [4], ant colony opti-
employing a hybridization of the Lyapunov-strategy-based ap-
mization [14], etc.
proach (LSBA) and particle swarm optimization (PSO) based
Lyapunov-based strategies have been utilized in conventional
approach, called the PSO-based approach (PSOBA) [8], [9],
control theory for a long time to design controllers that can
[11]. It attempts to combine strong points of both methods to
guarantee closed-loop stability of the system [21]–[24]. In [5]
evolve a superior method. The objective of this design strategy
and [6], two of the earliest design methods of Lyapunov-based
is to perform simultaneous adaptation of both the FLC struc-
direct adaptive fuzzy controllers were proposed. The main prob-
ture [15], [22] and its free parameters, so that two competing
lem in Lyapunov-based method is that it requires a considerable
requirements can be fulfilled: 1) to guarantee stability of the con-
amount of time to track the reference input and also the inherent
troller design and 2) to achieve very high degree of automation
approximation error in striving to achieve an unknown theoreti-
in the process of controller design by employing a global search
method. The LSBA helps to guarantee closed-loop stability with
Manuscript received November 29, 2007; revised September 15, 2008; ac- improved speed of convergence. On the other hand, the PSOBA
cepted December 1, 2008. First published January 6, 2009; current version pub-
lished April 1, 2009. This work was supported by the University Grants Commis- helps to achieve desired automation for the design methodology
sion, India, under Major Research Project Scheme [Grant 32-118/2006(SR)]. of the AFLC by getting rid of many manually tuned parameters
K. D. Sharma is with the Department of Electrical Engineering, Future Insti- in LSBA design procedure. However, a well-known drawback of
tute of Engineering and Management, West Bengal University of Technology,
Kolkata 700150, India, and also with Jadavpur University, Kolkata 700032, these global optimization techniques is that they are essentially
India (e-mail: kaushikdassharma@yahoo.com). slow in their search procedure to optimize a criterion function.
A. Chatterjee and A. Rakshit are with the Department of Electrical Engineer- This motivated us to choose PSO among these available stochas-
ing, Jadavpur University, Kolkata 700032, India (e-mail: cha_ami@yahoo.co.in;
anjan_rakshit@yahoo.com). tic methods because of their simple functioning, and generally,
Digital Object Identifier 10.1109/TFUZZ.2008.2012033 they can provide relatively faster, satisfactory solution.
1063-6706/$25.00 © 2009 IEEE
In this paper, three types of hybridization of locally opera- where w(t) is the excitation signal input to the reference
tive LSBA and a global version of PSOBA have been proposed model, and the state vector of the reference model is xm =
(n −1)
and simulated for benchmark nonlinear plants and also tested (xm , ẋm , . . . , xm )T ∈ Rn .
for a real-life experiment. These approaches are called hybrid- The control objective is to force the plant output y(t) to follow
adaptation-strategy-based approaches (HASBA). The first two a given bounded reference signal ym (t) under the constraints
hybrid approaches are the cascade and concurrent combinations that all closed-loop variables involved must be bounded to guar-
of LSBA and PSOBA. In cascade hybrid approach, the structure antee the closed-loop stability of the system. Thus, the tracking
of the controller and the values of the free parameters for LSBA error e = ym − y.
are determined by the PSOBA, i.e., the tedious manual tuning
part of the conventional LSBA is replaced. So, the drawbacks
of conventional LSBA are not fully taken care of. In concurrent A. Control Objective
hybrid approach, the LSBA performs within the supervision of Our objective is to design a stable adaptive fuzzy controller
PSOBA, with the objective of keeping the inherent approxima- for the system described earlier. AFLCs have evolved as popular
tion error limited and the tracking of the input reference signal, control solutions for those classes of plants whose input–output
may be a temporal one, and is expected to be good enough. characteristics are not precisely known. In case of the system in
In some cases, our experimentations show that PSOBA alone (1), this is given by the condition when f (·) and b are not pre-
performs better than the concurrent hybrid approach. Hence, we cisely known, which is widely prevalent in practical situations.
have proposed a third hybrid strategy, called preferential hybrid Here, we need to find the structure of the fuzzy controller as
approach. In this case, the movements of the particles in the well as a feedback control strategy u = u( x| θ), using a fuzzy
multidimensional global search space are sometimes governed logic system and an adaptive law for adjusting the parameter
by the concurrent hybrid method and sometimes by the PSOBA vector θ such that the following conditions are satisfied.
method, depending on the performance of the corresponding The closed-loop system must be globally stable in the
fitness evolutions. Therefore, the best candidate solution in this sense that all variables, x(t), θ(t), and u( x| θ) must be uni-
preferential hybrid approach evolves from a second level of hy- formly bounded, i.e., |x(t)| ≤ Mx < ∞, |θ(t)| ≤ Mθ < ∞,
bridized process that, according to our knowledge and belief, and |u( x| θ)| ≤ Mu < ∞, where Mx , Mθ , and Mu are set by
is a novel idea to design direct AFLCs. The proposed hybrid the designer [5].
approaches have all been implemented for different benchmark The tracking error e(t) should be as small as possible under
case studies and also experimented on a real-life problem, and the constraints in (1) and the e − θ space should be stable in the
the three variants of HASBA are compared with controllers large sense for the system [6].
designed by LSBA and PSOBA alone. The obtained results il- Now, the ideal control law for the system in (1) and (2) is
lustrate, on the whole, the superiority of the preferential hybrid given as [2], [12]
approach over the other approaches.
The rest of the paper is organized as follows. Section II 1
u∗ = −f (x) + ym
(n )
+ kT e (3)
presents a discussion about the design of stable adaptive fuzzy b
controllers. Section III presents the conventional adaptation (n )
strategies and the newly proposed hybrid adaptation strategies, where ym is the nth derivative of the output of the refer-
while in Section IV, simulation and experimental case studies ence model, e = (e, ė, . . . , e(n −1) )T is the error vector, and
for both conventional and proposed hybrid adaptation strategies k = (k1 , k2 , . . . , kn )T ∈ Rn is the vector describing the desired
are demonstrated. Section V concludes the paper. closed-loop dynamics for the error.
This definition implies that u∗ guarantees perfect tracking,
i.e., y(t) ≡ ym (t) if Lt e(t) = 0. It indicates that as t ap-
II. DESIGN OF STABLE ADAPTIVE FUZZY CONTROLLERS t→∞
proaches infinity, e(t) tends to zero, and under that condition,
Let us consider that our objective is to design an adaptive strat- output y(t) becomes equal to the reference ym (t) and perfect
egy for an nth-order single-input–single-output (SISO) nonlin- tracking is achieved.
ear plant given as In practical situations, since f and b are not known precisely,
the ideal u∗ of (3) cannot be implemented in real practice. Thus,
x(n ) = f (x) + bu a suitable solution can be to design a fuzzy logic system to
(1)
y=x approximate this optimal control law.
Now, to ensure stability, we assume that the control u(t) is
where f (·) is an unknown continuous function, u ∈ R and given by the summation of a fuzzy control uc ( x| θ) and an
y ∈ R are the input and output of the plant, and b is an unknown additional supervisory control us (x) [2], [5], [25]. Thus, u(t) is
positive constant. We assume that the state vector is given as given as
x = (x1 , x2 , . . . , xn )T = (x, ẋ, . . . , x(n −1) )T ∈ Rn . The sys-
tem under control has a reference model given as u(t) = uc ( x| θ) + us (x). (4)
(n )

xm = f (xm , w) Let us assume that the AFLC is constructed using a zero-
(2)
ym = xm order Takagi–Sugeno (TS) fuzzy system. Then, uc ( x| θ) for the
DAS SHARMA et al.: HYBRID APPROACH FOR DESIGN OF STABLE ADAPTIVE FUZZY CONTROLLERS USING LYAPUNOV THEORY AND PSO 331
Fig. 1. Flowchart representation of LSBA algorithm. k: plant simulation counter; PST: plant simulation time; ∆tc : controller simulation sampling time.
AFLC is given in the form III. ADAPTATION STRATEGIES FOR THE AFLC
N A. Lyapunov-Strategy-Based Approach
l=1 θl ∗ αl (x)
uc ( x| θ) = = θT ∗ ξ(x) (5)
N
α (x) Design methodologies for a stable direct adaptive fuzzy con-
l=1 l
troller, based on the Lyapunov method, are given in [5], [6],
and [21]. The flowchart representation of LSBA is shown in
1 θ2 · · · θN ] = the vector of the output singletons,
T
where θ = [θr l Fig. 1. The design methodology can be divided into two parts:
αl (x) = i=1 µi (xi ) = the firing degree of rule “l”, N is the
1) offline calculations and 2) online adaptation.
total number of rules, µli (xi ) is the membership value of the
The offline calculation deals with the choice of the val-
ith input MF in the activated lth rule, ξ(x) = vector contain-
ues of k. A typical method of specifying k is given in
ing normalized firing strength of all fuzzy IF–THEN rules =
[6]. The P matrix is chosen with the constraint P > 0.
(ξ1 (x), ξ2 (x), . . . , ξN (x))T , and
The adaptation of θ is carried out using (7), if (|θ| < Mθ )
αl (x) or(|θ| = Mθ and eT pn θT ξ(x) ≤ 0). Otherwise, a projection
ξl (x) = . (6) operator has to be used for adapting θ [5].
N
l=1 αl (x)
Let us define a quadratic form of tracking error as Ve =

1/2eT P e where P is a symmetric positive definite matrix sat- B. PSO-Based Approach (PSOBA)
isfying the Lyapunov equation [2]. With the use of us (x), it
can be shown that V̇e ≤ −1/2eT Qe ≤ 0, where Q is a positive PSO, originally developed by Kennedy and Eberhart in 1995,
definite matrix [5], [13]. is a population-based swarm algorithm [7], [8], [16]. In the
Thus, as P > 0, boundedness of Ve implies the boundedness PSO computational algorithm, population dynamics simulates
in x. Hence, the closed-loop stability is guaranteed. bioinspired behavior, i.e., a “bird flock’s” behavior that involves
The zero-order T–S-type fuzzy control uc (x|θ) can be so sharing of information and allows particles to take profit from
constructed that it will produce a linear weighted combination their own discoveries and previous experience (pBest—personal
of adapted parameter vector θ. Thus, a simple singleton-based best) as well as that of all other particles (gBest—global best)
adaptation law as proposed in [2] and [12] can be given as during the search for food.
To design a direct adaptive fuzzy controller with completely
θ̇ = νeT pn ξ(x) (7) adaptive controller structure and free parameters, we have to
choose the values of the scaling gains, the settings of controller
where ν > 0 is the adaptation gain or learning rate and pn is the structure (i.e., number of MFs in each variable, number of rules,
last column of P . etc.), the supports of input MFs, the positions of output single-
It should be noted here that when the reference signal ym tons, etc. The PSOBA design methodology not only tunes the
is the output of a reference model, excited with a given input singletons properly but it can also optimize the values of the scal-
signal w, and the reference plant has the same dynamics as that ing gains and can determine the optimum controller structure.
of the controlled plant and can be expressed by k, then the input In PSOBA design, a particle’s position in solution space is a
states include w along with x [6], [12]. vector containing all required information to construct a fuzzy
controller [12] as lated and the particles’ position and velocity in the search space
are updated by the global version PSO method [7], [8], [10], as
X= shown in the flowchart representation of the PSOBA algorithm
[structural flags for MFs | center locations of the MFs | in Fig. 2. The PSOBA algorithm will stop searching the solution
space when the number of iterations specified by the designer
scaling gains | learning rate | position of the output singletons]. is reached or a prespecified error is attained by the controller.
(8) C. Hybrid-Adaptation-Strategy-Based Approach
Each structural flag bears the information about the existence In hybrid design methods, the complementary features of
or nonexistence of an MF in its corresponding variable. It can LSBA and the PSOBA are combined in three different ways
only take binary values where 0 indicates the nonexistence and to get a superior systematic design procedure of direct AFLCs
1 indicates the existence of the corresponding MF. However, for a class of SISO systems. The proposed hybrid strategies are
PSOBA is an algorithm where each entry in the particle vector described now.
can take continuous values. Hence, for each structural flag in 1) HASBA-Concurrent Model: In this design process, the
the vector, the universe of discourse is set as [0, 1] and the flag LSBA and the PSOBA run concurrently or in parallel to op-
is set to 0 if the continuous value of the variable is less than 0.5 timize the: 1) structure of the controller; 2) scaling gains; 3)
and the flag is set to 1 if the variable is greater than or equal to learning rate (or adaptation gain); and 4) positions of the output
0.5. For each MF of an input variable, there is a corresponding singletons.
structural flag assigned. A specific MF for a given input vari- In this method, a particle X is divided into two subgroups,
able will only become active when its corresponding structural unlike PSOBA design methodology, and is given as
flag is set to 1. When this flag becomes 0, in any given itera-
X = [ Ψ| θ] (9)
tion, the corresponding MF becomes inactive and all rules, in
the fuzzy rule base, containing that MF in antecedent clauses, where
become inactive. The total number of 1’s in the structural flags
Ψ = [structural flags for MFs | center locations of
keeps changing in each iteration, and hence, the total number
of MFs in which an input variable is fuzzified also changes the MFs | scaling gains | learning rate]
on each iteration. This changes the structure of the AFLC in
θ = [position of the output singletons]. (10)
each iteration, thus changing the total number of rules of the
fuzzy rule base, and hence, changes the total number of ac- The partition of X vector separates the parameters accord-
tive output singletons. Each input is fuzzified using triangular ingly as Ψ has nonlinear influence and θ has linear influence on
MFs and the center locations of the MFs in the particle vector uc , respectively [12]. In LSBA, the values of the free parameters
store the peaks of the MFs. Each input is fuzzified in the range in the vector Ψ are defined a priori and the value of the corre-
[−1, 1] with two fixed triangular MFs having their peaks fixed sponding θ minimizing the tracking error can be obtained by the
at −1 and 1, respectively. All intermediate MFs for that input adaptation law as in (7). Here, Ψ is set by a hand-tuned trial-and-
variable are flexible in nature. They can be either active or in- error method. In this paper, PSOBA is applied to optimize Ψ and
active during an iteration and their peaks are also adapted by θ in tandem by updating the particle’s position, containing both
PSOBA in each iteration. For each MF, the peaks of the imme- Ψ and θ, as shown in the flowchart representation of PSOBA in
diate adjacent active MFs on either side of its own peak form Fig. 2. In addition to that, adaptation law, as in (7), is applied to
the left and right base support of it. Whether the immediate ad- every updated candidate controller to adjust the value of θ only.
jacent MF is an active MF or not is determined by the content So, in this method, θ explores both the local search space and
of its corresponding structural flag. Hence, some of the center global search space simultaneously, in a twofold manner [29].
locations of MFs are ignored while evaluating the AFLC out- The flowchart representation of HASBA-concurrent model is
put, because their corresponding entries in the structural flags shown in Fig. 4.
are zero. A similar logic holds true for the output singletons in 2) HASBA-Cascade Model: In this model, LSBA and
a particle vector. Here, those singletons become inactive whose PSOBA optimizations are used in cascade. Therefore, in this
antecedent parts contain one or more input MFs that become model, the structure [22] and the free parameters of the con-
inactive in that iteration. troller are optimized by the PSOBA method as well as the po-
To design an AFLC utilizing PSOBA algorithm first, the pop- sitions of the output singletons are primarily tuned in global
ulation of particles in the swarm is chosen, then the algorithm search space by the PSOBA method. So, after a primary global
will decide the structure of the candidate controller according to tuning of θ by the PSOBA method, LSBA optimizes θ locally
its structural flags. The candidate controller simulation (CCS) to further improve the performance of the controller. Thus, in
algorithm, as shown in Fig. 3, is carried out on the candidate this hybrid model also, the LSBA method is enhanced by the
controller to calculate the fitness function, the IAE, which can PSOBA method. However, the basic difference lies in the fact
be defined as IAE = PST n =0 e(n)τ , where PST = plant simula- that the LSBA algorithm is applied to the best solution vector de-
tion time and τ = step size or sampling time. Then, according termined by the PSOBA algorithm. Hence, the LSBA algorithm
to the value of the fitness function IAE of each particle in each runs only once at the completion of the PSOBA algorithm. It is
iteration, the pBest and gBest candidate controllers are calcu- in contrast to the HASBA-concurrent model, where the LSBA
Fig. 2. Flowchart representation of PSOBA algorithm. j: particle counter; NOP: number of particles (population size); g: PSO iteration counter; iterm a x :
maximum number of PSO iterations.
Fig. 3. Flowchart representation of candidate controller simulation (CCS) algorithm. k: plant simulation counter; PST: plant simulation time; ∆tc : controller
simulation sampling time.
algorithm runs as a submodule of the PSOBA algorithm and at- The flowchart representation of HASBA-cascade is shown in
tempts to locally fine-tune each candidate solution vector, called Fig. 5.
“particle.” In the cascade model, the learning rate is adjusted 3) HASBA-Preferential Model: In this proposed model, both
in such a fashion that LSBA adaptation searches only the near the PSOBA method and the HASBA-concurrent method are em-
neighborhood of the best θ produced by the PSOBA method. ployed to keep the tracking error as minimum as possible. Here,
Fig. 4. Flowchart representation of HASBA—concurrent model algorithm. j: particle counter; NOP: number of particles (population size); g: PSO iteration
counter; iterm a x : maximum number of PSO iterations.
each particle in each iteration, so that the same particle may fol-
low different update rules in different iterations. The flowchart
representation of HASBA-preferential is shown in Fig. 6.
IV. SIMULATION AND EXPERIMENTAL STUDIES

To evaluate the effectiveness of the proposed schemes, we
considered four benchmark case studies, previously utilized
in several research works and a real-life experiment. The pro-
cess models are simulated each using a fixed-step fourth-order
Runge–Kutta method with sampling time T = 0.01 s.
Fig. 5. Flowchart representation of HASBA—cascade model algorithm.

A. Case Study I
The controlled plant under consideration is a second-order dc
each particle in the swarm, i.e., each candidate controller, eval- motor containing nonlinear friction characteristics described by
uates the CCS algorithm and the LSBA algorithm separately at the following model [6], [12]:
the same time. The fitness values of the candidate controller, 
evaluated by employing the CCS algorithm and the LSBA al- ẋ1 = x2 

gorithm separately, are then compared, and then, the algorithm ẋ2 = − f (xJ 2 ) + CJT u (11)
showing a better performance is taken as the guiding factor for 

y = x1
updating the velocity and the position of that specific particle
in that specific iteration, according to the PSO method. If the where y = x1 is the angular position of the rotor (in radians),
LSBA algorithm shows a better performance for that particle in x2 is the angular speed (in radians per second), and u is the
that iteration, then the system will follow the algorithm for the current fed to the motor (in amperes). The plant parameters
HASBA-concurrent model; otherwise, the system will follow are CT = 10 N·m/A, J = 0.1 kg·m2 , and the nonlinear friction
the PSOBA model in that iteration. Thus, in this process, the torque is defined as
updates for each particle in a swarm switch between two algo-
rithms, even within the same iteration, to search the solution
space. This process of evaluation is performed separately for f (x2 ) = 5 tan−1 (5x2 ) N · m. (12)
Fig. 6. Flowchart representation of HASBA—preferential model algorithm. j: particle counter; NOP: number of particles (population size); g: PSO iteration
counter; iterm a x : maximum number of PSO iteration.
The control objective is to make the angular position y follow The controller is trained by applying a signal w(t) that
a reference signal given by changes its value randomly in the interval (−1.5, 1.5) in ev-
ery 0.5 s. This signal is chosen as it contains infinitely many
ÿm = 400w(t) − 400ym − 40ẏm (13) frequencies and leads to good exploration of controller input
(i.e., the reference model has a double pole at s1,2 = −20), space [6].
where w(t) is a square wave of random amplitude, as shown in The plant is simulated for 210 s. The adaptation starts at
Fig. 7(a), and ym is the output of w(t) filtered by the second- t = 0 s and continues up to t = 200 s when the adaptation is
order linear filter [given by (13)] as shown in Fig. 7(b). switched off. Then, the plant is tested for 10 s with this reference
The control signal obtained as the output from the fuzzy signal. During this period, adaptation is turned off (i.e., ν = 0)
controller, in response to input signals, is given as follows: and the controller operates with already adapted parameters.
b) PSOBA: In this case, all the parameters of X = [Ψ | θ]
uc = uc (x, w | X ), for PSOBA and HASBA are optimized by the PSOBA. To perform this simulation, a pop-
(14)
uc = uc (x, w | θ ), for LSBA. ulation size of ten particles is taken. For each simulation, 200
iterations of PSOBA is set. In fixed structure controller design, a
1) Comparison of Design Strategies: In this section, the fol- fixed number of evenly distributed input MFs are used for initial
lowing design strategies are compared. population of PSOBA. The PSOBA algorithm further tunes only
a) LSBA: In this case, only θ is optimized and the static the free parameters and the positions of the output singletons
Ψ is manually set a priori. In our simulation, input MFs are to minimize the tracking error. In variable structure controller
distributed evenly on the corresponding universe of discourse. design, the structural flags, are utilized to automatically vary
The other parameters are chosen as the structure of the candidate controller in each iteration. The
input MFs and the corresponding position of the output single-
p 0 p1 5 0.01
P = = tons are set according to the active structural flags, as stated in
p2 p2 0.01 0.001
Section III-B. IAE between the reference signal ym and the
k = [ 400 40 ] motor output y is taken as the fitness function for optimization.
c) HASBA: It consists of the following models.
such that p1 k1 > 0 and p2 > p1 /k2 . The value of adaptation i) Concurrent hybrid model: In this simulation pro-
gain is chosen as ν = 5, giving a good tradeoff between adap- cess, the PSOBA-based optimization of parameters
tation rate and small amplitudes of parameter oscillations that for X = [Ψ | θ] and the LSBA algorithm, which only
are due to inherent approximation error. The other manually adjusts θ, run concurrently. The position vector X
set parameters are b = 100, bL = 95, and f U = 1.5. The in- now additionally includes the adaptation gain ν in its
put and output scaling gains of the AFLC are chosen so as to Ψ portion. The gain ν is varied over the range 14 000–
map the ranges of the corresponding variables into the interval 16 000. The structural flags are set and reset in the
[−1, 1].
TABLE I
COMPARISON OF SIMULATION RESULTS OF CASE STUDY I
(EVALUATION TIME = 10 S)
The simulations are carried out for three fixed structures

of 3 × 3 × 3, 5 × 5 × 5, and 7 × 7 × 7 input MFs and a vari-
able structure with a potentially biggest possible structure of
7 × 7 × 7 MFs. Table I presents comparison of three possible
HASBA controllers vis-à-vis LSBA and PSOBA controllers for
case study I. Fig. 8 shows temporal variations of system re-
sponses for the sample case of fixed structure controllers with
Fig. 7. (a) Input signal w(t). (b) Reference signal y m (t).
5 × 5 × 5 input MFs. Fig. 9(a) shows the temporal variations of
system responses of the variable structure HASBA-preferential
same manner as is done in PSOBA and utilized to model for 100 intermediate iterations. In Fig. 9(b)–(d), the sys-
automatically find out the structure of the candidate tem response for 10 s evaluation phase, the control effort, and the
controller. error plot for this AFLC are shown as a sample case. For fixed
ii) Cascade hybrid model: This simulation process starts structure controllers, it can be seen that the performances of
with the PSOBA simulation. The result obtained from the controllers (demonstrated by the performance index of IAE)
PSOBA is then passed through the Lyapunov-based improve with increase in the number of rules from a 3 × 3 × 3
adaptation strategy to further adjust the positions of to 5 × 5 × 5 configuration. However, for a 7 × 7 × 7 configu-
the output singletons θ. The adaptation gain ν during ration, results are not better than that for a 5 × 5 × 5 configura-
the LSBA is set to 5 to explore the near neighborhood tion. Hence, we can conclude that merely increasing the num-
of the best θ as determined by the PSOBA. ber of rules in fuzzy controllers does not necessarily improve
iii) Preferential hybrid model: In this hybrid simula- their performance. This is in conformation with the observa-
tion process, both concurrent hybrid model and pure tions presented in [12]. For a given controller MF configuration
PSOBA model are implemented for the same particle (e.g., 5 × 5 × 5 or 7 × 7 × 7), among the competing controllers,
separately at the same time, and the candidate con- the HASBA-preferential scheme shows the best performance.
trollers, i.e., the particles of PSO, are updated accord- Even for the variable structure scheme, HASBA-preferential
ingly. The same particle can take different searching controller emerges as the best possible solution. Although the
algorithms to search the solution space depending on IAE value is higher for the variable structure case, it can be
the value of its fitness function IAE during the entire seen that the best result, in terms of IAE, is obtained with only
optimization process. 100 rules. This is a significant reduction in the total number of
The results of simulation studies are tabulated in Table I. In rules and such a simplified structure of an FLC will be very
all these simulations, the plant is evaluated for 10 s after the useful from the point of view of implementation in real life.
training is completed by different adaptation strategies so as to Although the HASBA-concurrent model could converge to an
compare the results. even smaller number of rules (80, in this case), a comparatively
Fig. 8. Responses of the fixed structure (MF: 5 × 5 × 5) adaptive fuzzy controllers of case study I. (a) LSBA. (b) PSOBA. (c) HASBA-concurrent.
(d) HASBA-cascade. (e) HASBA-preferential.
Fig. 9. Responses of the variable structure HASBA—preferential adaptive fuzzy controller of case study I. (a) Training phase response after 100 PSO iterations.
(b) Evaluation phase response. (c) Evaluation phase control effort. (d) Evaluation phase error plot.

much larger IAE prompted us to conclude that the HASBA- hence, bL = 0.95 and f U = 12 + x31 , K = [2 1], [18]. The
preferential gave the overall superior performance. LSBA, PSOBA design strategies, and different hybrid meth-
B. Case Study II ods are simulated in a similar fashion as in the earlier case
study.
In this case study, the Duffing’s forced oscillation system, The adaptation gain ν is set to 2 for LSBA and HASBA-
which is itself a chaotic system if unforced, is considered and is concurrent, 0.5 for HASBA-cascade, and 1 for HASBA-
given as  preferential. In a similar fashion of case study I, the simulations
ẋ1 = x2 
 are carried out for three fixed structures of 3 × 3, 5 × 5, and
ẋ2 = −0.1x2 − x1 + 12cos(t) + u(t) .
3
(15) 7 × 7 input MFs and a variable structure. In this case study also,


y = x1 the HASBA-preferential design strategy emerges as the best
solution, for both fixed structure and variable structure config-
The control objective is to track the reference signal ym ,
urations. In fact, for this case study, the best performance result
where ym = sin(t). The control signal fed to the fuzzy controller
for HASBA-preferential controller for variable structure config-
is uration is obtained with simultaneous achievement of minimum
uc = uc (x |X ), for PSOBA and HASBA
. (16) number of rules (here, it is 15). Hence, this design strategy
uc = uc (x |θ ), for LSBA
evolves as the most superior solution from both the points
1) Comparison of Design Strategies: The choice of param- of view of best performance and simplest implementation.
eters for this simulation process is: Table II shows the results for sample case of variable struc-
ture controllers. Fig. 10 shows the response of variable-structure
1 0.01
P = HASBA-preferential controller for case study II as a represen-
0.01 0.001
tative manner.
b=1
TABLE II TABLE III

COMPARISON OF SIMULATION RESULTS OF CASE STUDY II COMPARISON OF SIMULATION RESULTS OF CASE STUDY III
(EVALUATION TIME = 10 S) (EVALUATION TIME = 10 S)
Fig. 11. Evaluation phase response of the variable structure HASBA—

Fig. 10. Evaluation phase response of the variable structure HASBA— preferential adaptive fuzzy controller of case study III.
preferential adaptive fuzzy controller of case study II.
that, very rarely, a single controller design strategy emerges as
the best solution for many types of systems under several types
C. Case Study III of input conditions. Table III shows the results for a sample case
In this case study, the Duffing’s forced oscillation system with of variable structure controllers. Fig. 11 shows the response of
disturbance is considered and is given as the variable-structure HASBA-preferential controller for case
 study III in a representative manner.
ẋ1 = x2 

ẋ2 = −0.1x2 − x31 + 12cos(t) + u(t) + d(t) (17) D. Case Study IV


y = x1 All the discussions presented so far are given with reference to
the system in (1), where the control input is given in “bu” form.
where the external disturbance d(t) is a square wave of random The discussions hold equally for a more general system, where
amplitude within the range [−1, 1] and a period of 0.5 s. The control input is given in “g(x)u,” as demonstrated in [6]. Now
control objective is to track the reference signal ym , where we demonstrate the utility of our proposed controllers for such
ym = sin(t). The control signal fed to the fuzzy controller is a system. We consider the inverted pendulum system [26], [28]
the same as in (16). as
1) Comparison of Design Strategies: Here also, we choose 
the same parameters for this simulation process as was cho- ẋ1 = x2 

sen in case study II. The design strategies of LSBA, PSOBA, ẋ2 = f (x) + g(x)u(t) (18)


and different hybrid methods are simulated in a similar fash- y = x1
ion as in case study II. The adaptation gain ν is set to 2 for
LSBA and HASBA-concurrent, 0.5 for HASBA-cascade, and where
1 for HASBA-preferential. In this case study, among the fixed mlx2 sin x1 cos x1 − (M + m)g sin x1
f (x) =
structure controllers, the HASBA-concurrent shows least av- ml cos2 x1 − 4/3 l(M + m)
erage IAE value for 7 × 7 configuration and PSOBA shows and
best performance for variable configuration. However, HASBA- − cos x1
preferential configuration emerges as the second best solution g(x) =
ml cos2 x1 − 4/3 l(M + m)
in each configuration. It also achieves the minimum number of
rules for the variable configuration (along with HASBA-cascade where g = 9.8 m/s2 , m = 0.1 kg, M = 1 kg, and l = 0.5 m
and PSOBA controllers). This is understandable from the fact [26], [28]. The control objective is to track the reference signal
TABLE IV
COMPARISON OF SIMULATION RESULTS OF CASE STUDY IV
(EVALUATION TIME = 10 S)
Fig. 12. Evaluation phase response of the variable structure HASBA—

preferential adaptive fuzzy controller of Case Study IV.
ym , where ym = sin(t). The control signal fed to the fuzzy

controller is the same as in (16).
1) Comparison of Design Strategies: The choice of param-
eters for this simulation process is

1 0.01
P =
0.01 0.001
b = g(x)
hence, bL = b – 0.1 and f U = 5, K = [2 1]. The design strate-
gies of LSBA, PSOBA, and different hybrid methods are simu-
lated in a similar fashion as in earlier case studies. The adaptation
gain ν is set to 5 for LSBA, 0.001 for HASBA-concurrent and
HASBA-preferential, and 0.01 for HASBA-cascade. The main
difference of this case study is that, here, the coefficient of the
control input u(t) is not a constant like in earlier cases, but is
a function of the states of the inverted pendulum system and is
varying with time. Table IV and Fig. 12 show results in tabular
and graphical form for case study IV in a similar manner as in
Fig. 13. (a) Real-life experimental arrangements. (b) and (c) Evaluation phase
case studies II and III. It is seen from Table IV that the IAE of response of HASBA—preferential adaptive fuzzy controller of Case Study IV
the HASBA-preferential model is close to that of the PSOBA with different reference signals.
and HASBA-concurrent model but the number of fuzzy rules is
minimum (15 in this case). So, the performance of the HASBA-
E. Case Study V
preferential model is good enough to justify it as a superior
control strategy over the other control strategies. To demonstrate the usefulness of the proposed algorithms,
Overall, from our four simulation case studies with different we have performed a real-life experiment where we attempt to
types of temporal reference input variations, it can be concluded perform the speed control of a real dc motor. Fig. 13(a) shows
that the HASBA-preferential controller emerges as a superior, the experimental arrangement where the PC-based controller
feasible control solution. controls the dc motor through a parallel port interface in real
TABLE V V. CONCLUSION
COMPARISON OF SIMULATION RESULTS OF CASE STUDY V
(EVALUATION TIME = 240 S) This paper has proposed a new hybrid approach for designing
the AFLCs employing both the conventional Lyapunov theory,
to locally optimize the positions of the output singleton, and
PSO, which is a stochastic optimization technique. In this pro-
posed hybridization process, while LSBA optimizes the posi-
tions of the output singleton, PSOBA is utilized to design the
structure of the AFLC and also evaluate the free parameters of
it. Three variants of the hybrid models have been proposed and
simulated for the benchmark case studies. The preferential hy-
brid model was evolved as a superior technique as compared to
time. The scheme employs a power-transistor-based chopper- the LSBA, PSOBA, and other hybrid models. The stability of
controlled mechanism to achieve the desired speed in armature the closed-loop system and the convergence of the plant output
control mode. At first, the motor parameters, i.e., armature re- to a desired reference signal are guaranteed precisely. The main
sistance (Ra ), armature inductance (La ), inertia constant (J), advantages of these proposed methods are that it requires no a
damping constant (B), etc., are identified by exciting the motor priori knowledge about the controlled plant and the approxima-
in open loop and collecting the input–output data. Here, the dc tion error is reduced greatly. Moreover, the preferential hybrid
motor characteristics are modeled as model is able to avoid the local minima. In this paper, we have
chosen PSO as the representative stochastic global optimization
algorithm for hybridization with Lyapunov theory. The ratio-

ẋ1 = x2


nale behind this is that PSO is a relatively simple algorithm
among stochastic optimization algorithms. However, it should
ẋ2 = BJ LRaa + Kb KT
x1 − Ra
+ B
x2 + K KT
u(t)
J La La J J La 
 be pointed out that other stochastic global algorithms like GA,
y = x1 differential evaluation, etc. [12], [19], [20], [27], can also be
(19)
potentially employed to design similar AFLCs. The authors in-
where [x1 x2 ] = [ω ω̇], the output y is the angular speed of the
tend to design similar AFLCs using other global optimization
motor (in radians per second), K is the constant of the chopper
algorithms and compare their performances in the near future.
circuit, and Kb and KT are the constants of the dc motor. The
unknown parameters are learned utilizing PSO, where the PSO
algorithm determines the best vector comprising the unknown REFERENCES
parameters for which discrepancy between the model output and [1] P. He and S. Jagannathan, “Reinforcement learning neural network based
the actual experimental output data, for the same input, is min- controller for non-linear discrete time systems with input constraints,”
imum, considering the entire set of input–output real-life data. IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 37, no. 2, pp. 425–436,
Apr. 2007.
When the state model of this real dc motor is identified, our con- [2] S. H. Zak, Systems and Control. New York: Oxford Univ. Press, 2003.
troller design is carried out offline for different control strategies [3] F. Cupertino, E. Mininno, D. Naso, B. Turchiano, and L. Salvatore, “On-
as discussed earlier. Fixed structure controllers with 5 × 5 in- line genetic design of anti-windup unstructured controllers for electric
drives with variable load,” IEEE Trans. Evol. Comput., vol. 8, no. 4,
put MFs are implemented for this real-life experimentation for pp. 347–364, Aug. 2004.
controlling the same motor under temporally varying reference [4] C. L. Chiang and C. T. Sub, “Tracking control of induction motor using
speed conditions for 240 s. Fig. 13(b) and (c) shows the perfor- fuzzy phase plane controller with improved genetic algorithm,” Electr.
Power Syst. Res., vol. 73, pp. 239–247, 2005.
mance of HASBA-preferential controller, as sample cases, in [5] L. X. Wang, “Stable adaptive fuzzy control of nonlinear system,” IEEE
tracking two different temporally varying reference speed sig- Trans. Fuzzy Syst., vol. 1, no. 2, pp. 146–155, May 1993.
nals. Table V shows the sample performances of the controllers [6] K. Fischle and D. Schroder, “An improved stable adaptive fuzzy control
method,” IEEE Trans. Fuzzy Syst., vol. 7, no. 1, pp. 27–40, Feb. 1999.
in the real implementation phase for the dc motor, for the case [7] R. C. Eberhart and J. Kennedy, “A new optimizer using particle swarm
as shown in Fig. 13(b). In each of these experimentations, the theory,” in Proc. 6th Int. Symp. Micromach. Hum. Sci., Nagoya, Japan,
speed measurement is carried out using shaft encoders having 1995, pp. 39–43.
[8] J. Kennedy and R. C. Eberhart, “Particle swarm optimization,” in Proc.
ten holes in the wheel mounted on the shaft. Then, the speed sig- IEEE Int. Conf. Neural Netw., 1995, vol. 4, pp. 1942–1948.
nal acquired is filtered in the software developed, employing an [9] G. Venter and J. S. Sobieski, “Particle swarm optimization,” presented at
exponential averaging technique, to remove noise pickup. It can the 43rd AIAA Conf., Denver, CO, 2002.
[10] Y. Shi and R. C. Eberhart, “A modified particle swarm optimizer,” in Proc.
be seen that the controllers developed were able to provide sat- IEEE Int. Conf. Evol. Comput., May 1998, pp. 69–73.
isfactory performance. The interrupt-driven software developed [11] R. C. Eberhart and Y. Shi, “Comparing inertia weights and constriction
for this sophisticated motor control scheme has a sampling rate factors in particle swarm optimization,” in Proc. 2000 Congr. Evol. Com-
put., pp. 84–89.
of 100 µs within the interrupt mode, whereas each PWM cycle [12] V. Giordano, D. Naso, and B. Turchiano, “Combining genetic algorithm
generated for control input is of duration 10 ms (i.e., ON-time + and Lyapunov-based adaptation for online design of fuzzy controllers,”
OFF-time). The controller output is used to suitably control this IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 36, no. 5, pp. 1118–1127,
Oct. 2006.
ON-time. Table V, reveals that, in this real-life experiment also,
[13] L. X. Wang and J. M. Mendel, “Three-dimensional structured networks
the HASBA-preferential controller emerges as the best design for matrix equation solving,” IEEE Trans. Comput., vol. 40, no. 12,
strategy among the LSBA, PSOBA, and other hybrid variants. pp. 1337–1346, Dec. 1991.
[14] M. Dorigo, V. Maniezzo, and A. Colorni, “Ant system: Optimization by Amitava Chatterjee received the B.E., M.E., and
a colony of cooperating agent,” IEEE Trans. Syst., Man, Cybern. B, Ph.D. degrees in electrical engineering from Jadavpur
Cybern., vol. 26, no. 1, pp. 29–41, Feb. 1996. University, Kolkata, India, in 1991, 1994, and 2002,
[15] A. A. A. Esmin, A. R. Aoki, and G. L. Torres, “Particle swarm optimization respectively.
for fuzzy membership functions optimization,” in Proc. IEEE Int. Conf. In 1997, he joined the Department of Electrical En-
Syst., Man, Cybern., 2002, pp. 108–113. gineering, Jadavpur University, where he is currently
[16] Y. Shi and R. C. Eberhart, “Parameter selection in particle swarm opti- a Reader. In 2003, he received the Japanese Gov-
mization,” in Evol. Program. VII: Proc. EP 1998. New York: Springer- ernment (Monbukagakusho) Scholarship and went
Verlag, pp. 591–600. to Saga University, Saga, Japan. In early 2004, he
[17] T. J. Procyk and E. H. Mamdani, “A linguistic self-organizing process was invited as a Teacher–Researcher in the Uni-
controller,” Automatica, vol. 15, pp. 15–30, 1979. versity of Paris XII, Val de Marne, France. From
[18] B. S. Chen, C. H. Lee, and Y. C. Chang,, “H∞ tracking design of uncertain November 2004 to November 2005, he was with the University of Electro-
nonlinear SISO systems: Adaptive fuzzy approach,” IEEE Trans. Fuzzy Communications, Tokyo, Japan, on a Japan Society for the Promotion of Science
Syst., vol. 4, no. 1, pp. 32–43, Feb. 1996. (JSPS) Post-Doctoral Fellowship for Foreign Researchers. His current research
[19] R. Storn and K. Price, “Differential evolution—A simple and efficient interests include intelligent instrumentation and control, signal processing, im-
adaptive scheme for global optimization over continuous spaces,” Int. age processing, and robotics. He has authored or coauthored more than 50
Comput. Sci. Inst., Berkley, CA, Tech. Rep., 1995. technical papers, including 30 international journal papers.
[20] K. S. Lee and Z. W. Geem, “A new meta-heuristic algorithm for continuous
engineering optimization: Harmony search theory and practice,” Comput.
Methods Appl. Mech. Eng., vol. 194, no. 36–38, pp. 3902–3933, 2005.
[21] L. X. Wang, A Course in Fuzzy Systems and Control. Englewood Cliffs,
NJ: Prentice-Hall, 1997.
[22] D. L. Tsay, H. Y. Chung, and C. J. Lee, “The adaptive control of nonlinear
systems using the Sugeno-type of fuzzy logic,” IEEE Trans. Fuzzy Syst.,
vol. 7, no. 2, pp. 225–229, Apr. 1999. Anjan Rakshit received the M.E. and Ph.D. degrees
[23] R. Chekkouri, L. Romeral, and M. Moreno, “Direct adaptive fuzzy con- in electrical engineering from Jadavpur University,
troller by changing the output membership functions for nonlinear motion Kolkata, India, in 1978 and 1987, respectively.
applications,” in Proc. IEEE ICIT 2003, Maribor, Slovenia, 2000, pp. 22– He is currently a Professor in the Department
26. of Electrical Engineering, Jadavpur University. His
[24] S. S. Chen, J. L. Wu, S. F. Su, J. C. Chang, Y. C. Chang, and T. T. Lee, current research interests include digital signal pro-
“On the analysis of direct adaptive fuzzy controllers,” in Proc. 3rd Int. cessing, Internet-based instrumentation, smart instru-
Conf. Mach. Learn. Cybern., Shanghai, China, 2004, pp. 4368–4373. mentation, intelligent control, and design of real-time
[25] S. H. Chin and M. J. Er, “Hybrid adaptive fuzzy controllers of robot systems. He has authored or coauthored almost 70
manipulators,” in Proc. 1998 IEEE/RSJ Int. Conf. Intell. Robots Syst., technical papers.
pp. 1132–1137.
[26] T. Wang and S. Tong, “Adaptive fuzzy output feedback control for SISO
nonlinear systems,” Int. J. Innovative Comput., Inf. Control, vol. 2, no. 1,
pp. 51–60, 2006.
[27] R. Toscano and P. Lyonnet, “Stabilization of systems by static output
feedback via heuristic Kalman algorithm,” J. Appl. Comput. Math., vol. 5,
no. 2, pp. 154–165, 2006.
[28] H. X. Li and S. Tong, “A hybrid adaptive fuzzy control for a class of
nonlinear MIMO system,” IEEE Trans. Fuzzy Syst., vol. 11, no. 1, pp. 24–
34, Feb. 2003.
[29] K. D. Sharma, A. Chatterjee, and F. Matsuno, “A Lyapunov theory and
stochastic optimization based stable adaptive fuzzy control methodology,”
in Proc. SICE Annu. Conf. 2008, Univ. Electro-Commun., Aug. 20–22,
pp. 1839–1844.
Kaushik Das Sharma received the B.Sc. (Hons.)

degree in physics, and the B.Tech. and M.Tech. de-
grees in electrical engineering from the University of
Calcutta, Kolkata, India, in 1998, 2001, and 2004, re-
spectively. He is currently working toward the Ph.D.
degree at Jadavpur University, Kolkata.
He is currently an Assistant Professor of the De-
partment of Electrical Engineering, Future Institute of
Engineering and Management, West Bengal Univer-
sity of Technology, Kolkata. His current research in-
terests include fuzzy control system design, stochas-
tic optimization applications, and embedded system design, among others.
Fuzzy Control for Nonlinear Uncertain

Electrohydraulic Active Suspensions
With Input Constraint
Haiping Du and Nong Zhang
Abstract—This paper presents a Takagi–Sugeno (T–S) model- hydraulic actuators in active suspensions. Generally speaking,
based fuzzy control design approach for electrohydraulic active the currently proposed control algorithms for electrohydraulic
vehicle suspensions considering nonlinear dynamics of the actua- active suspensions can be divided into two main groups: one is
tor, sprung mass variation, and constraints on the control input.
The T–S fuzzy model is first applied to represent the nonlinear the two-loop control strategy, in which the outer loop is used to
uncertain electrohydraulic suspension. Then, a fuzzy state feed- provide the desired forces and the inner loop is used to make
back controller is designed for the obtained T–S fuzzy model with the electrohydraulic actuators track the desired forces; the other
optimized H∞ performance for ride comfort by using the parallel- is the sliding-mode-based control strategy. As proved in [11],
distributed compensation (PDC) scheme. The sufficient conditions pure proportional-integral differential (PID)-like controllers are
for the existence of such a controller are derived in terms of linear
matrix inequalities (LMIs). Numerical simulations on a full-car not capable of giving satisfactory performance in the actuator
suspension model are performed to validate the effectiveness of the force tracking problem, and more sophisticated control schemes
proposed approach. The obtained results show that the designed should be employed. Hence, some attempts have been made to
controller can achieve good suspension performance despite the compensate for this shortcoming through advanced inner loop
existence of nonlinear actuator dynamics, sprung mass variation, force control algorithms, for example [11]–[14]. Nevertheless,
and control input constraints.
due to their highly nonlinear dynamics, using electrohydraulic
Index Terms—Electrohydraulic actuator, input constraint, non- actuators to track the desired forces is fundamentally limited
linear dynamic system, Takagi–Sugeno (T–S) fuzzy modeling, un- when interacting with a dynamic environment [11]. On the other
certainty, vehicle active suspension.
hand, the chattering phenomenon is inevitable in sliding mode
control, and it may excite unmodeled high-frequency dynamics,
I. INTRODUCTION which degrades the performance of the system and may even
lead to instability. Techniques such as adaptive fuzzy sliding
CTIVE suspensions are currently attracting a great deal of
A interest in both academia and industry for improving ve-
hicle ride comfort and road holding performance [1], [2]. Since
control [5] and self-organizing fuzzy sliding control [15] were
then proposed to smooth the chattering phenomenon. However,
these approaches need a complicated learning mechanism or a
active suspensions need actuators to provide the required forces,
specific performance decision table, which is designed by a trial
one practical consideration in real-world applications involves
and error process, and presents certain difficulties in application.
choosing appropriate actuators that can fit into the suspension
In practice, the vehicle sprung mass varies with the loading
packaging space, and satisfy the practical power and bandwidth
conditions, such as the payload and number of vehicle occu-
requirements. It has been noted that electrohydraulic actuators
pants. The control performance of a vehicle suspension will be
are regarded as one of the most viable choices for an active sus-
affected if the sprung mass variation is not considered in the
pension due to their high power-to-weight ratio and low cost.
controller design process. In spite of its importance, this prob-
Therefore, in recent years, many studies have focused on elec-
lem has not been explicitly dealt with in any previous studies
trohydraulic active suspensions, and various control algorithms
on electrohydraulic suspensions. Furthermore, the constraint on
have been proposed to deal with the involved highly nonlinear
control input voltage sent to the actuator servo-valve has not
dynamics of electrohydraulic actuators [3]–[10].
yet been considered for controller design, although the input
It is still a challenge to develop an appropriate control strat-
power provided to an electrohydraulic actuator is, in practice,
egy for dealing with the highly nonlinear dynamics of electro-
limited. Thus, it is surely necessary to develop a new controller
design approach that aims at improving the performance of elec-
Manuscript received December 18, 2007; revised June 19, 2008 and trohydraulic active suspensions while considering the actuator
September 11, 2008; accepted December 8, 2008. First published December
31, 2008; current version published April 1, 2009. This work was supported nonlinear dynamics, sprung mass uncertainty, and control input
in part by the Early Career Research Grant, University of Technology, Sydney, voltage limitations.
and by the Australian Research Council’s Discovery Projects funding scheme Following the earlier discussion on electrohydraulic active
under Project DP0773415.
The authors are with the Faculty of Engineering, University of Technology, suspensions, in this paper, a fuzzy state feedback controller
Sydney, N.S.W. 2007, Australia (e-mail: hdu@eng.uts.edu.au; nong.zhang@ design method is presented to improve the ride comfort per-
uts.edu.au). formance of vehicles with electrohydraulic active suspensions
at http://ieeexplore.org. through a Takagi–Sugeno (T–S) fuzzy model approach. In recent
Digital Object Identifier 10.1109/TFUZZ.2008.2011814 decades, fuzzy logic control has been proposed as an alternative
1063-6706/$25.00 © 2009 IEEE
approach to conventional control techniques for complex · refers to either the Euclidean vector norm or the induced ma-
nonlinear systems. It was originally introduced and developed trix 2-norm, I is used to denote the identity matrix of appropriate
as a model-free control design approach, and it has been applied dimensions, and to simplify notation, ∗ is used to represent a
to active suspensions [5], [8], [16] to deal with the nonlinear- block matrix that is readily inferred by symmetry.
ities associated with the actuator dynamics, shock absorbers,
suspension springs, etc. However, the model-free fuzzy logic II. ELECTROHYDRAULIC SUSPENSION MODEL
control suffers from a number of criticisms, such as the lack of
A full-car electrohydraulic suspension model, as shown in
systematic stability analysis and controller design. It also faces
Fig. 1, is considered in this paper. This is a 7-DOF model where
a challenge in the development of fuzzy rules. Recent research
the sprung mass is assumed to be a rigid body with freedoms
on fuzzy logic control has, therefore, been devoted to model-
of motion in the vertical, pitch, and roll directions, and each
based fuzzy control systems that guarantee not only stability,
unsprung mass has freedom of motion in the vertical direction.
but also performance of closed-loop fuzzy control systems [17].
In Fig. 1, zs is the vertical displacement at the center of gravity, θ
The T–S fuzzy system is one of the most popular systems in
and φ are the pitch and roll angles of the sprung mass, ms , mu f ,
model-based fuzzy control. It is described by fuzzy IF-THEN
and mu r denote the sprung and unsprung masses, respectively,
rules that represent local linear input–output relations of a non-
and Iθ and Iφ are pitch and roll moments of inertia. The front
linear system. The overall fuzzy model of the nonlinear system
and rear displacements of the sprung mass on the left and right
is obtained by fuzzy “blending” of the linear models. The T–S
sides are denoted by z1f l , z1r l , z1f r , and z1r r . The front and
model is capable of approximating many real nonlinear systems,
rear displacements of the unsprung masses on the left and right
e.g., mechanical systems and chaotic systems. Since it employs
sides are denoted by z2f l , z2r l , z2f r , and z2r r . The disturbances,
linear models in the consequent part, linear control theory can
which are caused by road irregularities, are denoted by wf l , wr l ,
be applied for system analysis and synthesis accordingly, based
wf r , and wr r . The front and rear suspension stiffnesses and the
on the parallel-distributed compensation (PDC) scheme [18].
front and rear tyre stiffnesses are denoted by ksf , ksr , and
The T–S fuzzy models are therefore becoming powerful engi-
ktf , ktr , respectively. The front and rear suspension damping
neering tools for the modeling and control of complex dynamic
coefficients are csf and csr . Four electrohydraulic actuators are
systems.
placed between the sprung mass and the unsprung masses to
To apply the T–S model-based fuzzy control strategy to elec-
generate pushing forces, denoted by Ff l , Fr l , Ff r , and Fr r .
trohydraulic active suspensions, in this study, the nonlinear un-
Assuming that the pitch angle θ and the roll angle φ are small
certain suspension is first represented by a T–S fuzzy model.
enough, the following linear approximations are applied
Then, a fuzzy state feedback controller is designed for the fuzzy
T–S model to improve the ride comfort performance by opti- z1f l (t) = zs (t) + lf θ(t) + tf φ(t)
mizing the H∞ performance of the transfer function from the
road disturbance to the sprung mass accelerations. To avoid the z1f r (t) = zs (t) + lf θ(t) − tf φ(t)
problem of having a large number of inequalities when the in- z1r l (t) = zs (t) − lr θ(t) + tr φ(t)
put saturation constraint is characterized in terms of the convex
hull of some linear combination of linear functions and satu- z1r r (t) = zs (t) − lr θ(t) − tr φ(t) (1)
ration functions [19], the norm-bounded approach [20], [21] is and a kinematic relationship between xs (t) and q(t) can be
used here to handle the saturation nonlinearity. The sufficient established as
conditions for the existence of such a controller are derived
as linear matrix inequalities (LMIs) that can be solved very xs (t) = LT q(t) (2)
efficiently by means of the most powerful tools available to
where q(t) = [ zs (t) θ(t) φ(t) ]T , xs (t) = [z1f l (t) z1f r (t)
date, e.g., MATLAB LMI Toolbox. The proposed fuzzy state
z1r l (t) z1r r (t)]T , and
feedback controller design approach is validated by simula-
 
tions on a full-car electrohydraulic suspension model. A com- 1 1 1 1
parison of the results shows that the designed controller can  
L =  lf lf −lr −lr  .
achieve good suspension performance regardless of the actuator
nonlinear dynamics, sprung mass variation, and control input tf −tf tr −tr
constraints. In terms of mass, damping, and stiffness matrices, the motion
The rest of this paper is organized as follows. Section II equations of the full-car suspension model can be formalized as
presents the model of a full-car electrohydraulic suspension.
The T–S fuzzy model of the nonlinear uncertain suspension is Ms q̈(t) = LBs (ẋu (t) − ẋs (t))
given in Section III. In Section IV, the computational algorithm + LKs (xu (t) − xs (t)) − LF (t)
for the fuzzy state feedback controller is provided. Section V
presents the design results and simulations. Finally, the study’s Mu ẍu (t) = Bs (ẋs (t) − ẋu (t)) + Ks (xs (t) − xu (t))
findings are summarized in Section VI. + Kt (w(t) − xu (t)) + F (t) (3)
The notation used throughout the paper is reasonably stan-
dard. For a real symmetric matrix W, the notation of W > 0 where xu (t) = [ z2f l (t) z2f r (t) z2r l (t) z2r r (t) ]T , w(t) =
(W < 0) is used to denote its positive (negative) definiteness, [ wf l (t) wf r (t) wr l (t) wr r (t) ]T , F (t) = [Ff l (t) Ff r (t)
DU AND ZHANG: FUZZY CONTROL FOR NONLINEAR UNCERTAIN ELECTROHYDRAULIC ACTIVE SUSPENSIONS WITH INPUT CONSTRAINT 345
Fig. 1. Full-car suspension model.
Fr r (t)]T , and the matrices are given as

Fr l (t) The state-space form of (4) can be expressed as
 
  mu f 0 0 0 ẋg (t) = Âxg (t) + B̂1 w(t) + B̂2 F (t) (5)
ms 0 0  0 
   mu f 0 0  where
Ms =  0 Iθ 0  Mu =  
 0 0 mu r 0  0 I
0 0 Iφ xg (t) = [zmT
(t) żm T
(t)]T Â = −1 −1
,
0 0 0 mu r −Mm Km −Mm Bm
   
csf 0 0 0 ksf 0 0 0 0 0
 0 c   0 
B̂1 = −1
B̂ 2 = −1
.
 sf 0 0  0 ksf 0  Mm Km t Mm Lm
Bs =   Ks =  
 0 0 csr 0  0 0 ksr 0  The electrohydraulic actuator dynamics can be expressed as
0 0 0 csr 0 0 0 ksr [4], [9], [11]–[13]
 
ktf 0 0 0 Ḟi (t) = −βFi (t) − αA2s (ż1i (t) − ż2i (t))
 0 0 
 ktf 0 
Kt =  . sgn(xv i (t))Fi (t)
 0 0 ktr 0  + γa As Ps − xv i (t),
As
0 0 0 ktr
1
Substituting (2) into (3), we obtain ẋv i (t) = (−xv i (t) + Kv ui (t)) (6)
τ
Mm z̈m (t) + Bm żm (t) + Km zm (t) = Km t w(t) + Lm F (t) where xv i (t) is the spool valve displacement, ui (t) is the control
(4) input voltage to the servo valve, i denotes f l, f r, rl, and rr, re-
where zm (t) = [ q T (t) xTu (t) ]T and spectively, As is the actuator ram area, Ps is the hydraulic
supply
pressure, α = 4βe /Vt , β = αCtm , and γa = αCd ωa 1/ρa ,
Ms 0 LBs LT −LBs where βe is the effective bulk modulus, Vt is the total actuator
Mm = Bm =
0 Mu −Bs LT Bs volume, Ctm is the coefficient of total leakage due to pressure,
Cd is the discharge coefficient, ωa is the spool valve area gradi-
LKs LT −LKs 0 −L
Km = K mt = Lm = . ent, and ρa is the hydraulic fluid density. τ is the time constant
−Ks L T
Ks + Kt Kt I
of the spool valve dynamics and Kv is the conversion gain.
The dynamics equation (6) of the electrohydraulic actuator

can be further modeled with the state-space form as
ẋai (t) = Aai (t)xai (t) + Âai xg (t) + Bai ui (t),
Fi (t) = Cai xai (t) (7)
where
 
−β γa As fi (t)
xai (t) = [ Fi (t) xv i (t) ]T Aai (t) =  1

0 −
τ
 
0
sgn(xv i (t))Fi (t)
fi (t) = Ps − Bai =  Kv 
As Fig. 2. Membership functions.
τ
Cai = [ 1 0 ]
earity” [18], fi (t) can be represented by
and Âai xg (t) describes the term of −αA2s (ż1i (t) − ż2i (t)).
Combining the actuator dynamics equation (7) with the sus- fi (t) = M1i (ξi (t))fm ax + M2i (ξi (t))fm in (10)
pension model (5), we obtain the electrohydraulic suspension where ξi (t) = fi (t) is a premise variable, M1i (ξi (t)) and
model in state-space form as M2i (ξi (t)) are membership functions, and
ẋ(t) = A(t)x(t) + B1 w(t) + B2 ū(t) (8) fi (t) − fm in fm ax − fi (t)
M1i (ξi (t)) = M2i (ξi (t)) =
.
where x(t) = [ xTg
(t) xaf l (t) xaf r (t) xar l (t) xar r (t) ] T fm ax − fm in fm ax − fm in
(11)
is the state vector and ū(t) is the bounded input voltage to the
Similarly, the uncertain sprung mass ms (t) is bounded by its
actuator servo valve. In real applications, the input voltage
minimum value msm in and its maximum value msm ax , and can
to the servo valve can be bounded as ū(t) =sat(u(t)), where
thus be represented by
sat(u(t)) is a saturation function of control input u(t), defined
as 1
 = N1 (ξm (t))mm ax + N2 (ξm (t))mm in (12)
ms (t)
 −ulim , if u(t) < −ulim

sat(u(t)) = u(t), if − ulim ≤ u(t) ≤ ulim (9) where ξm (t) = 1/ms (t) is also a premise variable, mm ax =

 1/msm in , mm in = 1/msm ax , and N1 (ξm (t)) and N2 (ξm (t))
ulim , if u(t) > ulim
are membership functions that are defined as
where ulim is the control input limit. The matrices are
1/ms (t) − mm in
N1 (ξm (t)) =
Â B̂2 Ca B̂1 0 mm ax − mm in
A(t) = B1 = B2 = .
Âa Aa (t) 0 Ba mm ax − 1/ms (t)
N2 (ξm (t)) = . (13)
It is noted that the system matrix A(t) is a nonlinear and mm ax − mm in
time-varying matrix due to the nonlinear time-varying behavior For description brevity, we name the aforementioned mem-
of the actuator dynamics and the variation of the sprung mass. bership functions M1i (ξi (t)), M2i (ξi (t)), N1 (ξm (t)), and
N2 (ξm (t)), shown in Fig. 2, as ‘big,” “small,” ‘light,” and
III. T–S FUZZY MODELING ‘heavy,” respectively. The nonlinear uncertain suspension model
(8) can then be represented by a T–S fuzzy model composed of
The full-car electrohydraulic suspension model (8) incorpo-
32 (25 ) fuzzy rules, as listed in Table I, where B, S, L, and H
rates well-characterized and essential actuator nonlinearities,
represent ‘big,” ‘small,” ‘light,” and ‘heavy,” respectively. To
and the controller design is required to consider both the param-
describe the T–S fuzzy model more clearly, several examples of
eter uncertainty and the control input constraint, which leads to
the fuzzy IF-THEN rules corresponding to Table I are explained
a challenging control problem. In order to design a controller for
as follows.
the model through the fuzzy approach, the T–S fuzzy modeling
Model Rule 1:
technique will be applied, and the idea of “sector nonlinear-
ity” [18] is employed to construct an exact T–S fuzzy model for IF ξi (t) (i denotes f l, f r, rl, and rr, respectively)
the nonlinear uncertain suspension system (8). are small and ξm (t) is light,
Suppose that the actuator force Fi (t) (where i denotes f l, f r, THEN ẋ(t) = A1 x(t) + B1 w(t) + B2 ū(t)
rl, and rr, respectively) is bounded in practice by its minimum
where matrix A1 is obtained from matrix A(t) in (8) by replacing
value Fim in and its maximum value Fim ax ; the nonlinear func-
fi with fm in and 1/ms with mm ax .
tion fi (t) is then bounded by its minimum value fm in and its ..
maximum value fm ax . Thus, using the idea of “sector nonlin- .
TABLE I ..
LIST OF FUZZY RULES .
h32 (ξ(t)) = M1f l (ξf l (t))M1f r (ξf r (t))M1r l (ξr l (t))
× M1r r (ξr r (t))N2 (ξm (t)),

32
hi (ξ(t)) ≥ 0, i = 1, 2, . . . , 32, and hi (ξ(t)) = 1.
i=1
In practice, the actuator force Fi (t), the spool valve position

xv i (t), and the sprung mass ms (t) can be measured; thus, the
T–S fuzzy model (14) can be realized.
It is noted that the T–S fuzzy model (14) is obtained via
the “sector nonlinearity” approach based on the analysis of
the nonlinear function fi (t) and the variation of sprung mass
ms , the bounds of which can be estimated in a real operat-
ing situation. The construction of a T–S fuzzy model from
a given nonlinear dynamic model can also utilize the idea
of “local approximation” or a combination of “sector non-
linearity” and “local approximation” [18]. In general, these
are analytic transformation techniques, which can be applied
Model Rule 16: only to models described analytically. Since analytic techniques
IF ξi (t) (i denotes f l, f r, rl, and rr, respectively) need problem-dependent human intuition and cannot be easily
are big and ξm (t) is light, solved in some cases, recently, a higher order-singular-value-
THEN ẋ(t) = A16 x(t) + B1 w(t) + B2 ū(t) decomposition (HOSVD)-based tensor product (TP) model
transformation approach was proposed to automatically and nu-
where matrix A16 is obtained from matrix A(t) in (8) by replac- merically transform a general dynamic system model into a TP
ing fi with fm ax and 1/ms with mm ax . model form, including polytopic and T–S model forms [22].
Model Rule 17: The TP model representation has shown various advantages
IF ξi (t) (i denotes f l, f r, rl, and rr, respectively) for LMI-based controller design [23], [24], and relaxed LMI
are small and ξm (t) is heavy, conditions can be further obtained for closed-loop fuzzy sys-
THEN ẋ(t) = A17 x(t) + B1 w(t) + B2 ū(t) tems with TP structure [25]. There is also a MATLAB Tool-
box for TP model transformation (available for download to-
where matrix A17 is obtained from matrix A(t) in (8) by replac- gether with documentation and examples at http://tptool.sztaki.
ing fi with fm in and 1/ms with mm in . hu/tpde).
.. For our problem, in fact, using the convex normalized (CNO)
.
Model Rule 32: type of TP model transformation can obtain the same member-
ship functions as those described before when fi (t) and 1/ms (t)
IF ξi (t) (i denotes f l, f r, rl, and rr, respectively)
are used as time-varying parameter variables. However, if ms (t)
are big and ξm (t) is heavy,
is used as a time-varying parameter variable instead of 1/ms (t),
THEN ẋ(t) = A32 x(t) + B1 w(t) + B2 ū(t)
different membership functions can be generated with the TP
where matrix A32 is obtained from matrix A(t) in (8) by replac- model transformation, as shown in Fig. 3. It can be seen from
ing fi with fm ax and 1/ms with mm in . Fig. 3 that the membership functions are nonlinear, and they
Thus, the T–S fuzzy model that represents exactly the non- are different from those shown in Fig. 2. Nevertheless, using
linear uncertain suspension model (8) under the assumption this new type of membership function does not alter the LMI-
of bounds on actuator forces Fi (t) ∈ [Fm in , Fm ax ] and sprung based controller design process or the design results obtained
mass ms (t) ∈ [msm in , msm ax ] is obtained as with the membership functions defined in (13). Generally, Fi (t)

32 and xv i (t) can be directly used as the time-varying parameter
ẋ(t) = hi (ξ(t))Ai x(t) + B1 w(t) + B2 ū(t) (14) variables to obtain the T–S model using the TP model transfor-
i=1 mation. However, the TP model transformation should consider
where the tradeoff between approximation accuracy and complexity.
For the studied problem, when using the derived membership
functions (11) and (13), only 32 fuzzy rules need to be applied.
× M2r r (ξr r (t))N1 (ξm (t)) Therefore, in this paper, the derived membership functions (11)
and (13) are used. And, despite the authors’ effort, no other
types of membership functions are found to yield better perfor-
× M2r r (ξr r (t))N1 (ξm (t)) mance than the derived membership functions (11) and (13),
controller via the PDC as

32
u(t) = hi (ξ(t))Ki x(t) = Kh x(t) (18)
i=1

where Kh = 32 i=1 hi (ξ(t))Ki and Ki is the state feedback gain
matrix to be designed.
Since ride comfort is an important performance requirement
for a vehicle suspension and it can usually be quantified by
the sprung mass acceleration, the sprung mass acceleration is
chosen as the control output, i.e.

32
z(t) = z̈s (t) = hi (ξ(t))Ci x(t) = Ch x(t) (19)
i=1

where Ch = 32 i=1 hi (ξ(t))Ci and Ci is extracted from the
eighth to the tenth row of matrix Ai , i = 1, 2, . . . , 32.
Fig. 3. Membership functions for sprung mass obtained from TP model In order to design an active suspension to perform adequately
transformation. in a wide range of shock and vibration environments, the L2
gain of the system (15) with (19) is chosen as the performance
although there may exist methods to automatically generate measure, which is defined as
affine decompositions. z2
Tz w ∞ = sup (20)
w 2 = 0 w 2
IV. FUZZY CONTROLLER DESIGN
∞ ∞
In order to avoid the problem associated with having a large where z22 = 0 z T (t)z(t)dt and w22 = 0 wT (t)w(t)dt,
number of inequalities involved in the controller design, the and the supermum is taken over all nonzero trajectories of the
norm-bounded approach [20], [21] is used to handle the satu- system (15) with x(0) = 0. Our goal is to design a fuzzy con-
ration nonlinearity defined in (9). Hence, (14) will be written troller (18) such that the fuzzy system (15) with controller (18)
as is quadratically stable and the L2 gain (20) is minimized.
To design the controller, the following lemma will be used.

32
Lemma 2: For any matrices (or vectors) X and Y with appro-
ẋ(t) = hi (ξ(t))Ai x(t) + B1 w(t) + B2 ū(t)
priate dimensions, we have
i=1

32 X T Y + Y T X ≤ X T X + −1 Y T Y
1+ε
= hi (ξ(t))Ai x(t) + B1 w(t) + B2 u(t)
2 where > 0 is any scalar.
i=1
Theorem 3: For a given number γ > 0, 0 < ε < 1, the T–S
1+ε fuzzy system (15) with controller (18) is quadratically stable and
+ B2 ū(t) − u(t)
2 the L2 gain defined by (20) is less than γ if there exist matrices
1+ε Q > 0, Yi , i = 1, 2, . . . , 32, and scalar > 0, such that, (21)
= Ah x(t) + B1 w(t) + B2 u(t) + B2 v(t) (15) and (22), as shown at the bottom of the next page.
2
Moreover, the fuzzy state feedback gains can be obtained as
where Ah = 32 i=1 hi (ξ(t))Ai and v(t) = ū(t) − 2 u(t), 0 <
1+ε
Ki = Yi Q−1 , i = 1, 2, . . . , 32.
ε < 1. And for designing the controller, the following lemma Proof: Let us define a Lyapunov function for the system (15)
will be used. as
Lemma 1 [20]: For the saturation constraint defined by (9),
as long as |u(t)| ≤ u lεi m , we have V (x(t)) = xT (t)P x(t) (23)

where P is a positive definite matrix. By differentiating (23),
ū(t) − 1 + ε u(t) ≤ 1 − ε u(t) (16)
2 2
we obtain
and hence V̇ (x(t)) = ẋT (t)P x(t) + xT (t)P ẋ(t)

T 2 T
1+ε 1+ε 1−ε 1+ε
ū(t) − u(t) ū(t) − u(t) ≤ uT (t)u(t) = Ah x(t) + B1 w(t) + B2
2
u(t) +B2 v(t) P x(t)
2 2 2
(17)
1+ε
where 0 < ε < 1. + xT (t)P Ah x(t) + B1 w(t) + B2 u(t)
2
The fuzzy controller design for the T–S fuzzy model (15) is
carried out based on the so-called PDC scheme [18]. For the
+ B2 v(t) . (24)
T–S fuzzy model (15), we construct the fuzzy state feedback
By Lemma 1, Lemma 2, and definition (18), we have can be inferred from (27) that if Π < 0, then V̇ (x(t)) < 0, and
T the fuzzy system (15) with the controller (18) is quadratically
1+ε
V̇ (x(t)) ≤ x (t) Ah P + P Ah + B2
T T
Kh P stable.
2 Pre- and postmultiplying (28) by diag( P −1 I ) and its trans-

1+ε pose, respectively, and defining Q = P −1 and Yh = Kh Q, the
+ P B2 Kh x(t) condition Π < 0 is equivalent to
2
+ wT (t)B1T P x(t) + xT (t)P B1 w(t) + v T (t)v(t)  1+ε T T 1+ε 
QATh + Ah Q + Yh B2 + B2 Yh
−1 T 
+ x (t)P B2 B2T P x(t)
 1 − ε2 2 2 B1 

Σ =+ YhT Yh + −1 B2 B2T + QChT Ch Q < 0.
≤ xT (t)Θx(t) + wT (t)B1T P x(t) + xT (t)P B1 w(t)  2 
(25) B1T −γ 2 I
(29)
where By the Schur complement, Σ < 0 is equivalent to (Ψ), as shown
T at the bottom of this page.
1+ε 1+ε
Θ = ATh P + P Ah + B2 Kh P + P B2 Kh By the definitions Ah = 32 i=1 hi (ξ(t))Ai , Ch =
2 2 32 32
2 h (ξ(t))C , Y = h (ξ(t))Y , and the fact
i=1 i i h
i=1 i i
1−ε T −1 T that hi (ξ(t)) ≥ 0 and 32 h
i=1 i (ξ(t)) = 1, Ψ < 0 is equivalent
+ Kh Kh + P B2 B2 P (26)
2 to (21).
and is any positive scalar. On the other hand, from (18), the constraint |u(t)| ≤ u lεi m can
Adding z T (t)z(t) − γ 2 wT (t)w(t) on both sides of (25) be expressed as
yields 32
u
V̇ (x(t)) + z (t)z(t) − γ w (t)w(t)
T 2 T lim
hi (ξ(t))Ki x(t) ≤ . (31)
ε
Θ + ChT Ch P B1 x(t) i=1
≤ [ x (t) w (t) ]
T T
. (27)
B1T P −γ 2 I w(t)
It is obvious that if |Ki x(t)| ≤ u lεi m , then (31) holds. Let
Let us consider
Ω(K) = {x(t)| |xT (t)KiT Ki x(t)| ≤ ( u lεi m )2 }; then the equiv-
Θ + ChT Ch P B1 alent condition for an ellipsoid Ω(P, ρ) = {x(t)| xT (t)P x(t) ≤
Π= <0 (28)
B1T P −γ 2 I ρ} being a subset of Ω(K), i.e., Ω(P, ρ) ⊂ Ω(K), is [19]
then, V̇ (x(t)) + z T (t)z(t) − γ 2 wT (t)w(t) ≤ 0 and the L2 gain −1
defined in (20) is less than γ > 0 with the initial condition P u 2
lim
Ki KiT ≤ . (32)
x(0) = 0 [26]. When the disturbance is zero, i.e., w(t) = 0, it ρ ε
 1+ε T T 
QATi + Ai Q + Yi B2 + B2 Yi + −1 B2 B2T YiT QCiT B1
 2 
 2 
 
 2 
 ∗ −−1 I 0 0 <0 (21)
 1−ε 
 
 
 ∗ ∗ −I 0 
∗ ∗ ∗ −γ 2 I
 u 2 
lim
I Yi
 ε  ≥ 0. (22)
−1
YiT ρ Q
 1+ε T T 
QATh + Ah Q + Yh B2 + B2 Yh + −1 B2 B2T YhT QChT B1
 2 
 2 
 
 2 
Ψ= ∗ −−1 I 0 0  < 0. (30)
 1−ε 
 
 
 ∗ ∗ −I 0 
∗ ∗ ∗ −γ 2 I
By the Schur complement, inequality (32) can be written as TABLE II

PARAMETER VALUES OF THE FULL-CAR SUSPENSION MODEL
 −1 
ulim 2 P
 I Ki 
 ε ρ 
 −1  ≥ 0. (33)
 P −1 P 
T
Ki
ρ ρ
Using the definitions Q = P −1 and Yi = Ki Q, inequality (33)
is equivalent to (22). This completes the proof.
The minimization of γ can be realized as
min γ subject to LMIs (21) and (22). (34) TABLE III
PARAMETER VALUES OF THE HYDRAULIC ACTUATOR
This problem can be solved very efficiently by means of the
MATLAB LMI Toolbox software.
Remark 1: It is noted that once the solution of problem (34)
is feasible, then the value of γ can be obtained using the LMI
Toolbox software and the designed controller can guarantee the
L2 gain (20) to be less than γ in terms of the LMI condition (21). mass is assumed to be varied between ms min = 1120 kg and
When the designed controller is applied to the system (14), the ms max = 1680 kg, which is ±20% variation of the nominal
real value of the L2 gain (20) for the system (14) under a given sprung mass. Using the controller design approach presented in
disturbance can be evaluated by measuring the control output Section IV, and choosing ε = 0.97 by trial and error, we obtain
response (19) and calculating the value using equation (20). the controller gain matrices Ki that consist of 32 matrices, each
Remark 2: In the paper, the common quadratic Lyapunov with dimensions of 4 × 22. It is noted that the controller gain
function approach, where the Lyapunov function candidate is matrices are constant matrices that do not need to be recalcu-
defined as in (23), is utilized to derive the controller synthesis lated in a real-time implementation and can be easily stored in
conditions. It has been noted that common quadratic Lyapunov a microprocessor memory (RAM or ROM). The calculation of
functions tend to be conservative and, even worse, might not ex- the control outputs, i.e., input voltages sent to the actuators, in
ist for some complex highly nonlinear systems [17] . This is one terms of the measured state variables and the calculated mem-
of the main limitations of this kind of approach. With regard bership functions is also quite straightforward. Therefore, the
to overcoming the drawback of common quadratic Lyapunov required computational power will not be very high, which en-
functions, piecewise quadratic Lyapunov functions and fuzzy ables the implementation of the controller on a DSP, e.g., the
Lyapunov functions have received increasing attention recently Texas Instruments TMS320C30.
[27], [28]. However, controller synthesis conditions based on To compare the suspension performance, an optimal H∞
piecewise quadratic Lyapunov functions and fuzzy Lyapunov controller is designed for the linear full-car suspension model
functions are generally given by bilinear matrix inequalities (5) without considering the electrohydraulic actuator dynamics
(BMIs), which have to be solved by way of, e.g., a cone com- (7). By defining the control output as the sprung mass acceler-
plementarity linearisation approach [28] or a descriptor system ation and using the bounded real lemma (BRL), this controller
approach [27]. Therefore, the computation cost and complex- gain matrix is obtained as (35), as shown at the bottom of the
ity of using piecewise quadratic Lyapunov function and fuzzy next page.
Lyapunov function approaches would be much higher in general. Since the optimal H∞ controller design theory can be found
In terms of the possible requirement on real-time computation in many references, the controller design process is omitted here
for a practical system and the feasible solutions checking on for brevity.
the given system, this paper keeps using the common quadratic In the simulation, a test road disturbance, which is given as
Lyapunov function approach regardless of its conservatism.
zr (t) = 0.0254 sin 2πt + 0.005 sin 10.5πt
V. APPLICATION EXAMPLE + 0.001 sin 21.5πt (m) (36)
In this section, we will apply the proposed approach to design is used first. This road disturbance is close to the car body
a fuzzy state feedback controller for a full-car electrohydraulic resonance frequency (1 Hz), with high-frequency disturbance
suspension model, as described in Section II. The full-car sus- added to simulate the rough road surface. To observe the roll
pension model parameter values are listed in Table II, and the motion, this road disturbance is assumed to pass the wheels on
parameter values for each hydraulic actuator used in the simu- the left side of the car only. The simulation program is realized
lation are given in Table III. by MATLAB/Simulink.
In this study, we suppose that the input voltage of each spool For the nominal sprung mass under the specified road distur-
valve is limited to ulim = 2.5 V, and each actuator output force bance (36), the time-domain responses for three suspensions,
is limited to 2000 N. The bounds of the nonlinear function fi (t) i.e., passive suspension, active suspension with H∞ controller
are estimated as fm in = 2800 and fm ax = 4000. The sprung (35), and active suspension with the electrohydraulic actuators
Fig. 4. Time-domain responses under a test road profile. Solid line is for active suspension with fuzzy controller. Dotted line is for active suspension with H ∞
controller. Dot-dashed line is for passive suspension.
and the designed fuzzy controller are compared. The responses, fort performance. Fig. 6 shows the control input voltages, and
consisting of the sprung mass heave acceleration, heave dis- Fig. 7 shows the nonlinear function outputs. It can be seen from
placement, pitch angle, and roll angle, are plotted in Fig. 4. It Figs. 6 and 7 that the control input voltages are within the de-
is observed from Fig. 4 that the proposed fuzzy control strategy fined input voltage range and the nonlinear function outputs are
reduces the sprung mass acceleration and displacement mag- located within the estimated bounds.
nitudes significantly compared to the passive suspension under Now, consider the case of an isolated bump in an otherwise
the same road disturbance. The active suspension with the de- smooth road surface. The corresponding ground displacement
signed fuzzy controller also achieves a very similar suspension for the wheel is given by
performance to the active suspension with an optimal H∞ con- a
2πv0

l
troller. This confirms that the proposed fuzzy control strategy 
 1 − cos t , 0≤t≤
2 l v0
can realize good suspension performance with highly nonlinear
zr (t) = (37)
electrohydraulic actuators. 
 l
 0, t>
Fig. 5 shows the actuator output forces for the active sus- v0
pension with the designed fuzzy controller. It is observed that
the four actuators provide different control forces in accordance where a and l are the height and the length of the bump. We
with the fuzzy rules and the measurements of the state variables. choose a = 0.1 m, l = 10 m, and the vehicle forward velocity
The proposed fuzzy control strategy provides effective actuator as v0 = 45 km/h.
forces that aim to optimize the sprung mass heave acceleration, For the nominal sprung mass under the road disturbance (37),
pitch acceleration, and roll acceleration to improve ride com- Fig. 9 shows the bump responses of the sprung mass heave

−1.0324 −0.2478 −1.2670 0.8961 0.2431 2.2499
 −1.0324 −0.2478 −2.0930
 1.2670 0.2431 0.8961
104 × 
 −1.3299 1.3735 −1.5342 0.4602 −0.4434 −0.6042
−1.3299 1.3735 1.5342 −0.4434 0.4602 1.6712
 (35)
−2.0930 0.2879 0.3718 0.0665 0.0044 −0.0023 0.0078 −0.0072
2.2499 0.2879 0.3718 −0.0665 −0.0023 0.0044 −0.0072 0.0078 

.
1.6712 0.2097 −0.3700 0.0504 0.0128 −0.0124 0.0110 −0.0081 
−0.6042 0.2097 −0.3700 −0.0504 −0.0124 0.0128 −0.0081 0.0110
Fig. 5. Actuator output forces for active suspension with fuzzy controller.
Fig. 6. Actuator control voltages for active suspension with fuzzy controller.
acceleration, pitch acceleration, and roll acceleration for the formance very similar to the active suspension with an optimal
passive suspension, the active suspension with H∞ controller H∞ controller. The active suspension performance is signifi-
(35), and the active suspension with the electrohydraulic actua- cantly improved compared to the passive suspension.
tors and the designed fuzzy controller. It is again observed that To illustrate the effect of sprung mass variation, Fig. 9 shows
the proposed fuzzy control strategy achieves suspension per- the bump responses of the sprung mass heave acceleration for
Fig. 7. Nonlinear function values for active suspension with fuzzy controller.
Fig. 8. Acceleration responses under a bump road profile. Solid line is for active suspension with fuzzy controller. Dotted line is for active suspension with H ∞
controller. Dot-dashed line is for passive suspension.
the passive suspension (passive) and the active suspension with lower peak and shorter settling time are obtained. Figs. 10 and 11
electrohydraulic actuator and fuzzy controller (active) when the show the bump responses of the sprung mass pitch acceleration
sprung mass is 1120 and 1680 kg. It is observed that, despite the and roll acceleration, respectively. It can be seen from Fig. 10
change in sprung mass, the designed fuzzy controller achieves that the sprung mass affects the pitch acceleration significantly.
significantly better performance on heave acceleration, where a However, the active suspension can keep the pitch acceleration
Fig. 9. Heave acceleration responses for different sprung mass.
Fig. 10. Pitch acceleration responses for different sprung mass.
low, regardless of the sprung mass variation. In Fig. 11, the be represented by
active suspension achieves lower roll acceleration compared to
żr (t) = 2πq0 G0 V ω(t) (38)

the passive suspension, although the sprung mass variation does
not affect roll acceleration due to the symmetric distribution of where G0 stands for the road roughness coefficient, q0 is the
the sprung mass about the vehicle’s roll axis. Figs. 9–11 indicate reference spatial frequency, V is the vehicle forward velocity,
that the improvement in ride comfort can be maintained by the and ω(t) is zero-mean white noise with identity power spectral
designed active suspension for large changes in load conditions. density. For a given road roughness G0 = 512 × 10−6 m3 and
When the road disturbance is considered as random vibration, a given vehicle forward velocity V = 20 m/s, the rms values
it is typically specified as a stationary random process that can for sprung mass heave acceleration, pitch acceleration, and roll
Fig. 11. Roll acceleration responses for different sprung mass.
Fig. 12. RMS ratios for sprung mass heave acceleration, pitch acceleration, and roll acceleration versus sprung mass.
acceleration are calculated as the sprung mass changes from fuzzy control strategy can realize good ride comfort perfor-
1120 to 1680 kg. The rms ratios between the active suspension mance for electrohydraulic suspension even when the sprung
with fuzzy controller and the passive suspension are plotted mass is varied significantly.
against sprung mass in Fig. 12. It can be seen that the designed
fuzzy controller maintains a ratio below 1, regardless of the
large variations in the sprung mass. When the road roughness VI. CONCLUSION
and vehicle forward velocity are given different values, very In this paper, we have presented a fuzzy state feedback control
similar results are obtained. For brevity, these results are not strategy for electrohydraulic active suspensions to deal with
shown. Fig. 12 further validates the claim that the proposed the nonlinear actuator dynamics, sprung mass variation, and
control input constraint problems. First, using the idea of “sector [17] G. Feng, “A survey on analysis and design of model-based fuzzy control
nonlinearity,” the nonlinear uncertain electrohydraulic actuator systems,” IEEE Trans. Fuzzy Syst., vol. 14, no. 5, pp. 676–697, Oct. 2006.
[18] K. Tanaka and H. O. Wang, Fuzzy Control Systems Design and Analy-
was represented by a T–S fuzzy model in defined regions. Thus, sis:Linear Matrix Inequality Approach. New York: Wiley, 2001.
by means of the PDC scheme, a fuzzy state feedback controller [19] Y.-Y. Cao and Z. Lin, “Robust stability analysis and fuzzy-scheduling
was designed for the obtained T–S fuzzy model to optimize the control for nonlinear systems subject to actuator saturation,” IEEE Trans.
Fuzzy Syst., vol. 11, no. 1, pp. 57–67, Feb. 2003.
H∞ performance of ride comfort. At the same time, the actuator [20] J. H. Kim and F. Jabbari, “Actuator saturation and control design for build-
input voltage constraint was incorporated into the controller ings under seismic excitation,” J. Eng. Mech., vol. 128, no. 4, pp. 403–412,
design process. The sufficient conditions for designing such a 2002.
[21] C.-S. Tseng and B.-S. Chen, “H ∞ fuzzy control design for nonlinear
controller were expressed by LMIs. Simulations were used to systems subject to actuator saturation,” in Proc. 2006 IEEE Int. Conf.
validate the effectiveness of the designed controller. Fuzzy Syst., Vancouver, BC, Canada, pp. 783–788.
[22] P. Baranyi, “TP model transformation as a way to LMI-based controller
design,” IEEE Trans. Ind. Electron., vol. 51, no. 2, pp. 387–400, Apr.
ACKNOWLEDGMENT 2004.
[23] P. Baranyi and Y. Yam, “Case study of the TP-model transformation in the
The authors would like to thank the editors and anonymous control of a complex dynamic model with structural nonlinearity,” IEEE
reviewers for their invaluable suggestions and comments on the Trans. Ind. Electron., vol. 53, no. 3, pp. 895–904, Jun. 2006.
improvement of the paper. They also would like to thanks. Mr. [24] Z. Petres, P. Baranyi, P. Korondi, and H. Hashimoto, “Trajectory tracking
by TP model transformation: Case study of a benchmark problem,” IEEE
W. Smith for his careful reading and English modification of the Trans. Ind. Electron., vol. 54, no. 3, pp. 1654–1663, Jun. 2007.
manuscript. [25] C. Arino and A. Sala, “Relaxed LMI conditions for closed-loop fuzzy
systems with tensor-product structure,” Eng. Appl. Artif. Intell., vol. 20,
pp. 1036–1046, 2007.
REFERENCES [26] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan, Linear Matrix
[1] D. Hrovat, “Survey of advanced suspension developments and related Inequalities in System and Control Theory. Philadelphia, PA: SIAM,
optimal control applications,” Automatica, vol. 33, no. 10, pp. 1781– Jun. 1994.
1817, 1997. [27] K. Tanaka, H. Ohtake, and H. O. Wang, “A descriptor system approach to
[2] R. A. Williams, “Automotive active suspensions,” Proc. Inst. Mech. Eng. fuzzy control system design via fuzzy Lyapunov functions,” IEEE Trans.
D: J. Automobile Eng., vol. 211, pp. 415–444, 1997. Fuzzy Syst., vol. 15, no. 3, pp. 333–341, Jun. 2007.
[3] A. Alleyne, P. D. Neuhaus, and J. K. Hedrick, “Application of nonlinear [28] H. Gao, Z. Wang, and C. Wang, “Improved H ∞ control of discrete-time
fuzzy systems: A cone complementarity linearization approach,” Inf. Sci.,
control theory to electronically controlled suspensions,” Vehicle Syst.
vol. 175, pp. 57–77, 2005.
Dyn., vol. 22, no. 5, pp. 309–320, 1993.
[4] A. Alleyne and J. K. Hedrick, “Nonlinear adaptive control of active sus-
pensions,” IEEE Trans. Control Syst. Technol., vol. 3, no. 1, pp. 94–101,
Mar. 1995.
[5] S.-J. Huang and W.-C. Lin, “Adaptive fuzzy controller with sliding surface
for vehicle suspension control,” IEEE Trans. Fuzzy Syst., vol. 11, no. 4, Haiping Du received the Ph.D. degree in mechanical
pp. 550–559, Aug. 2003. design and theory from Shanghai Jiao Tong Univer-
[6] S. Chantranuwathana and H. Peng, “Adaptive robust force control for sity, Shanghai, China, in 2002.
vehicle active suspensions,” Int. J. Adaptive Control Signal Process., He is currently a Research Fellow in the Fac-
vol. 18, no. 2, pp. 83–102, 2004. ulty of Engineering, University of Technology, Syd-
[7] Y. Zhang and A. Alleyne, “A practical and effective approach to active ney, N.S.W., Australia. He was a Postdoctoral Re-
suspension control,” Vehicle Syst. Dyn., vol. 43, no. 5, pp. 305–330, 2006. search Associate at the University of Hong Kong,
[8] S.-J. Huang and H.-Y. Chen, “Adaptive sliding controller with self-tuning Hong Kong, from 2002 to 2003, and at the Imperial
fuzzy compensation for vehicles suspension control,” Mechatronics, College London, U.K., from 2004 to 2005. His cur-
vol. 16, pp. 607–622, 2006. rent research interests include robust control theory
[9] P. C. Chen and A. C. Huang, “Adaptive sliding control of active suspension and engineering applications, soft computing, dy-
systems with uncertain hydraulic actuator dynamics,” Veh. Syst. Dyn., namic system modeling, model and controller reduction, and smart materials
vol. 44, no. 5, pp. 357–368, 2006. and structures.
[10] C. Kaddissi, J.-P. Kenne, and M. Saad, “Identification and real-time control Dr. Du was the recipient of the Excellent Ph.D. Thesis Prize from Shanghai
of an electrohydraulic servo system based on nonlinear backstepping,” Provincial Government in 2004.
IEEE/ASME Trans. Mechatronics, vol. 12, no. 1, pp. 12–22, Feb. 2007.
[11] A. Alleyne and R. Liu, “On the limitations of force tracking control for
hydraulic servosystems,” J. Dyn. Syst., Meas. Control, vol. 121, no. 2,
pp. 184–190, 1999.
[12] A. Alleyne and R. Liu, “A simplified approach to force control for electro-
hydraulic systems,” Control Eng. Pract., vol. 8, no. 12, pp. 1347–1356, Nong Zhang received the Doctorate degree from the
2000. University of Tokyo, Tokyo, Japan, in 1989.
[13] A. Alleyne and R. Liu, “Systematic control of a class of nonlinear systems In 1989, he joined the Faculty of Engineering, Uni-
with application to electrohydraulic cylinder pressure control,” IEEE versity of Tokyo, as a Research Assistant Professor. In
Trans. Control Syst. Technol., vol. 8, no. 4, pp. 623–634, Jul. 2000. 1992, he joined the Engineering Faculty, University
[14] A. G. Thompson and B. R. Davis, “Force control in electrohydraulic active of Melbourne, as a Research Fellow. Since 1995, he
suspensions revisited,” Vehicle Syst. Dyn., vol. 35, no. 3, pp. 217–222, has been with the Faculty of Engineering, University
2001. of Technology, Sydney, N.S.W., Australia. His cur-
[15] Y.-S. Lu and J.-S. Chen, “A self-organizing fuzzy sliding-mode controller rent research interests include experimental modal
design for a class of nonlinear servo systems,” IEEE Trans. Ind. Electron., analysis, rotor dynamics, vehicle power train dynam-
vol. 41, no. 5, pp. 492–502, Oct. 1994. ics, and hydraulically interconnected suspension and
[16] N. Al Holou, T. Lahdhiri, D. S. Joo, J. Weaver, and F. Al Abbas, “Sliding vehicle dynamics.
mode neural network inference fuzzy logic control for active suspension Dr. Zhang is a Member of the American Society of Mechanical Engineers
systems,” IEEE Trans. Fuzzy Syst., vol. 10, no. 2, pp. 234–246, Apr. 2002. (ASME) and a Fellow of the Society of Automotive Engineers, Australasia.
Rapid Load Following of an SOFC Power System via

Stable Fuzzy Predictive Tracking Controller
Tiejun Zhang, Member, IEEE, and Gang Feng, Fellow, IEEE
Abstract—The solid oxide fuel cell (SOFC) is widely accepted of deficient fuel supply [8], [9]. Therefore, an effective control
for clean and distributed power generation use, but critical opera- system is in great demand to ensure that the fuel cell system
tion problems often occur when the stand-alone fuel cell is directly meets the time-varying power load demand with high process
connected to the electricity grid or the dc electric user. In order to
address these problems, in this paper, a data-driven fuzzy modeling operation efficiency [2]–[10].
method is employed to identify the dynamic model of an integrated Model predictive control (MPC) is one of the most powerful
SOFC/capacitor system. A novel offset-free input-to-state stable optimized control approaches to wide-range-constrained mul-
fuzzy predictive controller is developed based on the obtained fuzzy tivariable industrial processes with slow dynamics [11]–[15],
model. Both the rapid power load following and safe SOFC oper- particularly for power industry [16]. In MPC, the control ac-
ation requirements are taken into account in the design of the
closed-loop control system. Simulations are also given to demon- tions are computed by solving a receding-horizon optimization
strate the load following control performance of the proposed problem at each sampling time instant. This is in contrast to con-
fuzzy predictive control strategy for the SOFC/capacitor power ventional optimal control, where only a predetermined control
system. law is employed. However, as for large-scale nonlinear pro-
Index Terms—Fuel cell, fuzzy systems, identification, input-to- cesses, MPC suffers from a couple of important drawbacks in
state stability, load following, output tracking, predictive control. practical application. First, it is difficult and time-consuming to
obtain a reliable and concise first-principles prediction model
for highly nonlinear and complex plants. Second, for nonlinear
I. INTRODUCTION plants, especially with fast dynamics, nonlinear MPC suffers
S ONE of the second generations of fuel cells, the solid greatly from its intensive computational burden of solving a
A oxide fuel cell (SOFC) has been demonstrated to be a
promising power generation technology, especially in stationary
nonlinear nonconvex optimization problem. Hence, it is desir-
able to exploit alternative approximate-model-based predictive
applications [1]–[6]. They are characterized by high operating control approaches [17]–[19] so as to alleviate heavy optimiza-
temperature range (600 ◦ C–1000 ◦ C) with the use of solid-state tion burden and restrictive requirement on the precise mathe-
ceramic electrolyte, which is easier to maintain due to the lack matical plant model in nonlinear predictive control [13]–[15].
of cell corrosion. The SOFC/gas turbine (GT) based distribution On the other hand, several offset-free MPC design methods
generation can provide ancillary services such as load follow- have been developed to meet the output tracking requirements
ing and regulation with respect to the current deregulation and of constrained linear systems [20]–[22]. In these tracking con-
unbundling of the energy market [4], [5]. However, load follow- trol systems, an integrating disturbance signal is estimated to
ing problems occur when the response of the fuel cell system account for the plant–model mismatch and/or unmodeled plant
cannot safely meet both the external power load demand and disturbances. Then, a steady-state target generator is constructed
the balance of parasitic plant power demand [6]. For example, to remove the effects of estimated disturbances, thus making
the phenomenon of oxygen starvation will appear in a fuel cell sure the zero steady-state offset output tracking through a linear
when the sudden power load changes greatly [7]. In that case, MPC. Obviously, an offset-free linear MPC may not be capable
the partial pressure of oxygen falls dramatically, accompanied of controlling strongly nonlinear processes operating in a wide
by rapid decrease in cell voltage, which would shorten the life range to achieve the steady-state zero offset.
of the fuel cell stack. On the other hand, the fuel cell may also be Takagi–Sugeno (T–S) fuzzy dynamic models have been
permanently damaged when the fuel starvation occurs in case widely accepted in control community for its capability to repre-
sent nonlinear dynamics [24]–[30]. Fuzzy systems are demon-
Manuscript received December 22, 2007; revised June 19, 2008 and
strated to have very good approximation and interpretation ca-
September 9, 2008. First published December 22, 2008; current version pub- pability to general nonlinear systems [27]–[29]. And there have
lished April 1, 2009. This work was supported in part by the Hong Kong been many offline state feedback and output feedback controller
Research Grant Council under CityU 112806 and by the National Natural Sci-
ence Foundation of China under Grant 50640460116.
design methods for T–S fuzzy systems [25], [26], [30], whereas
T. J. Zhang was with the Department of Manufacturing Engineering and Engi- the transient response of the closed-loop control system has
neering Management, City University of Hong Kong, Kowloon, Hong Kong. He rarely been considered. As a matter of fact, transient perfor-
is now with the Center for Automation Technologies and Systems, Rensselaer
Polytechnic Institute, Troy, NY 12180-3590 USA (e-mail: tjzhang@ieee.org).
mance is of much more concern in industrial process control
G. Feng is with the Department of Manufacturing Engineering and Engi- and economical plant operation. Fortunately, several T–S-fuzzy-
neering Management, City University of Hong Kong, Kowloon, Hong Kong model-based predictive control approaches have been developed
(e-mail: megfeng@cityu.edu.hk).
most recently, such as [31]–[35], where the improvement of tran-
at http://ieeexplore.ieee.org. sient closed-loop control performance in real time is one of the
Digital Object Identifier 10.1109/TFUZZ.2008.2011135 main design objectives.
1063-6706/$25.00 © 2009 IEEE
Fig. 2. Schematic view of an SOFC/capacitor power system.
ternal steam reformer, methane can be converted into hydrogen

(H2 ) even in the presence of some extent of carbon monoxide
(CO). And syngas from biomass and coal gasification is also
able to be transformed to hydrogen based on a water–gas shift
reaction (WGSR). In fact, syngas itself may also serve as the
fuel gas for SOFC more than pure hydrogen [37]. The following
reactions will occur at the anode:
2H2 + 2O2− −→ 2H2 O + 4e (1)
2CO + 2O 2−
−→ 2CO2 + 4e (2)
Fig. 1. System diagram of an SOFC/GT combined cycle plant. while the reaction at the cathode is
Although the control of nonlinear SOFC system is challeng- O2 + 4e −→ 2O2− . (3)
ing due to the slow response under tight safe operation con-
straints, several predictive control strategies have been proposed It should also be noted that a desulfurizer is included in Fig. 1
for rapid load following. A novel structure of data-driven lin- for the compressed fuel gas supplied to fuel cell since sulfur is
ear MPC was used in [8] to deal with systems without full a poison for SOFCs.
online measurement of all output variables, i.e., in the absence Although the SOFC is widely accepted for stationary power
of output measurements for fuel utilization, H2 /O2 ratio, and generation use, critical operation problems often occur in cases
fuel cell pressure difference between the anode and cathode. where the fuel cell is directly connected to the electricity grid
As reported in [36], a fuzzy Hammerstein model was identified or the dc electric user. The reason is that the voltage of the
from the input–output operation data of an SOFC stack, then stand-alone fuel cell would generally fall rapidly with rising
the associated standard predictive controller was applied to the current density. In fact, fuel cells are particularly badly regu-
fuel control of the stand-alone SOFC stack to meet stepwise lated, whereas most electronic and electrical equipments require
power load demands. Unfortunately, the stability issues of these a fairly constant voltage. Moreover, due to the slow response,
closed-loop MPC systems have been largely ignored. However, the fuel cell cannot be directly connected to the load in general.
the control system stability plays a fundamental and vital role A parallel capacitor can be integrated so as to compensate
in plant operation and safety, especially for high-temperature the negative effect of varying external load on the stand-alone
fuel cells. Moreover, the model-reality discrepancy also needs SOFC stack operation, as shown in Fig. 2. The following points
to be considered in the model-based controller design, like the can be noted.
peer research on input-to-state stable (ISS) MPC [23]. In this 1) The capacitor will increase the response time of the SOFC
paper, we will focus on the rapid load tracking problem of the stack/capacitor power system to its fuel/oxygen input
SOFC/capacitor power system via offset-free fuzzy MPC strat- changes. Larger capacitance leads to slower dynamic re-
egy, where the fuzzy dynamic model is identified from the plant sponse of the integrated power system.
operation data and the closed-loop control system stability is 2) The capacitor will guarantee the load tracking of the stack
maintained. current. In steady state, the stack voltage VS becomes con-
Notation: x(k + i|k), i > 0, are the predicted system states at stant, V̇S = 0, and no current flows through the capacitor,
time k + i based on the current state estimation x (k|k) or x
(k). IC = 0; thus, IS = ID + IC = ID .
And for any two vectors x, y ∈ n , x y, means xj − yj ≥ As the first step of our advanced control research, we con-
0 ∀ j = 1, . . . , n, where xj , and yj denote the jth component sider here the ideal fuel processor to produce pure hydrogen
of x, and y, respectively. for the feed of SOFCs as in [2]–[4], and [8]. Here, hydrogen
and oxygen pass through the anode and cathode of fuel cells,
respectively. The hydrogen is consumed by an electrochemical
II. SOFC/CAPACITOR SYSTEM
reaction to generate the water and electricity, and the resulting
The typical diagram of an SOFC/GT system is shown in partial pressures are expressed as pH 2 , pO 2 , pH 2 O . Thus, one has
Fig. 1. SOFC has the outstanding fuel flexibility. With the in- the following schematic SOFC/capacitor dynamics with respect
ZHANG AND FENG: RAPID LOAD FOLLOWING OF AN SOFC POWER SYSTEM VIA STABLE FUZZY PREDICTIVE TRACKING CONTROLLER 359
TABLE I
PARAMETERS IN THE SOFC SYSTEM
to inlet fuel flow rate qF , inlet oxygen flowrate qO , and external

current load ID , as in [2]–[4], and [8]–[10]
1
V̇S = (V0 − VS − rID ) Fig. 3. Voltage/power–current characteristics of open-loop SOFC.
rC
1
q̇H = (qF − qH )
TF
1
ṗH 2 = (qH − 2Kr IS − pH 2 KH 2 ) (4)
τ H 2 KH 2
1
ṗH 2 O = (2Kr IS − pH 2 O KH 2 O )
τ H 2 O KH 2 O
1
ṗO 2 = (qO − Kr IS − pO 2 KO 2 )
τ O 2 KO 2
where VS denotes the stack voltage, qH the hydrogen flow rate,

IS = (V0 − VS )/r

R 0 T0 pH 2 (pO 2 /101.325)1/2
V0 = N0 E0 + ln (5)
2F0 pH 2 O
and Kr = N0 /4F0 . Note that (5) is the well-known Nernst

equation [1], and E0 is the electromotive force (EMF) at stan- Fig. 4. SOFC system response to stepwise fuel flow rate change (+0.1 mol/s).
dard pressure and it is dependent on the change in molar
Gibbs free energy of formation at different stack temperature
T [1], which can be approximated by an empirical formula
E0 = 1.2586 − 0.000252T [10]. Table I contains the SOFC
plant parameters. The nominal stationary voltage/power–current
characteristics of a stand-alone SOFC stack are depicted in
Fig. 3, which shows that SOFC exhibits nonlinear behavior
in a wide operating regime. Obviously, the stack voltage is usu-
ally under significant changes, especially in the low and high
current load, and the overloaded current would lead to the rapid
deterioration of operating stack voltage. Moreover, the stack
can maintain its maximal power output at some operating point,
e.g., when the current load is around 300 A from Fig. 3. As the
load current becomes larger, the power output would decrease
abruptly.
Furthermore, the open-loop dynamic responses of the
SOFC/capacitor system to stepwise changes of inputs qF , qO ,
and ID are given in Figs. 4–6, respectively.
The following variables are also defined for performance Fig. 5. SOFC system response to stepwise oxygen flow rate change
monitoring during SOFC operation. (+0.1 mol/s).
rules as in [24]–[30]
Rl : IF ξ1 (k) is M1l and · · · and ξr (k) is Mrl , THEN

m
l (q −1 )y(k) +
y(k + 1) = A h (q −1 )uh (k),
Bl
h=1
h = 1, . . . , m, l = 1, . . . , L (6)
where Rl denotes the lth fuzzy inference rule, L the number
of inference rules, Mjl are fuzzy sets, y(k) any system output
variable, u(k) ∈ m the system input variables, ξ(k) ∈ r the
premise measurable variables, (A l , B
h ) is the lth local model
l
with the shift operator q defined by q −1 y(k) = y(k − 1), and
−1
l (q −1 ) = âl1 + âl2 q −1 + · · · + âln q −n y +1

A (7)
y
h (q −1 ) = b̂h + b̂h q −1 + · · · + b̂h q −n u +1 .

B (8)
l l1 l2 ln u
Let µl (ξ(k)) be the normalized membership function

Fig. 6. SOFC system response to stepwise load current change (+8 A).
of the inferred fuzzy set M l , where M l = rj =1 Mjl and
L
l=1 µl (ξ) = 1. By using a center-average defuzzifier, product
1) Fuel utilization Uf stands for the ratio between the con- inference, and a singleton fuzzifier [24]–[26], the fuzzy model
sumed hydrogen flow in reaction and the inlet hydrogen (6) can be expressed as follows:
flow, i.e., Uf = qHr /qHin .
m
2) Hydrogen/oxygen ratio rH/O stands for the ratio between µ (q −1 )y(k) +
y(k + 1) = A µh (q −1 )uh (k)
B (9)
inlet hydrogen flow and oxygen flow, i.e., rH/O = qHin /qO
in
. h=1
3) Pressure difference ∆p denotes the difference of hydro-
where µ := µ(ξ) := (µ1 , . . . , µL ) and
gen/oxygen partial pressures in the fuel cell anode and
cathode compartments, i.e., ∆p = pH 2 − pO 2 .
L
L
In order to prevent damage to the electrolyte, the fuel cell µ =

A l (q −1 )
µl A µh =
B lh (q −1 ).
µl B
pressure difference ∆p between the anode and cathode com- l=1 l=1
partments should be kept below 4 kPa under normal operation Thus, (9) also reads
and 8 kPa under transient conditions [4]. Although the stoichio-
metric ratio of hydrogen to oxygen in overall fuel cell reaction
L
y(k + 1) = µl ψ T (k)θl (10)
is 2:1, oxygen excess is always expected in practice to make
l=1
sure hydrogen reacting with oxygen more completely. For this
SOFC system, it is desirable to maintain the ratio of hydrogen where
to oxygen rH/O around 1.145, as in [4] and [8]. In addition, the
ψ(k) = [y(k) · · · y(k − ny + 1) u1 (k) · · ·
normal operating point of fuel utilization Uf is about 85%, and
too low or high utilization degree Uf would lead to potential u1 (k − nu + 1) · · · um (k) · · · tum (k − nu + 1)]T
problems: if Uf < 70%, a large portion of fuel is underused;
(11)
thus, the economical efficiency of SOFC would decrease; other-
wise, when Uf > 90%, the cells may suffer from fuel starvation θl = [âl1 · · · âln y b̂1l1 ··· b̂1ln u ··· b̂m
l1 · · · b̂m
ln u ]
T
.
and might be at the risk of permanent damage.
(12)
III. T–S FUZZY MODELING AND INPUT-TO-STATE STABILITY Note that the fuzzy model (6) or (9) is in the so-called con-
trolled autoregressive moving average (CARMA) form.
In this section, T–S fuzzy system model and the input-to
state stability concept that will facilitate our discussions in the B. Input-to-State Stability
subsequent sections are introduced.
Let , ≥0 , N , and N≥0 denote the real, nonnegative real,
integer and nonnegative integer numbers, respectively.
A. T–S Fuzzy Dynamic Model Definition 1 (Input-to-state stability [39]): A discrete-time
Consider a general multiple-input–multiple-output (MIMO) nonlinear system x(k + 1) = F (x(k), v(k)) is ISS if there exist
nonlinear system with m inputs, u ∈ m , and p outputs, y ∈ p . a KL-function β and a K-function γ such that
This system can be approximated by a set of coupled multiple- x(k) ≤ β(x(0), k) + γ(v) (13)
input–single-output (MISO) discrete-time T–S fuzzy dynamic
models with linear consequents, which is described by L fuzzy for each k ∈ N≥0 and the initial state x(0).
Definition 2 (ISS Lyapunov Function [39]): A continuous where u(k) = [u1 (k) · · · um (k)]T and
function V : n → ≥0 is called an ISS Lyapunov function for
a system x(k + 1) = F (x(k), v(k)) if there exist some K∞ - l (q −1 ) = b̂l1 + b̂l2 q −1 + · · · + b̂ln q −n u +1
B u
functions α1 , α2 , and α3 and a K-function σ such that
b̂lj = [ b̂1lj ··· b̂m
lj ] , j = 1, . . . , nu .
α1 (x) ≤ V (x) ≤ α2 (x) (14)
Note that the system inputs in (19) can include both the manip-
V (F (x, v)) − V (x) ≤ −α3 (x) + σ(v). (15) ulated variables and the reference variables, which are for the
For the definitions of K, K∞ , and KL-functions, refer to [38]. purposes of feedback and feedforward controls, respectively.
Lemma 1 (ISS Lyapunov Characterization [39]): If a system For the subsequent controller design, they will be separated
x(k + 1) = F (x(k), v(k)) admits an ISS Lyapunov function, explicitly as
then the system is ISS.
Rl : IF ξ1 (k) is M1l and · · · and ξr (k) is Mrl , THEN
IV. FUZZY-MODEL-BASED PREDICTIVE TRACKING l (q −1 )y(k) + B
y(k + 1) = A l (q −1 )u(k)
CONTROLLER SYNTHESIS l (q −1 )v(k) + d(k),
+H l = 1, . . . , L (20)
A. Identification of Fuzzy Model and its State-Space Form
where u(k) denote the manipulated inputs for feedback control,
Defining
v(k) the reference inputs for feedforward control, and
    
µ1 µ1 (k)ψ(k) θ1
 .   ..   .  l (q −1 ) = ĥl1 + ĥl2 q −1 + · · · + ĥln q −n v +1 .
H
φ(k) =  ..  ⊗ ψ(k) =  .  Θ =  ..  v
µL µL (k)ψ(k) θL In fact, the previous model with feedforward (20) may also
be transformed into the equivalent state-space form as follows:
where ⊗ is the Kronecker product, it follows that the system of
the SOFC can be approximated by a fuzzy model Rl : IF ξ1 (k) is M1l and · · · and ξr (k) is Mrl , THEN

y(k + 1) = φT (k)Θ. (16) x(k + 1) = Al x(k) + Bl u(k) + Hl w(k) + El d(k)
y(k) = Cl x(k) + Dl d(k), l = 1, . . . , L
When the MISO dataset (u1 (k), . . . , um (k), y(k)), k =
1, . . . , N , associated with plant operation is frequency-rich (21)
enough in the sense that the inverse defined in (17) exists, it
is straightforward to apply any standard parameter identifica- where x(k) ∈ n denotes the system state variables, d(k) ∈ q
tion method such as the least squares method to identify its the modeling error signals or model–plant discrepancies, and
local dynamic model parameters, θl , l = 1, . . . , L, i.e. (Al , Bl , Cl , Dl , El , Hl ) is the associated local linear model. In
particular, the MISO state-space model in observable compan-
Θ = (ΦT Φ)−1 ΦT Y (17) ion form becomes
   
where 0 · · · 0 âln b̂ln
   y(2)     
φT (1)  1 . . . 0 âln −1   b̂ln −1 
   
 ..    Al =  . .  Bl =  . 
Φ= .  Y =  ...  . . .
. . .. .
..   . 
.   . 
φT (N − 1) y(N ) 0 ··· 1 âl1 b̂l1
Note that the resulting modeling performance can be measured    
0 0 0 0
by the variance accounted for (VAF) index [31] given by  0  0
 0 0   
var(Y − Y ) Hl = 
 .. .. .. 
 El =  
 .. 
VAF = 1 − × 100% (18)  . . .  .
var(Y )
ĥl1 ··· ĥln v 1
where Y is the sequence of real plant output measurements and Cl = C = [ 0 · · · 0 1], Dl = D = [0], l = 1, . . . , L
Y the fuzzy model output predictions, with Y = ΦΘ.
Then, from the estimated parameters θl and (12), one can when ny = nu = n and the reference input sequence
identify the fuzzy CARMA model with a signal of modeling
error d(k) w(k) = [v(k)T v(k − 1)T · · · v(k − nv + 1)T ]T .
Rl : IF ξ1 (k) is M1l and · · · and ξr (k) is Mrl , THEN Note that Cl in this case is constant and taken as C, and Dl is
l (q −1)y(k) + B
l (q −1)u(k) + d(k), zero.
y(k + 1) = A l = 1, . . . , L
With the same fuzzy inference principle as for the fuzzy
(19) CARMA model (9), the earlier fuzzy state-space model (21)

can be expressed as follows: 2) Function F(x) = ri=1 fi (x) for each bounded function
fi (x) satisfies the Lipschitz condition
x(k + 1) = Aµ x(k) + Bµ u(k) + Hµ w(k) + Eµ d(k) r

y(k) = Cx(k) + Dd(k) F(x1 ) − F(x2 ) ≤ Mi− Li x1 − x2 (28)
(22) i=1
r
where where Mi− = j =1,j = i Mj and Mj = sup fj (x).
3) Function F(x) = f1 [f2 (x)] satisfies the Lipschitz

L
L
condition
Aµ = µl Al Bµ = µl Bl
l=1 l=1 F(x1 ) − F(x2 ) ≤ (L1 L2 )x1 − x2 . (29)

L
L
Assumption 4: For the T–S fuzzy system F (x, u) =
Eµ = µl El Hµ = µl Hl L
l=1 µl (ξ)[Al x + Bl u], each fuzzy membership function
l=1 l=1
Mil (x) is assumed to be bounded and satisfies the Lipschitz
µ := µ(ξ) := (µ1 , . . . , µL ). (23) condition.
In particular, the bounded triangle and trapezoid membership
functions are easily proved to be Lipschitz. Based on Lemma 2
B. Assumptions and Lipschitz Conditions
and Assumption 4, we can obtain the following result.
In this paper,
we use the nominal T–S fuzzy dynamic model Theorem 1 [35]: For the T–S fuzzy system x(k + 1) =
F (x, u) = Ll=1 µl (ξ) (Al x + Bl u) with an additive uncer- F (x(k), u(k)) with state and input constraints x ∈ X , u ∈ U, if
tainty d to approximate the original nonlinear plant G(x, u) the premise variable functions ξi = Hi (x), i = 1, . . . , r, satisfy
as the Lipschitz condition, there exists a positive finite Lipschitz
constant LF such that
x(k + 1) = G(x(k), u(k)) = F (x(k), u(k)) + Ed(k) (24)
F (x1 , u) − F (x2 , u) ≤ LF x1 − x2 (30)
where d(k) ∈ D is supposed to be bounded. for all x1 , x2 ∈ X , and u ∈ U.
The following assumptions are made here as in [20]–[22].
Assumption 1 (Reference): The trajectory reference yr (k) C. Augmented Fuzzy Observer
is known currently and unknown in the future, and it would
For the fuzzy system with disturbance (21), consider the cor-
reach an unknown steady-state value asymptotically, i.e.,
responding simple augmented fuzzy observer as
limk →∞ yr (k) = ȳr .
Assumption 2 (Disturbance): The disturbance d(k) is al- (k + 1)
x L
Al El (k)
x
ways unknown for each time constant k, but it would reach = µl
+ 1)
d(k O I
d(k)
an unknown steady-state value asymptotically, i.e., limk →∞ l=1
¯
d(k) = d. Bl Hl
Assumption 3 (Detectability): The augmented pair + u(k) + w(k)
O O
1
Al El F
(C D), , l = 1, . . . , L (25) + y (k) − y(k)]
[ (31)
O Iq F2
is detectable, and the following condition holds: y(k) = C x

(k) + Dd(k) (32)
where F = [F 1
T T
F 2 ]T is the overall augmented observer
I − Al −El gain. Defining the estimation error x(k) = x(k) − x(k) and
rank = n + q. (26)
C D
d(k) = d(k) − d(k), one can easily obtain the following ob-
server error dynamics:
Before verifying that the T–S fuzzy dynamic model (21) sat-
(k + 1)
x L
Al El
isfies the Lipschitz condition, the following lemma is introduced = µl
first. + 1)
d(k O I
l=1
Lemma 2 [35]: If functions fi (x), i = 1, . . . , r, satisfy 1
the Lipschitz condition fi (x1 ) − fi (x2 ) ≤ Li x1 − x2 , i = F (k)
x
+ (C D) . (33)
1, . . . , r, then the following
hold. F2
d(k)
1) Function F(x) = ri=1 fi (x) satisfies the Lipschitz con-
dition Then, we have the following stability conditions of the fuzzy
error dynamic system (33).
r
Theorem 2: The error dynamic system (33) is exponentially
F(x1 ) − F(x2 ) ≤ Li x1 − x2 . (27) stable with a decay rate α ∈ (0, 1] if there exists a positive
i=1
definite matrix Z and a matrix G satisfying the following set of subject to

linear matrix inequalities (LMIs):    
I − Aµ −Bµ Eµ d(k) + Hµ w(k)
  tx  
−α2 Z + Qe A T GT
T Z T + C  C O 
=  ȳr − Dd(k)  (36)
l
< 0, l = 1, . . . , L ut
l + GC
ZA −Z O [1 − rH/O ] O
(34) um in ≤ ut ≤ um ax (37)
where ȳr is the desired output trajectory reference, w(k) are

where
the feedforward sequence of the trajectory reference inputs v,
is the current estimate of the disturbance signal d(k), and
d(k)
Al El
l =
A , = (C
C D) (um in , um ax ) are the input constraints.
0 I In particular, as for the stepwise reference trajectory, the ref-
erence inputs v(k) are supposed to remain constant at the current
and Qe ≥ 0 is a given matrix parameter. Moreover, the overall
point within a finite prediction horizon.
observer gain is given by
So, when this target generation problem is feasible, one has

F1
xt = Aµ xt + Bµ ut + Hµ wt + Eµ d(k). (38)
F = = Z −1 G.
F2
By subtracting (38) from the first equation of (22), the resulting
Proof: It is easy to prove this theorem by resorting to invoking transformed system becomes
the Lyapunov stability theorem [40], which is also a special case
and the observer gain F
χ(k + 1) = Aµ χ(k) + Bµ ω(k) + Eµ d(k) (39)
of [35, Th. 2] since both the matrix C
are constant here. Thus, the proof is omitted.
where χ := x − xt , ω := u − ut , and d := d − d.
Then, one
Remark 1: As shown in Theorem 2 and its proof, the sta-
can use the nominal model of the transformed system for pre-
ble overall fuzzy observer error system implies that each local
l = 1, . . . , L, is also stable.
l + F C), diction as follows:
observer error system (A
In fact, we also have the following result for the augmented χ(k + i + 1|k) = Aµ χ(k + i|k) + Bµ ω(k + i|k), i≥0
fuzzy observer. (40)
Lemma 3: Under Assumption 3, if the number of disturbance where χ(k|k) = χ(k). Its infinite horizon predictive control per-
states is equal to the number of outputs, q = p, then the partial formance cost can be defined as
observer gain F 2 in (31) is of full rank.
∞

Proof: The lemma can be proved by the argument of contra-
J0∞ (k) = W (χ(k + i|k), ω(k + i|k))
diction similar to that for [35, Lemma 4], and is thus omitted.
i=0

N −1
D. Fuzzy Predictive Tracking Control = W (χ(k + i|k), ω(k + i|k))
i=0
In order to achieve the offset-free output tracking perfor-
∞

mance, a constrained target generator has been developed in
[20]–[22], where the desired targets of system states and control + W (χ(k + i|k), ω(k + i|k)) (41)
i=N
inputs in steady state are obtained with respect to the lumped
disturbance estimates. Here, we extend this idea to the target where W (χ(k + i|k), ω(k + i|k)) = χ(k + i|k)T Qχ(k + i|k) +
generation for fuzzy systems. ω(k + i|k)T Rω(k + i|k).
Meanwhile, from the viewpoint of the SOFC operation, the Define a quadratic function V (χ(k + i|k)) = χ(k + i|k)T
fuel cell pressure difference between the anode and cathode P (k + i|k)χ(k + i|k), and suppose it satisfies
compartments should be below 4 kPa under normal operation
and 8 kPa under transient conditions [4]. It is desirable to V (χ(k + i + 1|k)) − V (χ(k + i|k))
maintain the ratio of hydrogen to oxygen rH/O around 1.145
≤ −W (χ(k + i|k), ω(k + i|k)) ≤ 0. (42)
to keep the fuel cell pressure difference close to zero, i.e.,
qF − rH/O qO = u1 − rH/O u2 = [1 − rH/O ]u ≈ 0. This con- Summing (42) from i = N to ∞ and requiring χ(∞|k) = 0 or
trol objective can also be embedded in the corresponding con- V (χ(∞|k)) = 0, it follows that
strained target generator.
∞

For the overall fuzzy system (22), its steady-state and input
target vectors xt and ut can be determined from the solution of JN∞ (k) = W (χ(k + i|k), ω(k + i|k))
i=N
the following quadratic programming (QP) problem:
≤ V (χ(k + N |k))
min [(yt − ȳr )T Qt (yt − ȳr )] (35) = χ(k + N |k)T P (k + N |k)χ(k + N |k). (43)
x t ,u t
By substituting (43) into (41), one can get Then, the MPC performance cost can be reformulated as
TA χ(k) + TB π(k)]

J(χ(k), π(k)) = [TA χ(k) + TB π(k)]T Q[

N −1
J0∞ (k) ≤ W (χ(k + i|k), ω(k + i|k)) + V (χ(k + N |k))
i=0
+ π(k)T Rπ(k) + [AN
µ χ(k) + AB π(k)]
T
=: J(χ(k), π(k)). (44) × P [AN

µ χ(k) + AB π(k)] (50)
Here, J(χ(k), π(k)) gives an upper bound for so we J0∞ (k),
where Q = IN ⊗ Q and R = IN ⊗ R. Moreover, with the feed-
can formulate MPC as an equivalent minimization problem on
J(χ, π) with respect to the optimal control sequence back control law ω(k + i|k) = Kχ(k + i|k), the closed-loop
fuzzy control system within the terminal region becomes
π ∗ (k) = {ω ∗ (k|k), ω ∗ (k + 1|k), . . . , ω ∗ (k + N − 1|k)}.
(45)
L
With (42) and (43), there always exists a terminal constraint set χ(k + i + 1|k) = µl (Al + Bl K)χ(k + i|k). (51)
E such that χ(k + N |k) ∈ E; thus, one has that the terminal con- l=1
trol law ω(k + i|k) = Kχ(k + i|k), i ≥ N , and the associated Before showing the input-to-state stability of the proposed
terminal cost V (χ(k + N |k)) = χ(k + N |k)T P χ(k + N |k). MPC system, we present two related lemmas.
For the closed-loop predictive control system, let Lemma 4 (Lipschitz Quadratic Functions): For the quadratic
Ψ(k) := J(χ(k), π ∗ (k)) = inf J(χ(k), π(k)). (46) function
π (k )
W (x, u) = xT Qx + uT Ru, Q, R > 0
Obviously, Ψ(k) is the optimal performance cost at time k. Also,
from the definition of J(χ(k), π(k)), at time instant k + 1, one there exists a finite Lipschitz constant LW > 0 such that
has

N W (x1 , u) − W (x2 , u) ≤ LW x1 − x2 (52)
J(χ(k + 1), π(k + 1)) = W (χ(k + i|k), ω(k + i|k))
i=1
for all x1 , x2 ∈ X , and u ∈ U. Similarly, for the quadratic func-
tion V (x) = xT P x, P > 0, there exists a finite Lipschitz con-
+ V (χ(k + N + 1|k)) (47)
stant LV > 0 such that
for the shifted control sequence π(k + 1) = {ω ∗ (k + 1|k), . . . ,
ω ∗ (k + N − 1|k), Kχ(k + N |k)}. V (x1 ) − V (x2 ) ≤ LV x1 − x2 (53)
With the overall predictive fuzzy dynamic model χ(k + i +
for all x1 , x2 ∈ X , and u ∈ U.
1|k) = Aµ χ(k + i|k) + Bµ ω(k + i|k) and an open-loop con-
Lemma 5 (Model Prediction Difference Bound): Consider
trol sequence π(k), we can easily derive the multistep-ahead
the predictive fuzzy model χ(k + i + 1|k) := F (k + i|k) =
state and output prediction
F (χ(k + i|k), ω(k + i|k)), where χ(k|k) = χ(k) such that
χ(k + N |k) = AN (48) Theorem 1 holds. For a sequence of control inputs, the mul-
µ χ(k) + AB π(k)
tistep nominal model prediction difference is bounded by
Y (k) = TA χ(k) + TB π(k) (49)
  χ(k + i|k + 1) − χ(k + i|k) ≤ Li−1
χ(k|k) F Eµ d(k) (54)
 χ(k + 1|k) 
  where LF is the Lipschitz constant of the T–S fuzzy system
Y (k) = 
 .. 
 denotes the disturbance estimation error from the aug-
 .  and d(k)
mented observer.
χ(k + N − 1|k) Proof: Due to the Lipschitz condition of the T–S fuzzy model,
  the result in (54) is readily proved from (39)–(40) and Theorem 2
ω(k|k)
  by the induction as in [35].
 ω(k + 1|k) 
π(k) = 
 .. 

Now we present the main result in this paper.
 .  Theorem 3 (Offset-Free Fuzzy MPC): For system (21) subject
to the input constraints um in u um ax , under Assumptions
ω(k + N − 1|k)
1–3, the closed-loop output feedback fuzzy predictive control
B = [ AN
A µ
−1
Bµ ··· Aµ Bµ Bµ ] system, with objective function J(χ(k), π(k)) in (44), aug-
   O ··· O O  mented fuzzy observer (31)–(32), and target generation pro-
In cedure (35)–(37) is ISS and achieves the offset-free reference
 A   Bµ ··· O O
 µ    tracking performance if all the following conditions are satis-

TA =   TB =  . fied.
 ..   .. .. 
 .   . . O O 1) The LMIs (34) for observer design in Theorem 2 admit a
−1 −2
AN
µ AN
µ Bµ ··· Bµ O positive definite matrix Z and a matrix G.
> 0, a matrix Y ,
2) There exist a positive definite matrix X the sufficient conditions (55) can be rewritten as
and γ > 0 satisfying the LMIs (Al X + Bl Y )T X
−1 (Al X
+ Bl Y ) − X
−1 + γ −1 X
T QX

 
X + γ −1 Y T RY < 0, l = 1, . . . , L. (61)
 
Al X + Bl Y X  Multiplying (61) on both sides by γ 1/2
X = γP −1
−1 with X
  > 0, l = 1, . . . , L

 X O γQ−1 
 and Y = K X yields
Y O O γR−1 (Al + Bl K)T P (Al + Bl K) − P + Q + K T RK < 0.

(62)
(55) Similar to the proof of [33, Th. 3.2], for the normalized member-

ū2j Uj Y ship function µl (ξ) and the predictive system states χ(k + i|k),
≥ 0, j = 1, . . . , m the earlier inequalities imply that (63), as shown at the bottom of
Y T UjT
X this page, holds where 2X T RY ≤ X T RX + Y T RY , R > 0,
(56) is used. By defining a global quadratic Lyapunov function
V (χ(k + i|k)) = χ(k + i|k)T P χ(k + i|k) (64)
such that there exists a constraint set
with the closed-loop fuzzy system (51), (63) implies
E = {z ∈ n |z T P z ≤ γ} (57)
V (χ(k + i + 1|k)) − V (χ(k + i|k))
where the tight input constraint ūj = min{|um ax − ut |j ,
< −W (χ(k + i|k), ω(k + i|k)) ≤ 0. (65)
|um in − ut |j }, and Uj is the jth unit row vector with the
nonzero element in the jth position Uj = [0 · · · 0 1 0 Summing (65) from i = N to ∞, with χ(∞|k) = 0 or
· · · 0]. V (χ(∞|k)) = 0, it follows that
3) There exist feasible solutions (xt , ut ) to the target tracking ∞
quadratic programming problem (35) subject to (36)–(37). W [χ(k + i|k), ω(k + i|k)]
4) There exists a feasible solution π to the following i=N
quadratic programming problem at time instant k < χ(k + N |k)T P χ(k + N |k) = V (χ(k + N |k)) ≤ γ (66)

1 T T which thus constitutes the constraint set (57) or {χ ∈ n |
min π Θπ + Ξ π
π 2 V (χ) ≤ γ}.
subject to Aπ B (58) Now, we check if each element of the control inputs satisfies
the constraint |uj (k + i|k)| ≤ min{|um ax |j , |um in |j }, i ≥ 0,
TB + R
Θ = 2(TBT Q +A
TB P A
B ) j = 1, . . . , m. For any i within the finite horizon N , the in-
put constraints are satisfied since Υ ⊗ [um in − ut ] π Υ ⊗
Ξ = 2[TBT Q(
TA χ(k)) + A
TB Ps (AN
µ χ(k))]
[um ax − ut ], Υ = [1 · · · 1]T . Otherwise, beyond the finite

IN ×m Υ ⊗ [um ax − ut ] horizon i ≥ N , χ(k + i|k) belongs to the constraint set E =
A= B= > 0}, which is guaranteed by the in-
−1 z ≤ 1, X
−IN ×m −Υ ⊗ [um in − ut ] {z ∈ n |z T X
equality (57). And in this case, there exists a feedback control
where χ(k) = x (k) − xt and Υ = [1 · · · 1]T . law ω(k + i|k) = Kχ(k + i|k) = Y X −1 χ(k + i|k) such that
Meanwhile, the real-time control action is chosen as u(k) =
max |ωj (k + i|k)|2
π ∗ (1 : m) + ut , and the feedback controller and fuzzy observer i≥N
gains are given by
= max |(Y X
−1 χ(k + i|k))j |2
P = γX K = Y X
−1 −1 (59)
i≥N
1 ≤ max |(Y X
−1/2 X
−1/2 z)j |2
F z ∈E
F = = Z −1 G. (60)
F2
≤ (Y X −1/2 z)j 22 ≤ Uj Y X
−1/2 )j 22 (X −1 Y T UjT
Proof: As shown in Theorem 2, the sufficient condition 1)
≤ ū2j , j = 1, . . . , m (67)
can guarantee the observer error system to be asymptotically
stable. Moreover, when both the sufficient conditions 2)–4) are where Uj = [0 · · · 0 1 0 · · · 0] and ūj = min{|um ax −
satisfied for the nominal system (40) with χ(k) = x(k) − xt ut |j , |um in − ut |j }. Therefore, the input constraints are satis-
and ω(k) = u(k) − ut , by resorting to the Schur complements, fied if the LMIs (56) are feasible.
T

L
L
µl (Al + Bl K)χ(k + i|k) P µj (Aj + Bj K)χ(k + i|k) − χ(k + i|k)T P χ(k + i|k)
l=1 j =1
< −χ(k + i|k)T Qχ(k + i|k) − [Kχ(k + i|k)]T R [Kχ(k + i|k)] < 0 (63)
As mentioned before, it can be found that the difference of Thus, the optimal cost Ψ(k) serves as an ISS Lyapunov function
the optimal cost Ψ(k) also satisfies for the closed-loop system by invoking Lemma 1. Hence, the
resulting fuzzy MPC control system is ISS.
∆Ψ(k) = J(χ(k + 1), π ∗ (k + 1)) − J(χ(k), π ∗ (k)) Then we turn to the proof of offset-free output tracking prop-
≤ V (χ(k + N + 1|k + 1)) − V (χ(k + N |k + 1)) erty. Suppose that x
(∞), and d(∞) denote as the steady-state
estimates of the plant states and lumped disturbances, u(∞)
+ V (χ(k + N |k + 1)) − V (χ(k + N |k))
is the steady-state control input by the regulator, and y(∞) is
+ W (χ(k + N |k + 1), ω(k + N |k + 1)) the measured steady-state process output. Hence, from the aug-
mented observer (31)–(32), we have
− W (χ(k|k), ω ∗ (k|k))
(∞) = Aµ x
x
(∞) + Bµ u(∞) + Eµ d(∞)

N −1
+ [W (χ(k + i|k + 1), ω(k + i|k + 1))
(∞) + Dd(∞)
+ F 1 [C x − y(∞)] (69)
i=1

d(∞)
= d(∞) + F 2 [C x
(∞) + Dd(∞) − y(∞)]. (70)
− W (χ(k + i|k), ω ∗ (k + i|k))].
Taking (54) of Lemma 5 into account, and with Lemma 4, one Since F 2 is shown to be of full rank from Lemma 3, (69) implies
can conclude that
(∞) + Dd(∞)
y(∞) = C x (71)
V (χ(k + N |k + 1)) − V (χ(k + N |k)) and accordingly
≤ LV LN −1
Eµ d(k) (∞) = Aµ x
x
(∞) + Bµ u(∞) + Eµ d(∞). (72)
F
= d(k) − d(k).
where d(k) From the definition of π(k + 1), it If the target generation procedure in (35) and (37) has a feasible
follows that ω(k + i|k + 1) = ω ∗ (k + i|k), i = 1, . . . , N − 1; solution at steady state, the following equality holds:
consequently, we have
xt = Aµ xt + Bµ ut + Eµ d(∞). (73)
W (χ(k + i|k + 1), ω(k + i|k + 1))
Subtracting (73) from (72) yields
− W (χ(k + i|k), ω ∗ (k + i|k))
(∞) − xt = Aµ [
x x(∞) − xt ] + Bµ [u(∞) − ut ]. (74)
≤ LW χ(k + i|k + 1) − χ(k + i|k)
From the property of stabilizing MPC, the feasibility of the
≤ LW Li−1
F χ(k + 1|k + 1) − χ(k + 1|k) regulator conditions implies that there is a feedback control law
≤ LW Li−1 x(∞) − xt ]
u(∞) − ut = K[
F Eµ d(k).
Then, one can get and substituting this feedback law into (74) results in
∆Ψ(k) ≤ [V (χ(k + N + 1|k + 1)) − V (χ(k + N |k + 1))

L
µl (Al + Bl K − I)[
x(∞) − xt ] = 0. (75)
+ W (χ(k + N |k + 1), ω(k + N |k + 1))] l=1
−1
As shown in the previous controller design, the closed-loop

N
+ LV L N −1
+ LW L i−1
Eµ d(k) system is guaranteed to be stable, and thus, the eigenvalues of
F F
i=1 each local system Al + Bl K stand within the unit circle. Hence,
− W (χ(k|k), ω ∗ (k|k)). (68) the only solution of (75) is
Since the stability constraint (65) also holds for i = N , the terms (∞) − xt = 0.
x
in the square brackets of (68) are thus less than zero. Then, one Also, with the feasible target generation procedure at steady
can conclude state, we have
−1
N −1 LN −1
ȳr = Cxt + Dd(∞). (76)
∆Ψ(k) ≤ LV LF + LW F
Eµ d(k)
LF − 1
So, the steady-state output tracking error between (71) and (76)
− [χ(k|k)T Qχ(k|k) + ω ∗ (k|k)T Rω ∗ (k|k)] becomes

≤ LJ d(k) − aχ(k)2 e = y(∞) − ȳr = C[
x(∞) − xt ] = 0. (77)
is the disturbance estimation error
where d(k) Therefore, the offset-free property of the closed-loop system is
guaranteed, and thus, the theorem is proved.
−1 −1
LJ = [LV LN
F + LW (LN
F − 1)/(LF − 1)]Eµ Remark 2: In Theorem 3, the tight input constraints are
considered in the computation of conservative terminal cost
and a = λm in (Q). And obviously would
since the state/disturbance estimation errors x(k)/d(k)
Ψ(k) ≥ λm in (Q)χ(k)2 . affect the scope of the associated terminal constraint set. So,
Fig. 7. Excitation input signals to SOFC system. Fig. 8. Output response signals of SOFC system.
the smaller feasible terminal region from the tight input con-
straints u ū is employed to prevent the transient predictive
states out of the hard limit um in , um ax . Also, due to the exis-
tence of estimation errors x
(k)/d(k), the transient performance
of model-based control system is only suboptimal.
Remark 3: Note that the comprehensive explicit reachability
analysis is a challenging issue in general for a T–S fuzzy sys-
tem due to its nonlinearity nature. It is well known that local
reachability and observability do not always imply the global
ones of the overall fuzzy system. In this paper, if the sufficient
conditions of Theorems 2 and 3 are satisfied, the reachability of
the system is implicitly ensured.
V. SIMULATION RESULTS
A. Fuzzy Model Identification of SOFC System
To identify the fuzzy model of the SOFC/capacitor sys-
tem, some amplitude-modulated pseudorandom binary signals Fig. 9. Monitored variables of SOFC system.
(APRBSs) are employed as excitation signals for the manipu-
lated variables, such as fuel input flow qF and oxygen input
flow qO , which are shown in Fig. 7. Here, both qF and qO are under the safe range; otherwise, it would destroy the cell itself.
supposed to be of the uniform distribution between limits of 0.5 The detailed safety region is normally provided by the manu-
and 0.8. Meanwhile, according to the stationary characteristics facturer. The maximal value of the pressure difference in this
of SOFC, another input variable ID may be chosen as study is assumed to be 8 kPa, as given in [4] and [8].
With the earlier excitation signals for inputs, the correspond-
qF Ufi ing output response signals for stack voltage VS , stack current
ID =
2Kr IS , and power output PO are presented in Fig. 8, while the
responses of the monitored variables, ∆p, Uf , and rH/O are
where Ufi is subject to the uniform distribution between
shown in Fig. 9. Obviously, the transient pressure difference
the lower and upper limits of fuel utilization, 0.7 and 0.9,
∆p is always less than 8 kPa, so SOFC still operates within the
respectively.
safety region during the identification process.
Here, the constraints on qF and qO are used to reflect possible
Choosing the membership functions for the premise variable
physical restriction in practice. Since the fuel cell stack capac-
ID as
ity is limited, it is impossible to inject an arbitrary amount of
fuel and oxygen into the anode and cathode with limited spaces. 

 1, if ID ≤ 200
When the fuel cell is shut down, the fuel/oxygen flow rate is 

supposed to be zero in theory. When the excitation signal is im- ID − 200
“Low” : M 1 = 1− , if 200 < ID < 250
posed on SOFC for identification purposes, it is still desirable 
 250 − 200


to keep the pressure difference between the anode and cathode 0, if ID ≥ 250.
Fig. 10. Membership function of current load ID in fuzzy SOFC model.

2
1 − M 1 , if ID ≤ 250
“Middle” : M =
1 − M 3 , if ID ≥ 300.


 0 if ID ≤ 250


ID − 250 Fig. 11. Dynamic fuzzy modeling performance of SOFC system (solid: orig-
“High” : M 3 = , if 250 < ID < 300 inal SOFC voltage; dashed: fuzzy model prediction).

 300 − 250


1, if ID ≥ 300.
which are shown in Fig. 10, and using the method described in
Section IV-A, it is easy to identify the T–S fuzzy dynamic model
(78) of the SOFC/capacitor power system as follows.
Rule 1: IF current load ID is Low, THEN
VS (k + 1) = 2.1219VS (k) − 1.3934VS (k − 1)
+ 0.27077VS (k − 2) + 0.34582qF (k)
+ 0.14722qF (k − 1) + 0.33459qF (k − 2)
+ 2.4031qO (k) − 1.051qO (k − 1)
− 1.6142qO (k − 2) − 0.069504ID (k)
+ 0.1123ID (k − 1) − 0.043102ID (k − 2).
Rule 2: IF current load ID is M iddle, THEN
VS (k + 1) = 2.0877VS (k) − 1.3433VS (k − 1) Fig. 12. Validation of fuzzy model for SOFC system under stepwise input
changes (solid: original SOFC voltage; dashed: fuzzy model prediction).
+ 0.25436VS (k − 2) + 0.032884qF (k)
+ 0.48617qF (k − 1) + 0.0030969qF (k − 2)
Then the VAF index in (18) is employed to evaluate the fuzzy
+ 2.6161qO (k) − 0.92155qO (k − 1) modeling performance. Here, with the series–parallel identifi-
− 1.6953qO (k − 2) − 0.069017ID (k) cation structure [42] between fuzzy model and plant, the corre-
sponding dynamic fuzzy modeling results are shown in Fig. 11
+ 0.10947ID (k − 1) − 0.040451ID (k − 2). with the associated VAF index value as 99%. Moreover, step-
Rule 3: IF current load ID is High, THEN wise input changes are imposed on both the SOFC plant and its
fuzzy model for further validation, with results shown in Fig. 12.
VS (k + 1) = 2.1041VS (k) − 1.3698VS (k − 1) It can be observed that the approximation performance of the
+ 0.26532VS (k − 2) + 0.22085qF (k) identified fuzzy SOFC model (78) is excellent.
+ 0.3106qF (k − 1) + 0.16448qF (k − 2)
B. Fuzzy Predictive Control of SOFC System
+ 2.6499qO (k) − 0.95128qO (k − 1)
As mentioned in Section II, the external load will directly
− 1.6032qO (k − 2) − 0.070544ID (k) affect the operation of the SOFC power plant. In fact, to be
an independent power source candidate, the output voltage of
+ 0.11077ID (k − 1) − 0.041867ID (k − 2).
the SOFC/capacitor system is expected to be at a desired con-
(78) stant value for stable operation of external electrical equipment.
Fig. 13. Diagram of fuzzy MPC/SOFC/capacitor system structure.
Therefore, it is desirable to integrate a controller into the stand-

alone SOFC plant to achieve the aforementioned objective.
To develop flexible advanced control strategies, the specific
closed-loop control structure diagram, shown in Fig. 13, has
been proposed in this paper. It should be noted in Fig. 13 that
three system inputs are classified as follows according to the
analysis of SOFC/capacitor system:
1) manipulated inputs: fuel flow qF , oxygen flow qO ;
2) feedforward input: load current demand ID . Fig. 14. Output responses in offset-free fuzzy predictive control system
The proposed fuzzy MPC strategy is based on the fuzzy of SOFC (solid: system output; dotted-dash: stack power output; dashed:
reference).
dynamic model (78), and the detailed predictive controller
design method has been presented in Theorem 3. The cor-
responding controller parameters are chosen in this case as
N = 10, Qt = 0.1I1 , Q = 0.1I2 , and R = 15I2 . Due to the
limited reaction capacity and space in the fuel cell stack, the
fuel and oxygen flow rates would be under some practical lim-
itations, and it is assumed here that the amplitude constraints
on manipulated inputs qF /qO are ujm ax = 1.2, ujm in = 0, with
tight constraints ūj = 0.5, j = 1, 2, for controller design. Then,
by solving the LMI conditions (55)–(56) in Theorem 3, we can
obtain the corresponding characteristic matrices P and K to
construct the terminal cost for MPC
 
8.0441 7.9584 7.8786
 
P =  7.9584 8.9429 9.4989 
7.8786 9.4989 11.0709

−0.2956 −0.3146 −0.3287
K= .
−0.0742 −0.1887 −0.2877
When α = 0.95 and Qe = 20I3 , the augmented observer gain Fig. 15. Manipulated input changes in offset-free fuzzy predictive control
system of SOFC.
can be obtained from Theorem 2 as
F = [ −0.2635 1.3689 −2.5093 −0.4062 ]T .
flow controller, the closed-loop power response to load current
To emulate the practical needs of dc electric users, the power changes is also much quicker than the open-loop case. In fact, the
load is supposed under stepwise changes while the voltage refer- settling time is about 10 and 100 s for the closed-loop and the
ence is assumed to be 245 V. Accordingly, the current reference open-loop systems, respectively. Meanwhile, the closed-loop
for the SOFC/capacitor power system will also change step- control input responses are given in Fig. 15, and the monitored
wisely, as shown in the dashed lines of Figs. 14 and 15. variables changes are kept within the safety range, as depicted
By applying the proposed fuzzy predictive controller design in Fig. 16.
in Theorem 3, the satisfactory control performances are achieved As for the tunning of controller parameters, larger Q/R ratio
for the closed-loop SOFC system with the suboptimal transient usually improves the tracking performance, but it sometimes
reference tracking ensured, as shown in Fig. 14. It can be ob- leads to oscillatory control actions and wastes more reactant
served from the third subfigure of Fig. 14 that the power output supply energy. And with larger prediction horizon length N ,
of the integrated SOFC/ccapacitor system (solid line) can follow better control performance may be achieved, but at greater op-
the power load change more quickly than stack power output timization cost.
(dotted-dash line) due to the existence of an additional capacitor, As shown in Table II (given at the top of the next page), the
and the latter is with larger transient overshoots and significant integral rms performance index does not change much, which
oscillations. On the other hand, with the proposed fuel/oxygen indicates that the proposed control system is not very sensitive
TABLE II
EVALUATION OF CONTROLLER PARAMETER CHANGES
tification. Both the rapid power load following and safe SOFC
operation requirements are taken into account in the proposed
predictive tracking control system design. The simulation re-
sults have validated the satisfactory load tracking performance
of the proposed SOFC control strategy even in the presence
of input constraints. One future research topic is to develop a
robust stable fuzzy model predictive control system for more
complex SOFC power systems under grid connection, such as
the control of plant-wide systems including SOFC stack with
temperature dynamics, buck–boost or boost dc/dc converters,
and forced-commutated dc-ac voltage source inverter. Another
possible research is to enhance the fuzzy modeling ability, e.g.,
adaptive model identification for time-varying dynamics. Mean-
while, it is desirable to study adaptive predictive control when
the plants are slowly time-varying.
REFERENCES
Fig. 16. Monitoring performances in offset-free fuzzy predictive control sys-
tem of SOFC. [1] J. Larminie and A. Dicks, Fuel Cell Systems Explained, 2nd ed. West
Sussex, U.K: Wiley, 2003.
TABLE III [2] V. Knyazkin, L. Soder, and C. Canizares, “Control challenges of fuel cell-
COMPARISON OF OF CONTROL METHODS driven distributed generation,” presented at the IEEE Bologna Power-Tech
Conf., Italy, Jun. 23–26, 2003.
[3] N. Akkinapragada and B. H. Chowdhury, “SOFC-based fuel cells for load
following stationary applications,” in Power Symp. (NAPS 2006), 38th
North Amer., Carbondale, IL, pp. 553–560.
[4] Y. Zhu and K. Tomsovic, “Development of models for analyzing the load-
following performance of microturbines and fuel cells,” Electr. Power
Syst. Res., vol. 62, pp. 1–11, 2002.
[5] R. Kandepu, L. Imsland, B. A. Foss, C. Stiller, B. Thorud, and O. Bolland,
to controller parameters. In fact, even for the most conservative “Modeling and control of a SOFC-GT-based autonomous power system,”
case, the stack power load following performance in Fig. 14 is Energy, vol. 32, pp. 406–417, 2007.
still quite satisfactory. [6] F. Mueller, F. Jabbari, R. Gaynor, and J. Brouwer, “Novel solid oxide
fuel cell system controller for rapid load following,” J. Power Sources,
Regarding the comparison in Table III, the fuel/hydrogen flow vol. 172, pp. 308–323, 2007.
rate qH and oxygen flow qO in [4] are fixed to be rH/O = 1.145, [7] J. T. Pukrushpan, A. G. Stefanopoulou, and H. Peng, Control of Fuel Cell
so their control system has only one active manipulated vari- Power Systems: Principles, Modeling, Analysis and Feedback Design.
New York: Springer-Verlag, 2004.
able and slow load following response. While in this paper and [8] X. Wang, B. Huang, and T. Chen, “Data-driven predictive control for solid
in [8], there are two independent control inputs qF and qO in oxide fuel cells,” J. Process Control, vol. 17, pp. 103–114, 2007.
data-driven control system, in general, more control freedom [9] J. Padulles, G. W. Ault, and J. R. McDonald, “An integrated SOFC plant
dynamic model for power systems simulation,” J. Power Sources, vol. 86,
offers more opportunities for performance improvement. For pp. 495–500, 2000.
the predictive controller in [8], evident steady-state load follow- [10] A. M. Murshed, B. Huang, and K. Nandakumar, “Control relevant mod-
ing errors was observed in [8, Figs. 9 and 11] when the power eling of planer solid oxide fuel cell system,” J. Power Sources, vol. 163,
no. 2, pp. 830–845, 2007.
demand takes stepwise changes. Since, in this paper, we have [11] E. F. Camacho and C. Bordons, Model Predictive Control, 2nd ed. New
taken into account the offset-free control performance in con- York: Springer-Verlag, 2004.
troller design and synthesis, so our proposed predictive control [12] J. M. Maciejowski, Predictive Control With Constraints. New York:
Prentice-Hall, 2002.
system can maintain the rapid load following ability with zero [13] S. J. Qin and T. A. Badgwell, “A survey of industrial model predictive
tracking error. control technology,” Control Eng. Practice, vol. 11, pp. 733–764, 2003.
[14] B. Kouvaritakis and M. Cannon, Nonlinear Predictive Control: Theory
VI. CONCLUSION and Practice. London, U.K.: The Institution of Electrical Engineers,
2001.
To improve the load tracking performance of a stand-alone [15] D. Q. Mayne, J. B. Rawings, and C. V. Rao, “Constrained model predictive
control: Stability and optimality,” Automatica, vol. 36, pp. 789–814, 2000.
SOFC stack, a constrained offset-free fuzzy predictive control [16] C. Aurora, L. Magni, R. Scattolini, P. Colombo, F. Pretolani, and G. Villa,
approach has been developed for the integrated SOFC/capacitor “Predictive control of thermal power plants,” Int. J. Robust Nonlinear
power system based on its fuzzy model from data-driven iden- Control, vol. 14, pp. 415–433, 2004.
[17] M. V. Kothare, V. Balakrishnan, and M. Morari, “Robust constrained [39] Z.-P. Jiang and Y. Wang, “Input-to-state stability for discrete-time nonlin-
model predictive control using linear matrix inequalities,” Automatica, ear systems,” Automatica, vol. 37, pp. 857–869, 2001.
vol. 32, no. 10, pp. 1361–1379, 1996. [40] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan, Linear Matrix
[18] L. Ozkan, M. V. Kothare, and C. Georgakis, “Model predictive control of Inequalities in System and Control Theory. Philadephia, PA: SIAM,
nonlinear systems using piecewise linear models,” Comput. Chem. Eng., 1994.
vol. 24, pp. 793–799, 2000. [41] M. J. Grimble, “Non-linear generalized minimum variance feedback, feed-
[19] M. Lazar, W. P. M. H. Heemels, S. Weiland, and A. Bemporad, “Stabilizing forward and tracking control,” Automatica, vol. 41, pp. 957–969, 2005.
model predictive control of hybrid systems,” IEEE Trans. Autom. Control, [42] K. S. Narendra and K. Parthasarathy, “Identification and control of dynam-
vol. 51, no. 11, pp. 1813–1818, Nov. 2006. ical systems using neural networks,” IEEE Trans. Neural Netw., vol. 1,
[20] K. R. Muske and J. B. Rawlings, “Model predictive control with linear no. 1, pp. 4–27, Mar. 1990.
models,” AIChE J., vol. 39, no. 2, pp. 262–287, 1993.
[21] K. R. Muske and T. A. Badgwell, “Disturbance modeling for offset-free
linear model predictive control,” J. Process Control, vol. 12, pp. 617–632,
2002.
[22] G. Pannocchia and J. B. Rawlings, “Disturbance models for offset-free
model predictive control,” AIChE J., vol. 49, pp. 426–437, 2003.
[23] D. Limon, T. Alamo, and E. F. Camacho, “Input-to-state stable MPC
for constrained discrete-time nonlinear systems with bounded additive Tiejun Zhang (S’04–M’08) was born in Wuxi,
uncertainties,” presented at the 41nd IEEE Conf. Decision Control, Las China, in 1979. He received the B.Eng. and M.Sc.
Vegas, NV, 2002. degrees in thermal power engineering from the
[24] T. Takagi and M. Sugeno, “Fuzzy identification of systems and its applica- Southeast University, Nanjing, China, in 2001 and
tion to modelling and control,” IEEE Trans. Syst., Man, Cybern., vol. 15, 2004, respectively, and the Ph.D. degree in systems
no. 1, pp. 116–132, Jan./Feb. 1985. and control from the City University of Hong Kong,
[25] S. G. Cao, N. W. Rees, and G. Feng, “Analysis and design for a class Kowloon, Hong Kong, in 2008.
of complex control systems, Part I: Fuzzy modelling and identification,” He is currently a Postdoctoral Research Associate
Automatica, vol. 33, no. 6, pp. 1017–1028, 1997. at Rensselaer Polytechnic Institute, Troy, NY. His
[26] S. G. Cao, N. W. Rees, and G. Feng, “Analysis and design for a class of current research interests include output tracking and
complex control systems, Part II: Fuzzy controller design,” Automatica, regulation, fuzzy systems and control, particularly
vol. 33, no. 6, pp. 1029–1039, 1997. dynamic modeling and predictive control of clean energy and power systems.
[27] J. Yen, L. Wang, and C. W. Gillespie, “Improving the interpretability Dr. Zhang was the recipient of the Outstanding M.Sc. Dissertation Award of
of TSK fuzzy models by combining global learning and local learning,” Jiangsu Province of China.
IEEE Trans. Fuzzy Syst., vol. 6, no. 4, pp. 530–537, Aug. 1998.
[28] T. A. Johansen, R. Shorten, and R. Murray-Smith, “On the interpretation
and identification of dynamic Takagi–Sugeno fuzzy models,” IEEE Trans.
Fuzzy Syst., vol. 8, no. 3, pp. 297–313, Jun. 2000.
[29] T. A. Johansen and R. Babuska, “Multiobjective identification of Takagi–
Sugeno fuzzy models,” IEEE Trans. Fuzzy Syst., vol. 11, no. 6, pp. 847–
860, Dec. 2003.
[30] G. Feng, “A survey on analysis and design of model-based fuzzy control
systems,” IEEE Trans. Fuzzy Syst., vol. 14, no. 5, pp. 676–697, Oct. 2006.
[31] J. A. Roubos, S. Mollov, R. Babuska, and H. B. Verbruggen, “Fuzzy
model-based predictive control using Takagi-Sugeno models,” Int. J. Gang Feng (S’90–M’92–SM’95–F’09) received the
Approximate Reason., vol. 22, pp. 3–30, 1999. B.Eng. and M.Eng. degrees in automatic control (of
[32] A. N. Venkat, P. Vijaysai, and R. D. Gudi, “Identification of complex electrical engineering) from Nanjing Aeronautical In-
nonlinear processes based on fuzzy decomposition of the steady state stitute, Nanjing, China, in 1982 and 1984, respec-
space,” J. Process Control, vol. 13, pp. 473–488, 2003. tively, and the Ph.D. degree in electrical engineering
[33] T. J. Zhang, G. Feng, and J. H. Lu, “Fuzzy constrained min-max model from the University of Melbourne, Melbourne, Vic.,
predictive control using piecewise Lyapunov functions,” IEEE Trans. Australia, in 1992.
Fuzzy Syst., vol. 15, no. 4, pp. 686–698, Aug. 2007. Since 2000, he has been with the the City Univer-
[34] T. J. Zhang, G. Feng, J. H. Lu, and W. G. Xiang, “Robust constrained sity of Hong Kong, Kowloon, Hong Kong, where he
fuzzy affine model predictive control with application to a fluidized bed is currently a Professor. From 1992 to 1999, he was a
combustion plant,” IEEE Trans. Control Syst. Technol., vol. 16, no. 5, Lecturer/Senior Lecturer in the School of Electrical
pp. 1047–1056, Sep. 2008. Engineering, University of New South Wales, Kensington, N.S.W., Australia.
[35] T. J. Zhang, G. Feng, and X.-J. Zeng, “Output tracking of constrained His current research interests include piecewise-linear systems, robot networks,
nonlinear processes with offset-free input-to-state stable fuzzy predictive and intelligent systems and control.
control,” Automatica, vol. 45, no. 4, Apr. 2009. Prof. Feng is an Associate Editor of the IEEE TRANSACTIONS ON AUTOMATIC
[36] F. Jurado, “Predictive control of solid oxide fuel cells using fuzzy Ham- CONTROL and the IEEE TRANSACTIONS ON FUZZY SYSTEMS, and was an Asso-
merstein models,” J. Power Sources, vol. 158, pp. 245–253, 2006. ciate Editor of the IEEE TRANSACTIONS ON SYSTEMS, MAN AND CYBERNETICS,
[37] R. S. Gemmen and J. Trembly, “On the mechanisms and behavior of coal PART C and the Journal of Control Theory and Applications, and on the Confer-
syngas transport and reaction within the anode of a solid oxide fuel cell,” ence Editorial Board of the IEEE Control System Society. He was the recipient
J. Power Sources, vol. 161, pp. 1084–1095, 2006. of the Alexander von Humboldt Fellowship Award in 1997–1998, and the 2005
[38] H. K. Khalil, Nonlinear Systems, 3rd ed. Upper Saddle River, NJ: IEEE TRANSACTIONS ON FUZZY SYSTEMS Outstanding Paper Award in 2007.
Prentice-Hall, 2002.
Diagnosability of Fuzzy Discrete-Event Systems:

A Fuzzy Approach
Fuchun Liu and Daowen Qiu
Abstract—In order to more effectively cope with the real-world applications such as biomedicine, in which vagueness, impre-
problems of vagueness, fuzzy discrete-event systems (FDESs) were ciseness, and subjectivity are typical features. For example, it is
proposed by Lin and Ying recently. Then we and Cao and Ying in- vague when a man’s condition of the body is said to be “good.”
vestigated the supervisory control of FDESs independently. In this
paper, we are concerned with another important issue of FDESs, Moreover, it is imprecise to say at what point exactly a man
the failure diagnosis. More specifically: 1) we propose a “fuzzy has changed from “good” state to “poor” state. In order to more
diagnosability” approach by introducing a fuzzy diagnosability effectively cope with those problems, Lin and Ying [11], [12]
function to characterize the diagnosability degree, which takes recently initiated significantly the study of fuzzy discrete-event
values in the interval [0, 1] rather than {0, 1}; 2) based on the systems (FDESs) by combining fuzzy set theory [31] with crisp
observability of events, we formalize the construction of the diag-
nosers that are used to perform fuzzy diagnosis; 3) a number of DESs [4]. Notably, FDESs have been applied to biomedical
basic properties of the diagnosers are investigated. In particular, we control for HIV/AIDS treatment [13], [29], [30] and decision
present a necessary and sufficient condition for failure diagnosis of making [14]. More recently, R. Huq et al. have proposed an in-
FDESs. Our results generalize the important consequences of the telligent sensory information processing technique using FDESs
diagnosability for crisp discrete-event systems (DESs) introduced for robotic control [8], [9].
by Sampath et al. The newly proposed approach allows us to deal
with the problem of diagnosability for both crisp DESs and FDESs; Just as Lin and Ying [12] pointed out, a comprehensive the-
4) in addition, a method for checking the fuzzy diagnosability for ory of FDESs still needs to be set up, including many important
FDESs is proposed. Also, some examples are provided to illustrate concepts, methods, and theorems, such as controllability, ob-
the application of the diagnosability of FDESs. servability, and optimal control. These issues have been partially
Index Terms—Discrete-event systems (DESs), failure detection, investigated. The supervisory control of FDESs was proposed
fuzzy diagnosability, fuzzy discrete-event systems (FDESs), fuzzy by Qiu [20], the authors [16], [21], and Cao and Ying [1]–[3],
finite automata. respectively. It is worth noting that there are great differences
between the above two frameworks. In the works of [1]–[3], the
I. INTRODUCTION observability of events is crisp, and the controllable and uncon-
ISCRETE-EVENT systems (DESs) are dynamical sys- trollable event sets are also crisp, although the controllability of
D tems whose state space is discrete and states can only
change when certain events occur, which have been successfully
events is not completely crisp. However, in [16] and [21], we
have established the centralized and decentralized supervisory
applied to provide a formal treatment of many technological and control of FDESs under partial observations, in which both the
engineering systems [4]. In order to guarantee performance to observability and controllability of events are fuzzy instead. As
a reliable system, the control engineers should design a system a continuation of our works, this paper is to investigate another
that runs safely within its normal boundaries. Therefore, failure important issue of FDESs, the failure diagnosis.
diagnosis in DESs, which is to detect and isolate the unobserv- In this paper, we propose a “fuzzy diagnosability” approach
able fault events occurring in a system within a finite delay, is for failure diagnosis of FDESs. Although we have respectively
of practical and theoretical importance, and has received con- dealt with the uncertain problems in decentralized diagnosis [15]
siderable attention in recent years [5], [10], [15], [17]–[19], and safe diagnosis of stochastic DESs [17] by means of the
[22]–[28]. framework in [27], and Kilic has also considered the uncer-
In most engineering applications, the states and events of tain problems in [10], fuzzy diagnosability has not received
DESs are crisp. However, this is not the case in many other enough attention since most of the research of diagnosability
in the literature has thoroughly focused on the centralized di-
Manuscript received March 29, 2008; revised October 6, 2008; accepted agnosis [24], [25], decentralized diagnosis [5], [22], distributed
January 12, 2009. First published January 27, 2009; current version published diagnosis [26], and its applications in telecommunication net-
April 1, 2009. This work was supported in part by the National Natural Science works [18], [23] and digital circuits [28], and so on.
Foundation under Grant 60573006 and Grant 60873055, in part by the Research
Foundation for the Doctoral Program of Higher School of Ministry of Education The framework in this paper is different from those in the
under Grant 20050558015, and in part by the New Century Excellent Talents literature. As a continuation of our works [16], [21], in this paper
(NCET) of China. we still consider the same model as that in [20], while the FDES
F. C. Liu is with the Faculty of Computer, Guangdong University of Tech-
nology, Guangzhou 510006, China, and also with the Department of Com- in [10] is modeled as a fuzzy expert system that contains fuzzy
puter Science, Zhongshan University, Guangzhou 510275, China (e-mail: IF–THEN rules. Furthermore, it is well known that the diagnoser
liufch@gdut.edu.cn). approach introduced by Sampath et al. [24] has been extensively
D. W. Qiu is with the Department of Computer Science, Zhongshan Univer-
sity, Guangzhou 510275, China (e-mail: issqdw@mail.sysu.edu.cn). adopted in the literature for its efficiency and adaptability [23].
Digital Object Identifier 10.1109/TFUZZ.2009.2013840 The fuzzy diagnosability proposed in this paper is exactly based
1063-6706/$25.00 © 2009 IEEE
LIU AND QIU: DIAGNOSABILITY OF FUZZY DISCRETE-EVENT SYSTEMS: A FUZZY APPROACH 373
on the diagnoser of FDESs, which also differs from that in [10]. language generated by G, denoted as L(G) or L. L/s stands
This paper is also different from [15], [17], and [27], since [15], for the set of all possible continuations of string s in L, i.e.,
[17], and [27] deal with the diagnosability of stochastic DESs, L/s = {t ∈ E ∗ : st ∈ L}.
which are modeled by stochastic automata with a probabilistic The event set E is partitioned into two disjoint subsets E =
structure appending to each transition of states, while this paper Eo ∪ Euo , where Eo is the set of observable events and Euo is
is to formalize the diagnosis of FDESs modeled by max–min the set of unobservable events. When a string of events occurs
systems with fuzzy states and fuzzy events [20], [21]. in a system, the sequence of observable events is filtered by the
The present paper attempts to address the diagnosability of projection P : E ∗ → Eo∗ , which is defined as: P () = , and
FDESs. We formalize a notion of the fuzzy diagnosability func- P (sσ) = P (s)P (σ) for σ ∈ E, s ∈ E ∗ , where
tion to characterize the diagnosability degree, which takes values
σ, if σ ∈ Eo ,
in the interval [0, 1] rather than {0, 1}, and a “fuzzy diagnos- P (σ) = (1)
ability” approach is proposed. The fuzzy diagnosability function , if σ ∈ Euo .
indicates that the system is diagnosable with a certain degree. The inverse projection of string t ∈ Eo∗ is given by
In particular, if the fuzzy diagnosability function equals one,
P −1 (t) = {s ∈ L : P (s) = t} . (2)
then the fuzzy system is completely diagnosable; if the fuzzy
diagnosability function equals zero, then the fuzzy system is Let Ef denote the set of failure events that are to be di-
completely nondiagnosable. Based on the observability degree agnosed. Usually, Ef is partitioned into a set of failure types
of events, a method for constructing the diagnosers used to per- f1 , f2 , . . . , fm , i.e.,
form fuzzy diagnosis is developed. After investigating a number
Ef = Ef 1 ∪ Ef 2 ∪ · · · ∪ Ef m . (3)
of basic properties of the diagnosers, we present a necessary and
sufficient condition for diagnosability of FDESs, which gener- Denote Ψ(Ef ) as the set of all traces that end with a failure
alizes the important consequences of failure diagnosis for crisp belonging to Ef , i.e.,
DESs introduced by Sampath et al. [24]. The newly proposed
Ψ(Ef ) = {s ∈ L : sl ∈ Ef } (4)
diagnosability approach allows us to deal with the problem of
failure diagnosis for both crisp DESs and FDESs. Furthermore, where sl denotes the last event of s.
a computing method to check the fuzzy diagnosability of FDESs Definition 1 [24]: Let L be a language generated by G =
is given. Also, some examples are provided to illustrate the ap- (Q, E, δ, q0 ) and let P : E ∗ → Eo∗ be a projection. L is said to
plication of the diagnosability of FDESs. be f -diagnosable if the following holds:
The rest of the paper is organized as follows. Section II recalls
(∃n0 ∈ N)(∀s ∈ Ψ(Ef ))(∀t ∈ L/s)(|t| ≥ n0 ⇒ D(st) = 1)
some preliminaries concerning diagnosability for crisp DESs
(5)
and the concept of fuzzy finite automata. In Section III, we for-
where the diagnosability function D : E ∗ → {0, 1} is defined
malize the notion of fuzzy diagnosability for FDESs. Then, in
as follows:
Section IV, a method for constructing the diagnosers used to per-
form fuzzy diagnosis is given. After that, some main properties 1, if ω ∈ P −1 (P (st)) ⇒ Ef ∈ ω
D(st) = (6)
of the diagnosers are investigated in Section V. In particular, we 0, otherwise
present a necessary and sufficient condition for diagnosability in which Ef ∈ ω represents that string ω contains a failure event
of FDESs. Moreover, a computing process to check the diagnos- from Ef .
ability condition is further provided. In Section VI, a number of Intuitively, L being f -diagnosable means that, for any trace
examples are provided to illustrate the potential applications to s that ends in a failure event and for its any sufficiently long
diagnosability for FDESs. To conclude, in Section VII, we sum- continuation t, every trace with the same observable record as
marize the main results of the paper and address some related st should contain a failure event too (i.e., D(st) = 1).
issues.
B. Fuzzy Discrete-Event Systems
II. PRELIMINARIES
If the crisp state set Q = {q0 , q1 , . . . , qn −1 }, then in the set-
In this section, we briefly recall some preliminaries concern- ting of FDESs, each fuzzy state q is a vector over the crisp state
ing diagnosability for crisp DESs proposed in [24] and the no- set, denoted as [a0 a1 · · · an −1 ], where ai ∈ [0, 1] represents
tions of fuzzy finite automata [20], [21]. the possibility of the current state being qi . Similarly, a fuzzy
event σ is a matrix [aij ]n ×n , in which aij ∈ [0, 1] indicates the
A. Diagnosability for Crisp DESs possibility of the system transforming from state qi to state qj
A crisp DES is usually described by a deterministic automaton when σ occurs.
G = (Q, E, δ, q0 ), where Q is the finite set of states with the Definition 2 [20]: A fuzzy finite automaton is a max–min
initial state q0 , E is the finite set of events, δ : Q × E → Q is system
the transition function. Each sequence over E is called a string. = (Q,
G E, q0 )
δ, (7)
E ∗ denotes the set of all finite strings over E. For s ∈ E ∗ , |s| is
the length of s. Especially, s is an empty string in case |s| = 0, where Q is a set of fuzzy states with the initial fuzzy state
which is denoted by . The behavior of G is described by the
q0 ; E is the set of fuzzy events; the state transition relation
δ : Q
×E →Q q, σ
is defined as δ( ) = q σ
. Note that is q ) be the set of all traces that originate from state q.
Let L(
max–min operation: for matrix A = [aij ]n ×m and matrix B = Denote
[bij ]m ×k , define A B = [cij ]n ×k , where L1 (
q, σ m ax−o ∪ {
) = (E :Σ
a∈E o ( o (
a) > Σ σ )}) ∩ L(
q)
m
cij = max min{ail , blj }. (8) (15)
l=1
L2 ( ) = {
q, σ u
a ∈ L(
q) :
a ∈ L1 ( o (
) ∧ Σ
q, σ o (
σ) ≥ M u)}
Remark 1: As a continuation of our works [16], [20], [21],
we still would like to consider max–min operation instead of (16)
max-product operation [12] in fuzzy finite automata. Moreover,
o ( o ( ∈u
}; and
δ can be regarded as a partial function in practice. where M u) = max{Σ σ) : σ
Example 1: Let G = (Q,
E, q0 ) be an FDES. If fuzzy state
δ,
L( ) = L1 (
q, σ ) ∪ L2 (
q, σ ).
q, σ
and fuzzy event
q = [0.7, 0.3] ∈ Q
Intuitively, L1 ( ) collects all of the single fuzzy events in
q, σ
0.4 0.5 L(q ) whose observability degree is the largest or larger than
=
σ ∈E
0.8 0.1 o (
Σ σ ); and L2 ( ) consists of the strings whose observability
q, σ
of each event is not larger than Σ o (
σ ) except for the last event.
occurs in q, then
Definition 3 [21]: For σ
∈ E, the σ-projectionP :E∗ → E ∗
σ
q, σ 0.4 0.5
δ( ) = q σ
= [0.7, 0.3] = [0.4, 0.5]. is defined as: P () = , and P s
(a) = P (
s)P a) for
( a∈E
0.8 0.1 σ σ σ σ
∗
and s ∈ E , where
III. DEFINITION OF DIAGNOSABILITY FOR FDESS m ax−o or Σ o ( o (

a, if a∈E a) > Σ σ)
Pσ
(
a ) = (17)
In this section, we formalize the notion of diagnosability for , otherwise.
FDESs, in which a fuzzy diagnosability function is introduced
The inverse projection operator is given by
to characterize the degree of diagnosability.
Let G = (Q, E, q0 ) be an FDES. In the framework of our
δ, P −1 (t) = {s∈E ∗ : (∃
q ∈ Q) q0 , s) = q ∧ P (
δ( s) = t}.

σ
σ
works [16], [20], [21], each fuzzy event simultaneously belongs (18)
to the observable event set, the unobservable event set, and the Remark 2: σ -projection is due to erasing the fuzzy events of
failure event set; only the degrees of belonging to those sets may strings whose observability is not larger than Σ o (
σ ). Especially,
be different. So we introduce three fuzzy subsets: the unobserv- if each event is either completely observable or completely un-
able event fuzzy subset Σ → [0, 1], the observable event
uo : E
observable, then all of the σ -projections are just the projection
fuzzy subset Σ o : E
→ [0, 1], and the failure event fuzzy subset P : E ∗ → Eo∗ in crisp DESs, which simply erases the unobserv-

Σf : E → [0, 1]. Intuitively, Σ f (
σ ) describes the possibility of able events.
failure occurring on σ
∈ E, and Σ o ( uo (
σ ) and Σ σ ) show the For the sake of making a correct decision of detection, it
observability degree of σ , which satisfies would be instructive to specify an upper bound Σ f (
σ ) for each

∈ E, if the possibility of the failure occurring on a string s
σ
uo (
Σ o (
σ) + Σ σ ) = 1. (9) f ( f (
exceeds the upper bound (i.e., Σ s) ≥ Σ σ )), then we think
Furthermore, we define that s is a failure string since the system is in a failed state. In the
following, we formalize an approach of “fuzzy diagnosability”
o (
Σ 2 · · · σ
σ1 σ o (
m ) = min{Σ σi ) : i = 1, 2, . . . , m} (10) to detect the occurrence of those failure strings that exceed the
f (
Σ 2 · · · σ
σ1 σ f (
m ) = max{Σ σi ) : i = 1, 2, . . . , m} (11) specified upper bound.
Denote E fail = { σ∈E :Σ f (
σ ) > 0}. For σ ∈E fail , let
where each σ
i ∈ E. Ψ f ) be the set of all traces that end with an event whose
(Σ
σ
In order to avoid the case that the event set of the diagnoser possibility of failure occurring is not less than Σ f (
σ ), i.e.
constructed later is null, we introduce the maximal observable
event set Em ax−o , which is composed of the events with the Ψ f ) = {
(Σ s∈E ∗ : (∃ q0 , s)
q ∈ Q)δ(
σ
largest observability degree, i.e.
f (
= q ∧ Σ f (
sl ) ≥ Σ σ )}. (19)
m ax−o = {
E : (∀
σ∈E Σ
σ
∈ E) o ( o (
σ) ≥ Σ σ
)}. (12)
Definition 4: For σ∈E fail , the fuzzy diagnosability function
denoted as L or L, is defined
The language generated by G, is defined as a partial function

G
as follows:
FD ∗ → [0, 1]
:E (20)
L = { ∗ : (∃
s∈E q ∈ Q) q0 , s) = q}.
δ( (13)
σ
where for s ∈ Ψ f ) and

(Σ t ∈ L/
s,
σ
The postlanguage of L after s is the set of continuations of s in
i.e.,
G, f (
min{Σ f (
σ ), Σ ∈ P −1 (P
ω) : ω s
(t))}
FD s
(t) =
σ σ
. (21)
s = {
L/ ∗ : (∃
t∈E q ∈ Q) q0 , s
δ( t) = q}. (14) σ f (
Σ σ)
Remark 3: We compare the fuzzy diagnosability function for Therefore, the fuzzy diagnosability for FDESs is a generaliza-
FDESs with the diagnosability function for crisp DESs [24] as tion of the diagnosability for crisp DESs introduced by Sampath,
follows. If the observable event set and the failure event set et al. [24].
o (
are crisp instead of fuzzy, i.e., Σ f (
a), Σ a) ∈ {0, 1} for each Case 2: If the diagnosability degree of G equals to λ with

a ∈ E, then for σ
∈ Efail (i.e., Σf ( σ ) = 1), by Remark 2, we respect to σ , then there are s and continuation t such that
have FD (
st ) = λ is minimal, that is, there is ω ∈ P −1
(P s
(t))
σ σ σ
f ( f (
ω ) = λΣf (
ω ) ≥ λΣf (
min{Σ σ ), Σ ∈ P −1 (P
ω) : ω s
(t))} satisfying Σf ( σ ), and Σf ( σ ) for other
FD s
(t) =
σ σ

∈ P −1 (P
ω (
s
t )).
σ f (
Σ σ)
σ σ
Case 3: In particular, if λ = 0, then there are s and t such that
f (
= min{1, Σ ω) : ω s
∈ P −1 (P (t))} FD (
st ) = 0, which means that P −1
(P (
s
t )) contains a trace
σ
σ
σ
f (
that Σf (
s
∈ P −1 (P (
1, if ω t)) ⇒ Σ ω) = 1 ω ω ) = 0, that is, the failure does not possibly occur on
= although ω
ω has the same observed record as s t and the failure
0, otherwise.
occurs on st.
Thus, in this case, the fuzzy diagnosability function FD : Definition 6: If G is completely f -diagnosable with respect to
σ
∗ → [0, 1] degenerates into the diagnosability function D :
E each σ
∈ Efail , then G is said to be completely f -diagnosable.
E ∗ → {0, 1} for crisp DESs described in Definition 1. If there exists σ
∈ Efail such that G is f -diagnosable with re-
spect to σ with degree λ, and the diagnosability degree of G
Definition 5: Let L be a language generated by FDES G.
with respect to other σ

∈ Efail is not less than λ, then G is

1) G is said to be completely f -diagnosable with respect to σ
equals one with respect said to be f -diagnosable with degree λ. If G is completely f -
(or the diagnosability degree of G
undiagnosable with respect to a certain σ
∈ Efail , then G is said
), if
to σ
to be completely f -undiagnosable.
(∃n0 ∈ N)(∀ s ∈ Ψ (Σ f ))(∀
t ∈ L/
s)
σ (22)

(|t| ≥ n0 ⇒ FD (
st) = 1).
σ IV. CONSTRUCTION OF DIAGNOSERS FROM AN FDES
2) G is said to be f -diagnosable with degree λ with respect to In this section, we present a method of constructing the di-
σ equals λ with respect agnosers from a given FDES with a single failure type. In Sec-
) where 0 ≤ λ < 1, if for any n0 ∈ N, we have
to σ tion V-D, we will extend it to the case of multiple failure types.
f ), Let G = (Q,
E, q0 ) be an FDES. We use the possible fail-
δ,
min{FD s
( t) : s ∈ Ψ (Σ t ∈ L/s, |
t| ≥ n0 } = λ.
σ σ
ure labels to label the states of diagnosers. For each σ∈E fail ,
(23)
denote the set of possible failure labels as
3) In particular, if the above λ = 0, i.e., there are s ∈ Ψ f )
(Σ
σ
and
t ∈ L/ s such that f (
= {F } ∪ {N µ : µ = Σ a),
a ∈ E} (26)
FD
σ
s
(t) = 0 (24) where F stands for the possibility of which the failure occurring
exceeds the specified degree Σ f (
σ ) and the system is in “failed”
then G is said to be completely f -undiagnosable with
equals zero state, while N µ means that the possibility of the failure occurring
respect to σ is µ but it does not exceed the specified degree, so the system is
with respect to σ). still in “normal” state.
Remark 4: Definition 5 shows that the fuzzy diagnosability For σ we define a subset of Q
∈ E, as
function FD ∗ → [0, 1] indicates the diagnosability degree
:E
σ
which can be interpreted as follows:
of G, = {
Q q0 } ∪ { : (∃
q∈Q a∈E
q
∈ Q)(∃ q
,
)δ( a) = q}

σ
σ
Case 1: If the diagnosability degree of G is equal to one with (27)
respect to σ
, then for any s that Σf (sl ) ≥ Σ f (
σ ) and its any where

continuation t, we have FD (
st) = 1, which makes clear that
σ =E
E m ax−o ∪ { :Σ
a∈E o ( o (
a) > Σ σ )}. (28)
the following inequality holds: σ
Σ f (
σ ) ≤ min{Σ f (
ω) : ω ∈ P −1 (P (s
t ))}. (25) consists of the initial state q0 and the states reach-

σ σ Intuitively, Q
σ
able along an event whose observability degree is the largest or
That is to say, if Σ f (
sl ) ≥ Σ f (
σ ), then for every trace ω that o (σ).
larger than Σ
produces the same observed record as s t, the possibility of the be an FDES and σ fail . The diagnoser
f ( Definition 7: Let G ∈E
failure occurring on ω is not less than Σ σ ), too. Obviously,
with respect to σ is constructed as a finite automaton
o (
if Σ f (
a), Σ a) ∈ {0, 1} for each then (25) degenerates
a ∈ E,
into G d , δd , 0 )
d , E
d = (Q (29)
f (
Σ f (
σ) = 1 ⇒ Σ ω) = 1
where Q d is the state set of the diagnoser, E d is the set of
−1
∈ P (P (
for all ω s
t)), which is exactly consistent with the
events, δd is the transition function, and 0 is the initial state.
concept of the diagnosability for crisp DESs (see Definition 1). More specifically
1) The state space Q d ⊆ 2Qσ × is composed of the states V. NECESSARY AND SUFFICIENT CONDITION OF
DIAGNOSABILITY FOR FDESS
reachable from 0 under δd . A state of Q
d is of the form
In this section, we give some properties of the diagnosers.
In particular, a necessary and sufficient condition of the fuzzy
= {(
q1 , 1 ), (
q2 , 2 ), . . . , (
qn , n )} (30)
diagnosability for FDESs is presented. Also, we propose an
approach to check the fuzzy diagnosability for FDESs.
in which each qi ∈ Q and i ∈ .

σ
2) The initial state 0 = {( is
q0 , N 0 )}, which means that G A. Properties of Diagnosers
normal to start with. d = (Q d , E d , δd , 0 ) be the diagnoser
, i.e., Property 1: Let G
3) The event set of the diagnoser is E
σ
with respect to σ ∈E fail and ∈ Q d . (q1 , 1 ), (q2 , 2 ) ∈ if

and only if there exist s1 , s2 ∈ E such that (∗
s1 )l = ( s2 )l ∈
Ed = E =E m ax−o ∪ { a∈E :Σ o (
a) > Σ o (σ )}. d , P ( d ( q0 , sk ) =

σ E
σ
s 1 ) = P
σ
(
s 2 ), δ 0 , P
σ
(s 1 )) =
, and δ(
(31) q0 , N 0 , sk ) = k for k = 1, 2.
qk , LP(
4) δd : Qd × E d → Q d is the transition function of the diag-
d , there are
Proof: Necessity: From ∈ Q a1 , a2 , . . . ,
aj ∈
noser, which will be defined in Definition 9.
××
Ed and 1 , 2 , . . . , j −1 ∈ Qd , such that δd ( i ,
ai+1 ) = i+1 ,
Definition 8: The label propagation function LP : Q
→ is defined as follows. Let q ∈ Q
∗ , ∈ , and s ∈ E
σ
∗ . where 0 ≤ i ≤ j − 1 and j = . Due to ( q1 , 1 ), (q2 , 2 ) ∈ ,
E
σ by Definition 9, there exist ( q1k , k1 ) ∈ j −1 , and tk1 ∈ La j ( )
q1k , σ
1) If = F , then LP( q , , s) = F .
f ( (k = 1, 2) such that for k = 1, 2
µ
2) If = N and Σf ( s) ≥ Σ σ ), then LP( q , , s) = F .
3) If = N , Σf (µ
s) < Σf ( σ ) and Σ f (s) > µ, then qk ,
qk = δ( k
q1k , k1 ,
tk1 ).
1 t1 ) k = LP(
LP( q , , s) = N
Σ f (
s)
.
f ( f ( f ( Similarly, note that δd ( j −2 ,
aj −1 ) = j −1 , and hence, there are
4) If = N µ , Σ s) < Σ σ ) and Σ s) ≤ µ, then q2k , k2 ) ∈ j −2 , and
( tk2 ∈ La j −1 ( ) (k = 1, 2) satisfying for
q2k , σ
LP( q , , s) = N . µ
k = 1, 2,
The above label propagation function is used to describe the
changes of label from one state of diagnoser to another, which qk ,
q1k = δ( k
q2k , k2 ,
k1 = LP( tk2 ).
2 t2 )
is interpreted as follows. Denote δ( q0 , s0 ) = q and δ( q , s) = q
.
1) If label F is added to state q (i.e., q is a “failed” state ......
for Σ f (s0 ) ≥ Σ f (σ )), then the successor state q
is also a With the analogous process, there are ( qjk−1 , kj−1 ) ∈ 1 and
“failed” state since Σ f ( s0 s) ≥ Σ f (σ ).
tkj ∈ La 1 ( ) (k = 1, 2) such that for k = 1, 2,
q0 , σ
2) If state q is a “normal” state, the possibility of the failure
occurring on s0 is µ that does not exceed the specified q0 ,
qjk−1 = δ( tkj ) q0 , N 0 ,
kj−1 = LP( tkj ).
degree Σ f (σ ), but if the possibility of the failure occurring
on s exceeds Σ f (σ ), then q
should be labeled F as a We take
f ( s0 s) ≥ Σ f (
“failed” state since Σ σ ). t1j
s1 = t12
t1j −1 · · · t11 t2j
s2 = t22
t2j −1 · · · t21 . (34)
3) If state q is a “normal” state labeled by N µ , where µ ≤
f (
Σ s) < Σ f (
σ ), then q
is also to be a “normal” state for Obviously, δd ( 0 , P (s )) = , ( s1 )l = ( s2 )l = aj ∈ E d , and
σ 1
f (
Σ s0 s) = µ, which does not exceed the specified degree, q0 , sk ) = qk , LP(
for k = 1, 2, we have δ( q0 , N 0 , sk ) = k .
but the label of q
should be changed to N Σf ( s)
. Moreover P σ 1
(
s ) =
a
a
1 2 · · ·
aj = P (
σ 2
s ).
4) If state q is a “normal” state labeled by N , and Σ µ f (
s) is Sufficiency: Assume that there exist s1 , s2 ∈ E ∗ such
smaller than both µ and Σf (
σ ), then q inherits the label that ( s1 )l = (

s2 )l ∈ Ed , P (
s ) = P (
s ), δd ( 0 , P (s )) =
σ 1 σ 2 σ 1
of q for Σ f (
s0 s) = Σ f ( s0 ) < Σ f (σ ).
, and δ( q0 , sk ) = qk , LP( q0 , N , sk ) = k for k = 1, 2. From
0
Definition 9: The transition function of the diagnoser G d is δd (

0 , P (
s )) = , we denote P s )=
( a1
a2 · · ·
aj , and then
σ 1 σ 1

a partial function δd : Qd × Ed → Qd , which is defined as: for
we can obtain a state sequence 1 , 2 , · · · , j −1 ∈ Qd satisfying
∈ Q d and a∈E d ,
δd (
i ,
ai+1 ) = i+1 , (0 ≤ i ≤ j − 1)

δd (
,
a) = qi , s), LP(
(δ( qi , i , s)) (32) where j = . Furthermore, from q0 , sk ) = qk ,
δ(
(
q i , i )∈ s ∈L (
q i ,
σ) and LP( q0 , N 0 , sk ) = k , (k = 1, 2), we have that
a
q2 , 2 ) ∈ by Definition 9.
q1 , 1 ), (
(
where La ( ) collects all of the strings in L(
qi , σ ) that end
qi , σ
Definition 10: Let Gd be the diagnoser with respect to σ ∈
with
a, i.e., Efail . A state ∈ Q d is said to be F -certain if = F for all
q , ) ∈ , or = F for all (
( q , ) ∈ ; otherwise, if there are
q2 , 2 ) ∈ such that 1 = F and 2 = F , then is
q1 , 1 ), (
(
La ( ) = {
qi , σ s ∈ L( ) : sl =
qi , σ a} . (33)
said to be F -uncertain.
Property 2: Let G d = (Q d , δd , 0 ) be the diagnoser with

d , E j and yjr form cycles in
b) the sequences of states x
∈E
respect to σ
fail and δd ( ) = .
0 , u G with
1) is F -uncertain if and only if there exist s1 , s2 ∈
∗ such that ( s2 )l ∈ E d , P ( xh , sh σ
j j ) = x
hj+1 ,
E s1 )l = ( s ) = P
σ 1
(
s ),
σ 2 δ(j

δd (
0 , P s )) = , and
(
σ 1 (j = 1, . . . , k − 1; h = 1, . . . , m)
xh , sh σ
f (
Σ f (
s1 ) ≥ Σ f (
σ) > Σ s2 ). (35) δ(k k k ) = x
h+1
1 ,
xm , sm σ
(h = 1, . . . , m − 1), δ( k k ) = x
11
2) is F -certain if and only if Σ f (
s) ≥ Σ f (
σ ) for all s ∈
k
−1
P ( u), or Σ f (
s) < Σ f (
σ ) for all s ∈ P −1 ( u), where

σ
σ and
sl ∈ Ed .
Proof: 1) Sufficiency: Suppose there exists s2 ∈ P −1 (P (
s )) yr ,
δ( r
j ) = yjr+1 ,
σ σ 1 j tj σ
such that Σ f (s1 ) ≥ Σ f (
σ) > Σ f (
s2 ), where ( s1 )l = (s2 )l ∈ (j = 1, . . . , k − 1; r = 1, . . . , n)

Ed and δd (0 , P s )) = . Denote
(
σ 1 yr ,
δ( r
k ) = y1r +1 ,
k tk σ
q0 , N 0 , s1 ) = 1
LP( q0 , N 0 , s2 ) = 2 .
LP( yn ,
(r = 1, . . . , n − 1), δ( n
k ) = y11
k tk σ
From Definition 8, we have 1 = F but 2 = F . By Property 1, where shj σ

j ∈ L( ),
xhj , σ j ∈ L(
trj σ ).
yjr , σ
we obtain ( q1 , 1 ), ( q2 , 2 ) ∈ , where δ( q0 , s1 ) = q1 and d is a cycle com-
Intuitively, a µ-F -indeterminate cycle in G

δ(q0 , s2 ) = q2 . That is, is F -uncertain. posed of F -uncertain states, h and

corresponding

to this cycle,
Necessity: It is clearly obtained from Definitions 9 and 10. there exist two sequences x j and yjr forming cycles of G,
2) It holds obviously since it is the inverse and negative propo- in which one carries the failure label F and the other carries the
h
sition of (1). failure label N µ . We call the above two state sequences x j
Property 3: Let G d = (Q d , E d , δd , 0 ) be the diagnoser with r
and yj to be “failed-state” cycle and “normal-state” cycle of

respect to σ ∈E fail . If δd (
1 , a) = 2 , where a∈E d , 1 and

respectively. All of the µ-F -indeterminate cycles are called

G,
2 are F -uncertain, and ( q1 , F ), ( q1 , N ) ∈ 1 , then there ex-
µ
by a joint name of F -indeterminate cycle.
),

ist s
a ∈ L( q1 , σ t
a ∈ L( q1
, σ
), and ( q2 , F ), (q2
, N µ ) ∈ 2
Theorem 1: Let L be a language generated by an FDES
satisfying E, q0 ) and σ
δ, fail . If the diagnoser G d with re-
G = (Q, ∈E
q1 , s q
, spect to σ contains a µ-F -indeterminate cycle, then the diag-
P s
(a) = P (
t
a) =
a δ( a) = q2 δ(1 t a) = q2
. with respect to σ
σ σ nosability degree of G , denoted as λ, satisfies
the following inequality:
Moreover, if Σ f (a) ≤ µ, then µ
= µ; if Σ
t f (a) > µ, then µ
=
t

Σf (t
a). µ
λ≤ . (36)
Proof: It is clearly obtained from Definitions 8 and 9.
Σf (
σ)
Property 4: Let G d be the diagnoser with respect to σ ∈E fail .
d , then all states in the cycle
If the set of states forms a cycle in G
Proof: Assume that the diagnosability degree of G with
have the same failure label. d contains a µ-F -indeterminate cy-
respect to σ is λ, and G
Proof: By Definition 8, it is easy to prove since any two states
d are reachable from each other. cle { 1 , 2 , . . . , k }. By Definition 11, there are two se-
in a cycle of G
quences of states { xhj } and { yjr }, which form two cycles in
and the corresponding strings {
G, j } and {
shj σ j } satisfy
trj σ
B. Necessary and Sufficient Condition of Diagnosability condition (b) of Definition 11, where ( h h
xj , j ), (yjr , drj ) ∈ j ,
for FDESs and hj = F but drj = N µ for all j = 1, . . . , k; h = 1, . . . , m;
Definition 11: Let G d = (Q d , E d , δd , 0 ) be the diag- r = 1, . . . , n.

noser of G with respect to σ ∈E fail . A set { 1 ,
1 , σ Since ( x11 , 11 ), (y11 , d11 ) ∈ 1 , by Property 1, there exist
2 , . . . , k , σ
2 , σ k , 1 } is said to form a µ-F -indeterminate ∗
s0 , t0 ∈ E such that P (
s ) = P ( q0 , s0 ) = x
t ), δ( 11 , and
σ 0 σ 0
cycle in G d , if q0 ,
δ( t0 ) = y11 . Notice that 11 = F , and drj = N µ for all
1) 1 , 2 , . . . , k ∈ Q d are F -uncertain states, and the set f ( f ( f ( f (
j, r. Therefore, Σ s0 ) ≥ Σ σ ), µ = Σ t0 ) < Σ σ ), and
{ 1 , σ1 , 2 , σ
2 , . . . , k , σ
k , 1 } forms a cycle in G d , i.e.,

Σf (tj σr
j ) ≤ µ for all j, r, i.e.

δd ( j , σ j ) = j +1 for j = 1, . . . , k − 1, and δd ( k ) =
k , σ
1 .
f (
Σ f (
s0 ) ≥ Σ f (
σ) > Σ f (
t0 ) ≥ Σ j )
trj σ (37)
2) For each above j , there exist ( yjr , drj ) ∈ j (j =
xhj , hj ), (
1, . . . , k; h = 1, . . . , m; r = 1, . . . , n) satisfying
a) hj = F and drj = N µ for all j, h, r; and f (
for all j, r, where Σ t0 ) = µ.
Let l be a sufficiently large integer. We consider the following cycle are F -uncertain because they have the same failure label
two traces: by Property 4. We denote the cycle of the F -uncertain states as
1 , 2 , . . . , k , where
1 = s0 (
ω 1 . . . s1k σ
s11 σ k s21 σ
1 . . . s2k σ
k . . . sm 1 . . . sm
1 σ k )ln
k σ
(38) δd ( j ) = j +1 ,

j , σ (j = 1, . . . , k − 1) δd ( k ) = 1
k , σ
(46)
2 =
ω t0 ( 1 . . .
t11 σ k
t1k σ 1 . . .
t21 σ k . . .
t2k σ 1 . . .
tn1 σ k )lm .
tnk σ j ∈ E
and σ d for each j = 1, . . . , k.
(39) Since δd ( 1 ) = 2 , and (
1 , σ q1 , F ), (q1
, N Σf (
ω)
) ∈ 1 , by
Property 3, there are ( q2 , F ), (

Σ (
q2 , N f ) ∈ 2 and s1 σ
ω )
1 ∈
Then
) and
q1 , σ
L( 1 ∈ L(
t1 σ q1
, σ
), such that
P (
ω1 ) = Pσ ( ω2 ) = P (
σ 0
t )( 2 . . . σ
σ1 σ k )lm n . (40)
σ q1 , s1 σ
δ( 1 ) = q2 q
,
δ( 1 ) = q2
.
1 t1 σ (47)
Notice that Σ f (
s0 ) ≥ Σ f (σ ). There is a prefix s of s0 such
that s ∈ Ψ
(Σf ). Take t ∈ L/ s satisfying ω 1 = s t. From (40), Similarly, from δd ( 2 ) = 3 , and (
2 , σ q2
, N Σf (
q2 , F ), ( ω)
)∈
σ
−1 2 , we know that there exist (

Σ f (
ω)
) ∈ 3 and
we have ω 2 ∈ P (P (
s t )). Furthermore, from (37) and (39), q3 , F ), (
q3 , N

σ σ
2 ∈ L(
s2 σ ) and
q2 , σ 2 ∈ L(
t2 σ q2
, σ
), such that
we know that
f ( f ( f ( j ) :
trj σ q2 , s2 σ
δ( 2 ) = q3 q
,
δ( 2 ) = q3
.
Σ ω2 ) = max{Σ t0 ), Σ 2 t2 σ (48)
j = 1, . . . , k; r = 1, . . . , n} (41) With the above analogous process, we can obtain two sequences
f ( of states {q1 , q2 , q3 , . . .} and { q1
, q2
, q3
· · ·}. Furthermore, the
i.e., Σ ω2 ) = µ. Therefore, from (21), the diagnosability de- and satisfy the condi-
with respect to σ two sequences form two cycles in G
gree λ of G satisfies tions of Definition 11 because { 1 , 2 , . . . , k } is a cycle of
f (
Σ ω2 ) µ F -uncertain states in G d . That is, { 1 , 2 , . . . , k } is a µ-F -
λ = FD s
(t) ≤ = . (42) d , where µ = Σ f (
σ f (
Σ σ) f (
Σ σ) indeterminate cycle in G ω ). From (44), we

have µ = λΣf ( σ ).

Definition 12: A µ-F -indeterminate cycle in diagnoser G d is In the following, we further prove that the above F -

called to be minimal, if for each µ -F -indeterminate cycle in indeterminate cycle is minimal. Suppose that there is another
d , the inequality µ
≥ µ always holds. µ
-F -indeterminate cycle in G d . From Theorem 1, the diagnos-
G
Now we are ready to present the necessary and sufficient ability degree λ of G with respect to σ satisfies the following
condition of the diagnosability for FDESs. inequality:
Theorem 2: Let L be a language generated by an FDES G µ
d = (Q d , δd , 0 ) be the diagnoser of G

d , E with respect λ≤ (49)
and G f (
Σ σ)
to σ∈E fail . G
is f -diagnosable with degree λ with respect to
σ, if and only if, there is a minimal µ-F -indeterminate cycle in i.e., µ
≥ λΣ f (
σ ) = µ, which shows that { 1 , 2 , . . . , k } is a
G d , where µ = λΣ f (σ ). minimal µ-F -indeterminate cycle.
Proof: Necessity: Assume that G is f -diagnosable with degree Sufficiency: Assume that there is a minimal µ-F -
λ with respect to σ . By Definition 5, for any n0 ∈ N, there are indeterminate cycle in G d , where µ = λΣ f (
σ ). We denote the
s ∈ Ψ ( f ) and
Σ t ∈ L/ s that |
t| ≥ n0 , we have
diagnosability degree of G with respect to σ as x. Then by the
σ
above proof of Necessity, there is a minimal µ
-F -indeterminate
min{Σ f ( f (
σ ), Σ ω) : ω ∈ P −1 (P s
(t))} d , where µ
= xΣ f (
cycle in G σ ). From the minimality of µ-F -
s
σ
FD (t) = σ
= λ.
σ f (
Σ σ) indeterminate cycle, we have µ
≥ µ. Likewise, from the mini-
(43) mality of µ
-F -indeterminate cycle, we have µ ≥ µ
. That is
∈ P −1 (P
Hence, there exists ω (
s
t )) such that f ( f (

σ σ λΣ σ ) = µ = µ
= xΣ σ) (50)
Σ f (
ω ) = λΣ f (
σ) < Σ f (
σ) (44) with re-
i.e., x = λ. Therefore, the diagnosability degree of G
and spect to σ is λ.
Theorem 3: Let L be a language generated by an FDES G.
Σ f (
ω ) = min{Σ f ( f (
σ ), Σ ω) : ω ∈ P −1 (P s
(t))} (45) is f -diagnosable with degree λ, if and only if the fol-
σ σ 1) G
f ( f ( lowing two conditions hold:
i.e., for any ω
∈ P −1 (P s
( t)), we have Σ ω) ≤ Σ ω
). d with re-
fail such that the diagnoser G
σ σ a) there is σ∈E
Denote δ( q0 , s
t) = q and δ( q0 , ω
) = q
. It is obvious that contains a minimal µ-F -indeterminate cycle,
spect to σ
δd (
0 , P s
( t)) is F -uncertain since ( q
, N Σf (
q , F ), ( ω)
) are its where µ = λΣ f (
σ );
σ
b) for any σ

∈ Efail , if the diagnoser with respect to
two elements. Note that the number of states of G is finite, and
δd (
0 , P s
( t)) and its all successor states form a cycle of G d
contains a minimal µ
-F -indeterminate cycle, then
σ
σ
µ
≥ λΣ f (σ
).
when n0 is sufficiently large. Moreover, all of the states in the
2) G is completely f -undiagnosable, if and only if there is construction of the diagnosers. In the following, we roughly
∈E
σ fail such that the diagnoser G d with respect to σ analyze the complexity of constructing these diagnosers. Sup-
pose that |Q| and |E| are the numbers of fuzzy states and
contains a 0-F -indeterminate cycle.
3) G is completely f -diagnosable, if and only if for any fuzzy events of fuzzy finite automaton G = (Q,
E, q0 ), re-
δ,
∈E
σ fail , the diagnoser G d with respect to σ does not
spectively. Let Gd = (Qd , Ed , δd , 0 ) be the diagnoser with
contain any F -indeterminate cycle. respect to σ ∈E fail . There are at most |E| + 1 possible fail-
Proof: (1) Necessity: Suppose that G is f -diagnosable with ure labels for the set of possible failure labels = {F } ∪
degree λ. By Definition 6, there exists σ ∈E
fail such that G
{N µ : µ = Σ f (
a), Due to Q
a ∈ E}. d ⊆ 2Qσ × , the number
is f -diagnosable with respect to σ with degree λ, and there is
not σ fail that the diagnosability degree with respect to σ

∈ E
of states in G d is at most 2|Q|×(|E|+1) . Notice that E d ⊆ E,
is less than λ. From Theorem 2, we know that the diagnoser so the maximum number of transitions in the diagnoser Gd
with respect to σ contains a minimal µ-F -indeterminate cycle, is 2|Q|×(|E|+1) × |E|.
That is, the complexity of constructing
f ( fail , if the diagnoser with
where µ = λΣ
∈ E
σ ); and for any σ one diagnoser is O(2|Q|×(|E|+1) × |E|). Since E fail ⊆ E, the

contains a minimal µ -F -indeterminate cycle, then

respect to σ overall complexity of the above computing process of testing
λ ≤ λ
. From Theorem 2 again, we obtain that
diagnosability of FDESs is, at most, O(2|Q|×(|E|+1) × |E| 2 ).
µ µ
λ= λ
= . (51)
f (
Σ σ) f (
Σ σ
) D. Diagnosability for Multiple Failure Types
Therefore, µ
≥ λΣ f (
σ
). In this section, we extend the fuzzy diagnosability of FDESs
fail such that the diag- with one single failure type to the case of multiple failure types.
Sufficiency: Assume that there is σ ∈E
Assume that a fuzzy system has events of multiple failure
noser with respect to σ contains a minimal µ-F -indeterminate fail = { :Σ f (
f ( fail , if the diag- types f1 , f2 , . . . , fm , and E σ∈E σ ) > 0}. For
cycle, where µ = λΣ
∈ E
σ ); and for any σ i i
fail , we define the fuzzy fi -diagnosability function
noser with respect to σ
contains a minimal µ
-F -indeterminate each σ∈E i
cycle, then µ
≥ λΣ f (
σ
). By Theorem 2, G is f -diagnosable as
with respect to σ with degree λ, and the diagnosability degree i
FD ∗ → [0, 1]
:E (53)
of G with respect to each other σ
∈ Efail is not less than λ. By σ
Definition 6, G is f -diagnosable with degree λ. f ) and

where i = 1, 2, . . . , m, and, for s ∈ Ψ (Σ t ∈ L/
s,
2) It is a special case of (1) when λ = 0. σ i
3) It holds clearly since it is the inverse and negative propo- f (

min{Σ f (
σ ), Σ ω) : ω ∈ P −1 (P s
(t))}
sition of (1).
i
FD s
(t) =
i i

σ σ
. (54)
σ f (
Σ σ)
i
C. Computing Process to Check the Diagnosability of FDESs = (Q,

E, q0 ) with multi-
δ,
Definition 13: Given an FDES G
According to Theorems 2 and 3, we give the following method ple failure types f1 , f2 , . . . , fm . Let L be a language generated
to check the diagnosability for FDESs. If
by G.
Step 1: Construct the diagnosers for each σ ∈E fail .
Let L be a language generated by an FDES G. For each (∃n0 ∈ N)(∀
s ∈ Ψ f ))(∀
(Σ t ∈ L/
s)
σ i
d = (Q
fail , we construct the diagnoser G d , δd , 0 ) by
d , E (55)
∈E
σ (|
t| ≥ n0 ⇒ FD
i
s
(t) = 1)
σ
means of Definitions 7 and 9. This specific procedure can be
seen in Section IV for the details. then G is said to be completely fi -diagnosable with respect to
Step 2: Check whether there exists F -indeterminate cycles in ; if for any n0 ∈ N, the following holds:
σ
these diagnosers.
If the diagnoser G d with respect to each σ ∈E fail does i
min{FD s
(t) : s ∈ Ψ f ),
(Σ s, |
t ∈ L/ t| ≥ n0 } = λ (56)
σ σ i

not contain F -indeterminate cycle, then G is completely f -
diagnosable; otherwise, we perform the next step. where 0 ≤ λ < 1, then G is said to be fi -diagnosable with
Step 3: Collect all of the minimal F -indeterminate cycles and degree λ with respect to σ ; if for any n0 ∈ N, there are s ∈
compute the diagnosability degree of G. Ψ ( f ) and
Σ t ∈ L/
s that |
t| ≥ n0 , we have
σ i
For σ
∈ Efail , denote the minimal F -indeterminate cycle as
µ-F -indeterminate cycle. Let
i
FD
σ
s
(t) = 0 (57)

µ fail . then G is said to be completely fi -undiagnosable with respect
λ = min ∈E
:σ (52)
f (
Σ σ) .
to σ
The diagnosers with multiple failure types can be similarly
If λ = 0, then G is completely f -undiagnosable; if 0 < λ < 1,
constructed, and we only need to use the following tensor prod-

then G is f -diagnosable with degree λ. uct to represent the possible failure labels
Remark 5: The main complexity of the above computing
process of checking the diagnosability for FDESs lies in the = 1 × 2 × · · · × m (58)
Fig. 2. Diagnosers with respect to θ and τ in Example 2.

Fig. 1. Crisp DES G in Example 2.
and any t1 ∈ L/s1 , i.e., s1 = αθ and t1 = β n (n ≥ 2), we have
in which i = {Fi } ∪ {Niµ : µ = Σ f (

a), for each i =
a ∈ E} P −1 (P (s1 t1 )) = P −1 (P (αθβ n )) = {αθβ n }.
i

1, 2, · · · , m. A state ∈ Qd of the diagnoser is of the form Therefore, D(s1 t1 ) = 1. Similarly, for the failure type f2 ,
Ψ(Ef 2 ) = {ατ }, for any n0 ∈ N, if we take s2 = ατ and
= {(
q1 , (11 , 12 , . . . , 1m )), . . . , (
qn , (n 1 , n 2 , . . . , n m ))} t2 = αn ∈ L/s2 , then
in which qj ∈ Q , j i = Fi or j i = N µ (µ ∈ [0, 1); j = P −1 (P (s2 t2 )) = P −1 (P (ατ αn )) = {ατ αn , αγαn }.

σ i
1, 2, . . . , n; i = 1, 2, . . . , m), and each pair (qj , j i ) is called So, D(s2 t2 ) = 0. By Definition 1, L is f1 -diagnosable but not
to be a component of state . f2 -diagnosable.
Remark 6: Each component of the states of the diagnoser is 2) Fuzzy Diagnosability Approach:
responsible for detecting one type of failure, and the failure of Next, we verify the above result by means of the fuzzy diag-
type fi is reported when a component of a certain state is labeled nosability approach for FDESs presented in Section V.
by Fi . Therefore, the approach to check the diagnosability with a The crisp DES G can be viewed as a special FDES with the
single failure type is also suitable for the case of multiple failure fuzzy states: q0 = [1, 0, 0, 0], q1 = [0, 1, 0, 0], q2 = [0, 0, 1, 0],
types. Since the complexity of testing the diagnosability with q3 = [0, 0, 0, 1]; and fuzzy events
a single failure type is O(2|Q|×(|E|+1) × |E| 2 ), the complexity 
0 1 0 0
 
0 0 0 0

of testing the diagnosability with m failure types is at most 0 0 0 0 0 0 0 0
α= β =  
O(2|Q|×(m (|E|+1)) × |E| 2 ). 0 0 1 0 0 0 0 0
0 0 0 0 0 0 0 1
VI. APPLICATIONS TO DIAGNOSIS OF CRISP DESS AND    
0 0 0 0 0 0 0 0
FDESS: SOME EXAMPLES 0 0 1 0 0 0 0 1
γ= θ =  
In this section, we give some examples to illustrate that the 0 0 0 0 0 0 0 0
fuzzy diagnosability for FDESs presented before may be applied 0 0 0 0 0 0 0 0
to the failure diagnosis for crisp DESs and the failure diagnosis  
0 0 0 0
for FDESs. 0 0 1 0
τ = .
0 0 0 0
A. Application to Diagnosis of Crisp DESs 0 0 0 0
As we know, for a crisp DES with n crisp states, if we regard From Eo = {α, β}, Ef 1 = {θ}, and Ef 2 = {τ }, we know
the crisp states as n-D unit state vectors, and the events are o are determined as
that the observable event fuzzy subset Σ
denoted as n × n matrices that every entry takes two possible o (γ) = Σ
Σ o (θ) = Σ
o (τ ) = 0, o (α) = Σ
Σ o (β) = 1
values 0 or 1, then the crisp DES can be viewed as a special
FDES. Therefore, the fuzzy diagnosability for FDESs presented f and Σ
and the failure event fuzzy subsets Σ f are determined
1 2
above may be used to cope with the problem of failure diagnosis as follows:
for crisp DESs. The following example verifies the view. Σ f (β) = Σ
f (α) = Σ f (γ) = Σ
f (τ ) = 0, f (θ) = 1
Σ
1 1 1 1 1
Example 2. Consider the crisp DES G described in Fig. 1, in
which the set of observable events Eo = {α, β}, and the two f (α) = Σ
Σ f (β) = Σ
f (γ) = Σ
f (θ) = 0, f (τ ) = 1.
Σ
2 2 2 2 2
failure event sets are Ef 1 = {θ} and Ef 2 = {τ }, respectively.
For event θ and event τ , from Definition 7, we know that the
In the following, we prove that L is f1 -diagnosable but not
diagnoser with respect to θ and that with respect to τ are the
f2 -diagnosable by means of two methods: one is the classical
same, which are shown as Fig. 2.
diagnosability approach for crisp DESs proposed in [24] (i.e.,
For the failure type f1 , notice that both of the diagnosers
Definition 1), another is the fuzzy diagnosability for FDESs
in Fig. 2 do not contain F1 -indeterminate cycle since all of
proposed in Section V.
the states are F1 -certain. However, for the failure type f2 , the
1) Classical Diagnosability Approach:
diagnosers contain a 0-F -indeterminate cycle { 3 , α, 3 , α},
First, we prove the result using the classical diagnosabil-
where
ity approach for crisp DESs. For the failure type f1 , due to
Ψ(Ef 1 ) = {αθ}, if we take n0 = 2, then for any s1 ∈ Ψ(Ef 1 ) 3 = {(q2 , (N10 , N20 )), (q2 , (N10 , F2 ))}.
Therefore, by Theorem 3, we obtain the result that G is com-

pletely f1 -diagnosable but G is completely f2 -undiagnosable,
which coincides with the result obtained by the classical diag-
nosability approach for crisp DESs.
B. Application to Diagnosis of FDESs

Fig. 3.
Treatment process of the animal modeled by FDES G.
FDESs combines fuzzy set theory [31] with crisp DESs [4],
and it has been successfully applied to the real-world complex
systems such as biomedical systems [11]–[13], [29], [30] and [7, p. 323], [29, pp. 668–669], [30, pp. 967-970])
robotic control [8], [9], in which vagueness, impreciseness, and
subjectivity are typical features. Therefore, the fuzzy diagnos-    
0.4 0.9 0.4 0.4 0 0
ability for FDESs presented above may better cope with the =  0 0.4 0.4  β =  0.9 0.4 0 
α
problem of uncertainty emerging from the issue of failure diag- 0 0 0.4 0.4 0.4 0.4
nosis. Next, we apply our framework formalized in Sections III    
and IV to the following two examples of diagnosis for FDESs 0.9 0.9 0.4 0.5 0 0
(Example 3 and Example 4). In particular, Example 3 arising =  0 0.4 0.4  θ =  0.1 0.1 0  .
γ
from medical diagnosis and treatment may be viewed as an 0 0 0.4 0.1 0.1 0.1
applicable background of the diagnosability for FDESs.
Example 3. Suppose that there is an animal growing sick from By the max–min operation (i.e., Definition 2), we can calcu-
a new disease. For the new disease, the doctor has no complete late the other states as
knowledge about it, but he (or she) believes, by experience,
that these drugs, such as theophylline, ipratropium bromide, q1 = [0.4, 0.9, 0.4] q2 = [0.9, 0.4, 0.4] q5 = [0.4, 0.5, 0.4]
erythromycin ethylsuccinate, and dopamine (denoted as α, β, γ
q3 = [0.9, 0.9, 0.4] q4 = [0.5, 0.1, 0] q5 = [0.4, 0.5, 0.4]
and θ, respectively) may be useful to the disease.
For simplicity, it is assumed that the doctor considers roughly q6 = [0.5, 0.4, 0.4] q7 = [0.5, 0.5, 0.4].
the animal’s condition to be three states, say “good,” “fair,” and
“poor.” As mentioned in Section I and [11], [12], and [21], it is As the continuation of our work [16], [21], each fuzzy event is
vague when the animal’s condition is said to be “good,” “fair,” or observable with a certain membership degree in the framework
“poor,” since the animal’s condition can simultaneously belong of FDESs. For example, some symptoms of the drugs such as
to “good,” “fair,” and “poor” with respective memberships in skin turning pale can be clearly seen by the doctor, but some
the real-life situation [12], [20], [21]. Therefore, when an FDES symptoms may be observable only by means of special medi-
is used to model the treatment process of the animal, a fuzzy cal instruments, even some potential side effects such as high
state is naturally denoted as a 3-D vector white blood cell count are undesired failures. Assume that high
white blood cell count is considered as undesired failure in this
good fair poor model, and the observability degree and the possibility of failure
(59)
q = [a1 a2 a3 ] occurring are evaluated as follows by the doctor:
which is represented as the possibility distribution of the ani-
mal’s condition over states “good,” “fair,” and “poor.” o (
Σ α) = 0.5, Σ = 0.4,
o (β) o (
Σ γ ) = 0.6, Σ = 0.3
o (θ)
Similarly, it is imprecise to say at what point exactly the f ( = 0.2,
f (β) f ( = 0.4.
f (θ)
Σ α) = 0.1, Σ Σ γ ) = 0.3, Σ
animal has changed from one state to another state after a drug
treatment (i.e., event), because each drug event occurring may
The problem of the diagnosis for FDES G is how to detect
lead a state to multistates with respective memberships [20,
p. 82]. Therefore, in the treatment process modeled by an FDES, the failure (i.e., high white blood cell count) in time and further
a fuzzy event is represented as a 3 × 3 matrix compute the possibility of failure occurring during the treatment
process.
good fair poor In order to explain the notion of fuzzy diagnosability for
good a11 a12 a13 FDESs formalized in Section III and illustrate the necessary and
(60)
=
σ fair  a21 a22 a23  . sufficient condition for diagnosis of FDESs (i.e., Theorems 2
poor a31 a32 a33 and 3), we present the following two methods: one is the defini-
tion approach (i.e., Definitions 5 and 6), another is the diagnoser
Suppose that the treatment process of the animal is modeled approach proposed in Sections IV and V.
by the following FDES G = (Q, E, q0 ), shown as Fig. 3, in
δ, 1) Definition Approach: First, we investigate the diagnosis
which the initial state vector is q0 = [0.9, 0.1, 0], and the drug problem for the treatment process G by means of the definition
events are evaluated as follows by means of medical theory and of diagnosability for FDESs.
the doctor’s experience (there are many methods for estimating For fuzzy event α , notice that Σ f (
α) is the least among
the memberships in the literature, for example, [6, pp. 256–260],
{Σf (
∈ Σ}, and therefore, by (21), for any s ∈ Ψα(Σ
σ) : σ f )
and s that |
t ∈ L/ t| ≥ n0 (n0 ∈ N), we have
f (
min{Σ f (
α), Σ ω) : ω s
∈ P −1 (Pα(t))}
s
FDα(t) =
α
=1
f (
Σ α) Fig. 4. and
Diagnosers with respect to α γ.
which means that the possibility of failure occurring (i.e., high

white blood cell count) on each treatment strategy ω is not less
(i.e., drug theophylline treatment).
than that on α
For fuzzy event β, take n0 = 3. Then for any s ∈ Ψ (Σ
f)
β
d with respect to β.
and s that | s
Fig. 5. Diagnoser G
t ∈ L/ t| ≥ n0 , we have FDβ( t) = 1 because all
−1
∈ P (Pβ(
of the strings ω s For simplicity,
or θ.
t)) contain γ

β
we only calculate the case that s = α β and
t=γ β ∈ L/
α s as
follows (the other cases can be similarly calculated):
s
P −1 (Pβ(t)) = { γ α
αβ , α γ α
β θα
β, γ α
β , θα γ α
β
β}.

β
d with respect to θ.
Fig. 6. Diagnoser G
Therefore, Σ f (
ω ) = 0.3 or Σ f ( ∈
ω ) = 0.4 for each ω
s
P −1 (Pβ(t)). By (21), we obtain , the event set of the diagnoser with respect to α
For α is

β
d = {
E γ }.
min{Σ Σ
f (β), f ( s
∈ P −1 (Pβ(
ω) : ω t))}

β
The states of the diagnoser with respect to α are 0 =
FDβ(
st) =

f (β)
Σ q0 , N 0 )}, and 1 = {(
{( q3 , F ), (
q7 , F )}. By Definition 7, the
min{0.2, 0.3} diagnoser with respect to α is constructed as in Fig. 4.
≥ =1 For β, the event set of the diagnoser with respect to β is

0.2
which indicates that we can detect the occurrence of events d = {
E }
α, γ
whose possibilities of failure occurring are not less than that on
and the states of the diagnoser with respect to β are as follows:
β within delay of three steps (i.e., n0 = 3).
The case of fuzzy event γ and all of
is similar to that of β, 0 = {(
q0 , N 0 )} 1 = {(
q1 , N 0.1 ), (
q5 , F )}
the strings ω −1
∈ P (P s
(t)) contain γ
or θ. By (21), we have

γ γ 2 = {(
q3 , F ), (
q7 , F )} 3 = {(
q1 , F ), (
q5 , F )}.
s
By Definition 7, the diagnoser with respect to β is constructed
FD γ
(t) = 1.
if we take s = θ and
For fuzzy event θ, t=α γ α
(β )n 0 ∈ L/
s as in Fig. 5.
(n0 ∈ N), then , it is not difficult to know from Definition 7 that the
For γ
γ α γ α diagnoser with respect to γ is the same as the diagnoser with
P −1 (P s
(t)) = P −1 (
α(β )n 0 ) = {
α(β )n 0 , θα γ α
(β )n 0 }.

θ θ
θ as shown in Fig. 4.
respect to α
the event set of the diagnoser is
For θ,
equals to 0.75 with
By (21), the diagnosability degree of G
respect to θ for d = {
E γ
α, β, }
min{Σ Σ
f (θ), f ( γ α
α(β f (θα
)n 0 ), Σ γ α
(β )n 0 )} and the states of the diagnoser with respect to θ are as follows:
FD s
(t) =
θ
Σ
f (θ) 0 = {(
q0 , N 0 )} 1 = {(
q1 , N 0.1 ), (
q5 , F )}
min{0.4, 0.3, 0.4} 2 = {(
q2 , N 0.2 ), (
q6 , F )} 3 = {(
q3 , N 0.3 ), (
q7 , F )}
= = 0.75.
0.4
4 = {(
q1 , N 0.3 ), (
q5 , F )} 6 = {(
q2 , N 0.3 ), (
q6 , F )}.
By Definition 6, G is f -diagnosable with degree 0.75. The
By Definition 7, the diagnoser with respect to θ is constructed
result shows that we cannot detect the failure occurring on the
as in Fig. 6.
treatment strategy ω whose possibility of failure occurring is
but we may ensure that the possibilities of
f (θ), Although there is not any F -indeterminate cycle in Figs. 4
not less than Σ and 5, but there is a minimal 0.3-F -indeterminate cycle
failure occurring on all of the treatment strategies are at least 5 , γ
{ , 4 , β,
3 , α , 3 } in Fig. 6, where
0.75 times that on θ.
2) Diagnoser Approach: In the following, we further verify 3 = {(
q3 , N 0.3 ), (
q7 , F )} 4 = {(
q1 , N 0.3 ), (
q5 , F )}
the above results by using the necessary and sufficient condition 5 = {(
q2 , N 0.3 ), (
q6 , F )}.
for the diagnosability of FDESs (i.e., Theorem 3).
First, we construct the diagnosers for each fuzzy event in That is, there is an F -indeterminate cycle only for θ ∈ E
fail ,

Efail = { γ
α, β,
, θ}.
and the diagnoser Gd with respect to θ contains a minimal
Fig. 7. in Example 4.
Fuzzy automaton G
µ-F -indeterminate cycle, where µ = λΣ Noticing that µ =

f (θ).
Fig. 8. Diagnosers with respect to each σ∈E fa il . (a) Diagnosers with

0.3 (see Fig. 6) and Σf (θ) = 0.4, we obtain that λ = 0.75. .
and
respect to α γ . (b) Diagnoser with respect to β
Therefore, by Theorem 3, G is f -diagnosable with degree 0.75.
In fact, from Fig. 6, we know that corresponding to the
5 , γ VII. CONCLUDING REMARK
F -indeterminate cycle { , 4 , β,
3 , α , 3 }, the “failed-
state” cycle {(q7 , F ), (
q5 , F ), (
q6 , F ), (q7 , F )} carries the fail- In this paper, we have dealt with the failure diagnosis of
ure label F , which shows that the possibility of failure occurring FDESs. We have presented a fuzzy diagnosability approach for
on the corresponding treatment strategy exceeds 0.4; however, FDESs. A fuzzy diagnosability function that takes values in the
the “normal-state” cycle interval [0, 1] rather than {0, 1} has been introduced to charac-
terize the diagnosability degree of the fuzzy systems. We have
{(
q3 , N 0.3 )(
q1 , N 0.3 )(
q2 , N 0.3 )(
q3 , N 0.3 )} formalized the construction of the diagnosers that are used to
perform fuzzy diagnosis. In particular, we have proposed a nec-
carries the label N 0.3 , which means that the possibility of fail- essary and sufficient condition for the diagnosability of FDESs,
ure occurring on the corresponding treatment strategy does not and a method for checking the diagnosability condition has been
exceed 0.4. Therefore, we cannot detect the failure occurring on given. The newly proposed diagnosability approach allows us to
the treatment strategies whose possibility of failure occurring deal with the problem of failure diagnosis for both crisp DESs
is not less than Σ but we may ensure that the possibili-
f (θ), and FDESs, which may better deal with the problems of fuzzi-
ties of failure occurring on all of the treatment strategies are at ness, impreciseness, and subjectivity in the failure diagnosis.
least 0.3, which coincides with the results obtained above by the With the results obtained in this paper, a further issue
Definition Approach. worthy of consideration is the I-diagnosability and the AA-
Example 4. Consider the FDES G represented in Fig. 7, where diagnosability of FDESs, as those investigated in the frame-
the fuzzy states and fuzzy events are, respectively, given as works of crisp DESs [24] and stochastic DESs [27]. Of course,
follows: since we have focused on the centralized diagnosis of FDESs
in this paper, the decentralized diagnosis together with the safe
q0 = [0.9, 0.1, 0] q1 = [0.4, 0.9, 0.4] q2 = [0.9, 0.4, 0.4] diagnosis of FDESs is another interesting issue, just as we have
dealt with them in stochastic DESs [15], [17]. We would like to
q3 = [0.9, 0.9, 0.4] q4 = [0.4, 0.1, 0] q5 = [0.4, 0.4, 0.4] consider these issues in the subsequent work.
   
0.4 0.9 0.4 0.4 0 0
= 0
α 0.4 0.4  β =  0.9 0.4 0  ACKNOWLEDGMENT
0 0 0.4 0.4 0.4 0.4
The authors would like to thank Professor N. R. Pal, Editor-
 
0.9 0.9 0.4 in-Chief, the Associate Editor, and the anonymous reviewers for
= 0
γ 0.4 0.4  . their invaluable suggestions and comments that greatly helped
0 0 0.4 to improve the quality of the paper.
Suppose that the observability degree of the events and the REFERENCES
possibility of failure occurring are defined as [1] Y. Cao and M. Ying, “Supervisory control of fuzzy discrete event systems,”
IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 35, no. 2, pp. 366–371,
Apr. 2005.
o (
Σ α) = 0.6 Σ = 0.4
o (β) o (
Σ γ ) = 0.6 [2] Y. Cao and M. Ying, “Observability and decentralized control of fuzzy
discrete-event systems,” IEEE Trans. Fuzzy Syst., vol. 14, no. 2, pp. 202–
f (
Σ α) = 0.1 Σ = 0.2
f (β) f (
Σ γ ) = 0.3. 216, Apr. 2006.
[3] Y. Cao, M. Ying, and G. Chen, “State-based control of fuzzy discrete-
event systems,” IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 37,
Then the diagnosers with respect to each σ ∈E fail are con- no. 2, pp. 410–424, Apr. 2007.
structed as in Fig. 8. [4] C. G. Cassandras and S. Lafortune, Introduction to Discrete Event Systems.
Boston, MA: Kluwer, 1999.
From Fig. 8, we know that all of the diagnosers do not con- [5] R. Debouk, “Failure diagnosis of decentralized discrete event systems,”
tain any F -indeterminate cycle. Therefore, by Theorem 3, G is Ph.D. dissertation, Elec. Eng. Comp. Sci. Dept., Univ. Michigan, Ann
completely f -diagnosable. Arbor, MI, 2000.
[6] D. Dubois and H. Prade, Fuzzy Sets and Systems: Theory and Applications. [27] D. Thorsley and D. Teneketzis, “Diagnosability of stochastic discrete-
New York: Academic, 1980. event systems,” IEEE Trans. Autom. Control, vol. 50, no. 4, pp. 476–492,
[7] J. W. Huang, C. M. Held, and R. J. Roy, “Hemodynamic management Apr. 2005.
with multiple drug using fuzzy logic,” in Fuzzy and Neuro-Fuzzy Systems [28] G. Westerman, R. Kumar, C. Stround, and J. Heath, “Discrete event system
in Medicine, H. N. Teodoresch, A. Kandel, and L. C. Jain, Eds. Boca approach for delay fault analysis in digital circuits,” in Proc. ’98 Amer.
Raton, FL: CRC, 1999, pp. 319–340. Control Conf., Jun., pp. 239–243.
[8] R. Huq, G. K. I. Mann, and R. G. Gosine, “Distributed fuzzy discrete [29] H. Ying, F. Lin, R. D. MacArthur, J. A. Cohn, D. C. Barth-Jones, H. Ye,
event system for robotic sensory information processing,” Expert Syst., and L. R. Crane, “A fuzzy discrete event system approach to determin-
vol. 23, no. 5, pp. 273–289, Nov. 2006. ing optimal HIV/AIDS treatment regimens,” IEEE Trans. Inf. Technol.
[9] R. Huq, G. K. I. Mann, and R. G. Gosine, “Behavior-modulation technique Biomed., vol. 10, no. 4, pp. 663–676, Oct. 2006.
in mobile robotics using fuzzy discrete event system,” IEEE Trans. Robot., [30] H. Ying, F. Lin, R. D. MacArthur, J. A. Cohn, D. C. Barth-Jones, H. Ye, and
vol. 22, no. 5, pp. 903–916, Oct. 2006. L. R. Crane, “A self-learning fuzzy discrete event system for HIV/AIDS
[10] E. Kilic, “Diagnosability of fuzzy discrete event systems,” Inform. Sci., treatment regimen selection,” IEEE Trans. Syst., Man, Cybern. B, Cybern.,
vol. 178, pp. 858–870, 2008. vol. 37, no. 4, pp. 966–979, Aug. 2007.
[11] F. Lin and H. Ying, “Fuzzy discrete event systems and their observability,” [31] L. A. Zadeh, “Fuzzy logic=computing with words,” IEEE Trans. Fuzzy
in Pro. Joint Int. Conf. 9th Int. Fuzzy Systems Assoc. World Congr. 20th Syst., vol. 4, no. 2, pp. 103–111, May 1996.
North Amer. Fuzzy Inform. Process. Soci., Vancouver, BC, Canada, Jul.
25–28, 2001.
[12] F. Lin and H. Ying, “Modeling and control of fuzzy discrete event sys-
tems,” IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 32, no. 4, pp. 408–
415, Aug. 2002.
[13] F. Lin, H. Ying, X. Luan, R. D. MacArthur, J. A. Cohn, D. C. Barth-Jones, Fuchun Liu received the M.S. degree in mathematics
and L. R. Crane, “Fuzzy discrete event systems and its applications to in 1997 from Jiangxi Normal University, Nanchang,
clinical treatment planning,” in Proc. 43rd IEEE Conf. Decision Contr., China, and the Ph.D. degree in computer science from
Budapest, Hungary, Jun. 25–29, 2004, pp. 197–202. Sun Yat-Sen University, Guangzhou, China, in 2008.
[14] F. Lin, H. Ying, R. D. MacArthur, J. A. Cohn, D. C. Barth-Jones, and He is an Associate Professor in the Faculty
L. R. Crane, “Decision making in fuzzy discrete event systems,” Inform. of Computer, Guangdong University of Technol-
Sci., vol. 177, no. 18, pp. 3749–3763, 2007. ogy, Guangzhou, and a Research Fellow in the
[15] F. Liu, D. Qiu, H. Xing, and Z. Fan, “Decentralized diagnosis of stochastic Department of Electrical and Computer Engineer-
discrete event systems,” IEEE Trans. Autom. Control, vol. 53, no. 2, ing, National University of Singapore, Kent Ridge,
pp. 535–546, Mar. 2008. Singapore.
[16] F. Liu and D. Qiu, “Decentralized supervisory control of fuzzy discrete His current research interests include automata
event systems,” Eur. J. Control, vol. 14, no. 3, pp. 234–243, 2008. theory and discrete event systems, mathematical logic, and rough set theory. He
[17] F. Liu and D. Qiu, “Safe diagnosability of stochastic discrete event sys- has authored or coauthored over 20 peer-reviewed papers in academic journals,
tems,” IEEE Trans. Automat. Control, vol. 53, no. 5, pp. 1291–1296, Jun. including two papers about stochastic discrete-event systems published by the
2008. IEEE TRANSACTIONS ON AUTOMATIC CONTROl, and some papers about fuzzy
[18] Y. Pencolé, “Decentralized diagnoser approach: Application to telecom- discrete-event systems published by the European Journal of Control and ac-
munication networks,” in Proc. 11th Int. Workshop Principles Diagnosis cepted by the IEEE TRANSACTIONS ON FUZZY SYSTEMS.
(DX’00), Jun., pp. 185–192.
[19] A. Paoli and S. Lafortune, “Safe diagnosability for fault-tolerant super-
vision of discrete event systems,” Automatica, vol. 41, pp. 1335–1347,
2005.
[20] D. Qiu, “Supervisory control of fuzzy discrete event systems: A formal
approach,” IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 35, no. 1,
pp. 72–88, Feb. 2005.
[21] D. Qiu and F. Liu, “Fuzzy discrete event systems under fuzzy observability Daowen Qiu received the M.S. degree in mathemat-
and a test-algorithm,” IEEE Trans. Fuzzy Syst., to be published. ics in 1993 from Jiangxi Normal University, Nachang,
[22] S. L. Ricker and J. H. van Schuppen, “Decentralized failure diagnosis with China, and the Ph.D. degree in mathematics from Sun
asynchronous communication between supervisors,” in Proc. Eur. Contr. Yat-Sen University, Guangzhou, China, in 2000.
Conf., 2001, pp. 1002–1006. He completed the postdoctoral research in com-
[23] L. Rozé and M. O. Cordier, “Diagnosing discrete event systems: Extending puter science at Tsinghua University, Beijing, China,
the ‘Diagnoser Approach’ to deal with telecommunication networks,” in 2002. Since May, 2004, he has been a Profes-
Discrete Event Dyn. Syst.: Theory Appl., vol. 12, pp. 43–81, 2002. sor of computer science at Sun Yat-Sen University.
[24] M. Sampath, R. Sengupta, S. Lafortune, K. Sinnamohideen, and His current research interests include automata the-
D. Teneketzis, “Diagnosability of discrete-event systems,” IEEE Trans. ory and discrete-event systems, fuzzy logic, models
Automat. Contr., vol. 40, no. 9, pp. 1555–1575, Sep. 1995. of quantum computation, quantum information, and
[25] M. Sampath, R. Sengupta, S. Lafortune, K. Sinnamohideen, and the theory of computation based on nonclassical logic. He has authored or
D. Teneketzis, “Failure diagnosis using discrete-event models,” IEEE coauthored over 40 peer-reviewed papers in academic journals, including In-
Trans. Autom. Control Syst. Technol., vol. 4, no. 2, pp. 105–124, Mar. formation and Computation, Artificial Intelligence, the Journal of Computer
1996. and System Sciences, Theoretical Computer Science, THE IEEE TRANSACTIONS
[26] R. Su and W. Wonham, “Global and local consistencies in distributed ON AUTOMATIC CONTROL, THE IEEE TRANSACTIONS ON SYSTEMS, MAN, AND
fault diagnosis for discrete-event systems,” IEEE Trans. Autom. Control, CYBERNETIC—PART B: CYBERNETICS, THE IEEE TRANSACTIONS ON FUZZY
vol. 50, no. 12, pp. 1923–1935, Dec. 2005. SYSTEMS, Physical Review A, The Journal of Physics A, and Science in China.
On Generalized Fuzzy Belief Functions

in Infinite Spaces
Wei-Zhi Wu, Yee Leung, and Ju-Sheng Mi
Abstract—Determined by a fuzzy implication operator, a gen- and upper probabilities induced by a multivalued mapping car-
eral type of fuzzy belief structure and its induced dual pair of fuzzy rying a probability measure defined over subsets of the do-
belief and plausibility functions in infinite universes of discourse main of the mapping [32]. The belief (respectively, plausibility)
are first defined. Relationship between the belief-structure-based
and the belief-space-based fuzzy Dempster–Shafer models is then function is a monotone Choquet capacity of infinite order (re-
established. It is shown that the lower and upper fuzzy probabil- spectively, alternating Choquet capacity of infinite order) [4]
ities induced by the fuzzy belief space yield a dual pair of fuzzy satisfying the subadditive (respectively, superadditive) property
belief and plausibility functions. For any fuzzy belief structure, at any order [32]. Subadditivity and superadditivity at any order
there must exist a fuzzy belief space such that the fuzzy belief form the essential properties of belief and plausibility functions,
and plausibility functions defined by the given fuzzy belief struc-
ture are just the lower and upper fuzzy probabilities induced by respectively.
the fuzzy belief space, respectively. Essential properties of the fuzzy One of the main directions of research in the Dempster–Shafer
belief and plausibility functions are also examined. The fuzzy belief theory of evidence is naturally the generalization of the belief
and plausibility functions are, respectively, a fuzzy monotone Cho- and plausibility functions. There are mainly two approaches
quet capacity and a fuzzy alternating Choquet capacity of infinite for the definitions of belief and plausibility functions, namely,
order.
the belief-space-based and belief-structure-based approaches. In
Index Terms—Belief functions, fuzzy implication operators, the belief-space-based approach, a belief space that consists of a
fuzzy rough sets, monotone capacity, rough sets. multivalued mapping (or a binary relation) carrying a probability
distribution on the domain of the mapping is a primitive notion.
The lower and upper probabilities are constructed by means of
I. INTRODUCTION the belief space. Unlike the belief-space-based approach, the
S A GENERALIZATION of Bayesian theory of subjec- belief-structure-based approach does not take belief spaces as
A tive judgment, the Dempster–Shafer theory of evidence
(also called the theory of belief functions) is a method devel-
primitive notion. It regards the abstract belief structure (defined
by axioms) as primitive notion. The belief and plausibility func-
oped to model and manipulate uncertain, imprecise, incomplete, tions are derived from the belief structure. In the finite universes
and even vague information. It was originated by Dempster’s of discourse, it has been proved that a belief space can yield
concept of lower and upper probabilities [8] and extended by a belief structure such that the lower and upper probabilities
Shafer [32] as a theory. The basic representational structure in in the belief space are, respectively, the belief and plausibility
this theory is a belief structure, which consists of a family of functions defined by the belief structure. Conversely, for a belief
subsets, called focal elements, with associated individual posi- structure defined by axioms, there must exist a belief space such
tive weights summing to 1. The fundamental numeric measures that the belief and plausibility functions defined by the belief
derived from the belief structure are a dual pair of belief and structure are, respectively, the the lower and upper probabilities
plausibility functions. Since its inception, evidential reasoning induced by the belief space [7], [32], [43], [51].
has emerged as a powerful methodology for pattern recognition, Another important method used to deal with uncertainty in
image analysis, diagnosis, knowledge discovery, information fu- intelligent systems characterized by insufficient and incom-
sion, and decision making. plete information is the theory of rough sets originated by
The original concepts of belief and plausibility functions in Pawlak [24], [25]. The basic structure of rough set theory is an
the Dempster–Shafer theory of evidence come from the lower approximation space consisting of a universe of discourse and a
binary relation imposed on it. Using the concepts of lower and
upper approximations in rough set theory, knowledge hidden
Manuscript received April 11, 2008; accepted December 4, 2008. First pub-
in information systems may be unravelled and expressed in the
lished January 27, 2009; current version published April 1, 2009. This work form of decision rules (see, e.g., the literature cited in [25]–[27]).
was supported by the Hong Kong Research Grants Council under the Ear- The lower and upper probabilities in the Dempster–Shafer the-
marked Grant CUHK 4126/04H.
W.-Z. Wu is with the School of Mathematics, Physics and Informa-
ory of evidence seem to have some natural correspondences
tion Science, Zhejiang Ocean University, Zhoushan 316004, China (e-mail: with the lower and upper approximations in rough set theory.
wuwz@zjou.edu.cn). The relationships between the Dempster–Shafer theory of evi-
Y. Leung is with the Department of Geography and Resource Management,
Center for Environmental Policy and Resource Management, Institute of Space
dence and rough set theory have received wide attention in the
and Earth Information Science, The Chinese University of Hong Kong, Shatin, research community [33]–[35], [43], [51]. In finite universes of
Hong Kong (e-mail: yeeleung@cuhk.edu.hk). discourse, it has been demonstrated that for each type of belief
J.-S. Mi is with the College of Mathematics and Information Science, Hebei
Normal University, Hebei 050016, China (e-mail: mijsh@263.net).
structure and its induced belief and plausibility functions, there
Digital Object Identifier 10.1109/TFUZZ.2009.2013634 must exist an approximation space such that its derived dual
1063-6706/$25.00 © 2009 IEEE
pair of lower and upper approximation operators may be used that belief functions satisfy the subadditive property at any
to interpret the belief and plausibility functions [43], [51]. It order.
can be observed that the belief and plausibility functions in the It is well known that there are a lot of implicators [31] that
Dempster–Shafer theory of evidence and the lower and upper have been widely used in fuzzy sets research. It should be noted
approximations in rough set theory, respectively, capture the that fuzzy inference results often depend upon the choice of
mechanisms of numeric and nonnumeric aspects of uncertain the implicator and choice of the triangular norm. For analyzing
knowledge. uncertainty in complicated fuzzy systems, lower and upper fuzzy
Shafer’s belief and plausibility functions are constructed un- rough approximations defined by arbitrary implicators in rough
der the assumption that the focal elements in the belief structure set theory were examined in a number of studies [5], [21], [22],
are all crisp. In some situations, it seems to be quite natural [28], [42], [54]. With reference to different requirements, more
that the evidence mass may be assigned to a fuzzy subset of the general fuzzy belief functions associated with various fuzzy
universe of discourse. In fact, combining the Dempster–Shafer implicators need to be developed to evaluate fuzzy events.
theory and fuzzy set theory has been suggested to be a way to
deal with different kinds of uncertain information in intelligent II. BRIEF NOTE ON RESULTS AND CONTRIBUTIONS
systems in a number of studies. Zadeh was the first to generalize
the Dempster–Shafer theory to the fuzzy environment. He de- The purpose of this paper is to develop a general type of
fined the concepts of expected certainty and expected possibility fuzzy belief structure induced by a fuzzy implication operator
as a generalization of Dempster–Shafer belief and plausibility in an infinite universe of discourse and the fuzzy belief and
functions. The generalization is based on the concept of in- plausibility functions through the generalization of the Shafer
formation granularity [56] and the theory of possibility [55]. approach to the fuzzy environment.
Following Zadeh’s work, the Dempster–Shafer theory has been In the next section, we review some basic notions of fuzzy
enriched in slightly different ways in order to evaluate the de- logical operators and fuzzy relations. In Section IV, we show
gree of belief in fuzzy events (see, e.g., [9], [11], [13], [15], [16], some useful properties of a fuzzy probability measure on an in-
[18], [20], [23], [29], [30], [36]–[39], [46], [47], [52], and [53]). finite universe of discourse defined by Chen et al. [3]. We show,
For example, based on the upper and lower expectations of a in particular, that Zadeh’s probability measures of fuzzy events
membership function, Smets [36] defined the probability of a can be extended to infinite countable universes of discourse
fuzzy event within the framework of belief functions and showed [see (11)]. In Section V, we define generalized fuzzy rough
that the fuzzy belief function satisfies the subadditive property sets determined by fuzzy implicators and examine properties
at any order in the fuzzy case. Like Zadeh’s expected certainty of the fuzzy rough approximation operators. In Section VI, by
that is constructed by a fuzzy belief structure and a measure using an arbitrary implicator, we define a measure of inclusion
of fuzzy inclusion, Ishizuka et al. [15] and Yager [46], respec- on fuzzy sets (Definition 11). We further define a dual pair of
tively, employed the Łukasiewicz implicator (fuzzy implication generalized fuzzy belief and plausibility functions constructed
operator) and the Kleene–Dienes implicator to define fuzzy in- by a fuzzy belief structure and the measure of fuzzy inclusion
clusions. As a result, they developed two types of fuzzy belief (Definition 12). We then introduce the main novel contributions
and plausibility functions. Analogous in spirit to conditional (Theorems 7, 8, and 10) of the present paper. In Theorem 7, we
probability, Ogawa and Fu [23] employed relative sigma count consider a fuzzy belief space arising from a pair of lower and
to compute the degree of inclusion for defining fuzzy belief upper fuzzy rough approximation operators and prove that as-
function. Although these extensions to the Dempster–Shafer sociated with every fuzzy belief space, it is possible to obtain a
theory of evidence arrive at frameworks in which both proba- fuzzy belief structure. In Theorem 8, by employing generalized
bilistic information and vague information can be handled, the fuzzy rough set theory, we show that for any belief structure,
universes of discourse are all finite. On the other hand, whether there must exist a countable approximation space and a proba-
or not the belief functions defined in [15], [23], and [46] satisfy bility measure imposed on it such that the probabilities of lower
the subadditivity property at any order have not been exam- and upper approximations of a fuzzy set with respect to the ap-
ined. Biacino [2] recently used the Kleene–Dienes implicator to proximation space are just the belief and plausibility measures
calculate the degree of fuzzy inclusion in infinite universe of dis- of the fuzzy set in the given belief structure. That is to say,
course. A fuzzy belief function on infinite universe of discourse for every belief structure, there exists a fuzzy belief space such
was then defined and the belief function was shown to satisfy that the arising belief structure coincides with the given one. In
the subadditivity property at any order. However, almost all of Theorem 10, we prove that every belief or plausibility function is
these extensions have not shown if the belief and plausibility infinitely monotone, i.e., fuzzy belief and plausibility functions,
functions can be represented as lower and upper probabilities respectively, satisfy the essential properties of subadditivity and
similar to the case of crisp sets. Biacino [2] considered fuzzy superadditivity at any order.
belief functions induced by an infinitely monotone inclusion
and she proved that they are lower probabilities. More recently, III. PRELIMINARIES
with reference to the R-implicator and S-implicator in fuzzy set
A. Fuzzy Logical Operators
theory, Chen et al. [3] explored two types of fuzzy belief and
plausibility functions, which are, respectively, the lower and up- A triangular norm, or t-norm in short, is an increasing, as-
per probabilities defined by a fuzzy belief space. They showed sociative, and commutative mapping T : [0, 1] × [0, 1] →
WU et al.: ON GENERALIZED FUZZY BELIEF FUNCTIONS IN INFINITE SPACES 387
[0, 1], satisfying the boundary condition: for all α ∈ [0, 1], of A and B as follows:
T (α, 1) = α.
A triangular conorm (t-conorm in short) is an increasing, (A ∩T B)(x) = T (A(x), B(x)), x∈X
associative, and commutative mapping S: [0, 1] × [0, 1] → (A ∪S B)(x) = S(A(x), B(x)), x∈X
[0, 1], satisfying the boundary condition: for all α ∈ [0, 1],
S(α, 0) = α. (A ⇒I B)(x) = I(A(x), B(x)), x ∈ X.
A negator N is a decreasing [0, 1] → [0, 1] mapping sat- An implicator I is called a border implicator (or it satisfies
isfying N (0) = 1 and N (1) = 0. The negator Ns (α) = 1 − α the neutrality principle [6]) iff for every α ∈ [0, 1], I(1, α) = α.
is usually referred to as the standard negator. A negator N is An implicator I is called an exchange principle (EP) impli-
said to be involutive if and only if (iff) N (N (α)) = α for all cator ([1], [31]) if it satisfies for all α, β, γ ∈ [0, 1]
α ∈ [0, 1]. It is well known that every involutive negator is
continuous [17]. I(α, I(β, γ)) = I(β, I(α, γ)). (2)
Given a negator N , a t-norm T and a t-conorm S are said be
An implicator I is called a confinement principle (CP) impli-
dual w.r.t. N iff the De Morgan’s laws are satisfied, i.e.
cator ([6]) if it satisfies for all α, β ∈ [0, 1]
S(N (α), N (β)) = N (T (α, β)) ∀α, β ∈ [0, 1] α ≤ β ⇐⇒ I(α, β) = 1. (3)
T (N (α), N (β)) = N (S(α, β)) ∀α, β ∈ [0, 1]. Several classes of implicators have been studied in the lit-
erature. We recall here the definitions of two main classes of
operators [6].
It is well known [17] that for an involutive negator N
Let T , S, and N be a t-norm, a t-conorm, and a negator,
and a t-conorm S, the function TS (α, β) = N (S(N (α),
respectively. An implicator I is called
N (β))), α, β ∈ [0, 1], is a t-norm such that TS and S are dual
1) an S-implicator based on S and N iff
w.r.t. N . In what follows, it will be referred to as a t-norm dual
to S w.r.t. N . I(α, β) = S(N (α), β) for all α, β ∈ [0, 1]. (4)
Let X be a nonempty set called the universe of discourse. By
a fuzzy set A in X, we mean a mapping A : X → [0, 1]. The 2) an R-implicator (residual implicator) based on a left-
class of all subsets of X (respectively, all fuzzy sets in X) will continuous t-norm T iff for all α, β ∈ [0, 1]
be denoted by P(X) [respectively, F(X)]. Zadeh’s fuzzy union I(α, β) = sup{λ ∈ [0, 1] : T (α, λ) ≤ β}. (5)
and fuzzy intersection will be denoted by ∪ and ∩, respectively,
and ∼N will be used to denote the fuzzy complement determined We will use θT to denote the R-implicator, i.e., θT (α, β) =
by a negator N , i.e., for every A ∈ F(X) and every x ∈ X, sup{λ ∈ [0, 1] : T (α, λ) ≤ β}.
(∼N A)(x) = N (A(x)). If N = Ns , we then write ∼ A instead From a t-conorm S, we define a binary operation σS on [0, 1]
of ∼N A. as follows: for all α, β ∈ [0, 1]
By an implicator (fuzzy implication operator), we mean a
function I: [0, 1] × [0, 1] → [0, 1] satisfying I(1, 0) = 0 σS (α, β) = inf{λ ∈ [0, 1] : S(α, λ) ≥ β}. (6)
and I(1, 1) = I(0, 1) = I(0, 0) = 1. An implicator I is called If S is the dual t-conorm of a t-norm T w.r.t. Ns , then σS is
left monotonic (respectively, right monotonic) iff for every α ∈ dual to θT w.r.t. Ns [21], i.e., for all α, β ∈ [0, 1]
[0, 1], I(·, α) is decreasing (respectively, I(α, ·) is increasing).
If I is both left monotonic and right monotonic, then it is called σS (1 − α, 1 − β) = 1 − θT (α, β). (7)
hybrid monotonic. I is semicontinuous if
Theorem 1 [1], [28]: Every S-implicator and every R-

implicator is a hybrid monotonic, border, and EP implicator.
I aj , bk = I(aj , bk ) (1) Every R-implicator is a CP implicator.
j k j,k
B. Fuzzy Relations
for all indexed families {aj : j ∈ J} and {bk : k ∈ K} of real Let X be a nonempty universe of discourse. For any fuzzy set
numbers in [0, 1]. A ∈ F(X), the α-level set of A, denoted by Aα , is Aα = {x ∈
It is easy to see that I(α, 1) = 1 for all α ∈ [0, 1], when I X : A(x) ≥ α}, where α ∈ [0, 1]. In what follows, α denotes
is a left monotonic implicator, and if I is right monotonic, then (x) = α for all x ∈ X.
the constant fuzzy set: α
I(0, α) = 1 for all α ∈ [0, 1]. Definition 1: Let U and W be two nonempty universes of
For a left monotonic implicator I, the function N (α) = discourse. A subset R ∈ P(U × W ) is referred to as a (crisp)
I(α, 0), α ∈ [0, 1], is a negator, called a negator induced binary relation from U to W . The relation R is serial if for each
by I. For example, the Łukasiewicz implicator IL (α, β) = x ∈ U , there exists a y ∈ W such that (x, y) ∈ R. If U = W ,
min{1, 1 − α + β} induces the standard negator Ns . then R is called binary relation on U . A fuzzy subset R ∈
Given a t-norm T , a t-conorm S, an implicator I, and two F(U × W ) is referred to as a fuzzy binary relation from U to
fuzzy sets A and B of a set X, we, respectively, define the W , R(x, y) is the degree of relation between x and y, where
corresponding generalized union, intersection, and implication (x, y) ∈ U × W . If for each x ∈ U , there exists a y ∈ W such
that R(x, y) = 1, then R is a serial fuzzy relation from U to W . Theorem 2: The fuzzy probability measure P in Definition 3
If U = W , then R is called a fuzzy relation on U . satisfies the following properties.
For a fuzzy relation R ∈ F(U × W ), we can define a 1) P (A) ∈ [0, 1] and P (A) + P (∼ A) = 1 for all A ∈
fuzzy-set-valued mapping F : U → F(W ) by F (u)(w) = F(U, A).
R(u, w), (u, w) ∈ U × W. Conversely, for a given fuzzy-set- 2) P is countably additive, i.e., for Ai ∈ F(U, A), i =
valued mapping F : U → F(W ), we can derive a fuzzy rela- 1, 2, . . . , Ai ∩ Aj = ∅ ∀i = j, then
tion R from U to W by setting R(u, w) = F (u)(w), (u, w) ∈ ∞ ∞
U × W . In such a case, F can be viewed as a fuzzy compatibility P Ai = P (Ai ). (10)
relation between U and W . Similar to the crisp case [52], the im- i=1 i=1
age F (u) of an element u in U under the fuzzy-set-valued map-
ping is called the fuzzy granule of u. It can be easily observed 3) A, B ∈ F(U, A), A ⊆ B =⇒ P (A) ≤ P (B).
that the fuzzy relation R is serial if and only if F : U → F(W ) 4) If U = {ui : i = 1, 2, . . .} is an infinite countable set and
is a normal fuzzy-set-valued mapping, i.e., for each u ∈ U , A = P(U ), then for all A ∈ F(U )
there exists w ∈ W such that F (u)(w) = 1. For a given fuzzy- 1
set-valued mapping F : U → F(W ) (or a fuzzy binary relation P (A) = P (Aα )dα = A(x)P (x). (11)
0
from U to W ), we define an operator j : F(W ) → P(U ) x∈U
j(A) = {u ∈ U : F (u) = A}, A ∈ F(W ). (8) 5) If U is a finite set with |U | = n (where |U | is the cardinality
of the set U ), A = P(U ), and P (u) = 1/n, then P (A) =
1
It can be easily shown that j satisfies the following.
0 P (Aα )dα = |A|/n for all A ∈ P(U ).
J1) A = B =⇒ j(A) ∩ j(B) = ∅. Proof: Properties 1)–3) are given in [3].
4) For any A ∈ F(U ), since U = {ui : i = 1, 2, . . .} is count-
J2) j(A) = U. able, the valued set {A(x) : x ∈ U } is also countable. Without
A ∈F(W ) loss of generality, we assume that {A(x) : x ∈ U } = {αi : i =
1, 2, . . .} such that
IV. PROBABILITIES OF FUZZY SETS
0 ≤ α1 < α2 < · · · < αk < · · · ≤ 1.
The notion of a fuzzy event and its probability are first defined
by Zadeh [57]. A fuzzy event is a fuzzy set whose membership Denote
function is Borel measurable. Its probability of occurrence is de- ∞

fined by the expectation of the membership function. Recently, α∞ = αk .
Chen et al. [3] defined probabilities of fuzzy sets in infinite k =1
universes of discourse. In this section, we present the formal Obviously, α∞ ≤ 1. Denote
definition of probabilities of fuzzy sets in infinite universes of
discourse and give some basic properties of the fuzzy probability Aα j = = {x ∈ U : A(x) = αj }.
measure. Let P (Aα j = ) = pj , then
Throughout this section, we always assume that (U, A, P ) is
a probability space, i.e., U is a nonempty set and may not be ∞

finite, A ⊆ P(U ) is a σ-algebra on U , and P is a probability Aα j = {x ∈ U : A(x) ≥ αj } = Aα k = (12)
measure on U . The pair (U, A) is called a measurable space. k =j
Definition 2: A fuzzy set A ∈ F(U ) is said to be measurable and by (10), P is countably additive. Hence
w.r.t. (U, A) if A : U → [0, 1] is a measurable function w.r.t. ∞ ∞ ∞
A − B([0, 1]), where B([0, 1]) is the family of Borel sets on
P (Aα j ) = P Aα k = = P (Aα k = ) = pk . (13)
[0, 1]. We denote by F(U, A) the family of all measurable fuzzy
k =j k =j k =j
sets of U w.r.t. A − B([0, 1]).
For any measurable fuzzy set A ∈ F(U, A), Aα , clearly Aα ∈ For any α ∈ (0, 1], if αk < α ≤ αk +1 , then
A for all α ∈ [0, 1], is a measurable set on the probability
Aα = {x ∈ U : A(x) ≥ α}
space (U, A, P ) and P (Aα ) ∈ [0, 1]. Since f (α) = P (Aα ) is
∞

monotone decreasing and left continuous, f (α) is integrable,
1 = {x ∈ U : A(x) = αj }
we denote the integrand as 0 P (Aα )dα.
j =k +1
Definition 3: If a fuzzy set A is measurable w.r.t. (U, A) and
P is a probability measure on (U, A). Denote ∞

1 = Aα j =
j =k +1
P (A) = P (Aα )dα (9)
0
and hence
where P (A) is called the probability of A. ∞
∞

For a singleton set {x}, we will write P (x) instead of P ({x}) P (Aα ) = P (Aα j =) = pj = P (Aα k + 1). (14)
for short. j =k +1 j =k +1
It follows that R = {(x, y) ∈ U × W : y ∈ F (x)}. The triple (U, W, R) is re-

αk + 1 αk + 1 ferred to as a generalized crisp approximation space. For any
P (Aα )dα = P (Aα k + 1 )dα set A ∈ P(W ), the lower and upper approximations, denoted as
αk αk
R(A) and R(A), w.r.t. (U, W, R) are defined, respectively, by
= P (Aα k + 1 )(αk +1 − αk ). (15)
R(A) ={x ∈ U : Rs (x) ⊆ A}
If α1 = 0, then of course
α1 R(A) ={x ∈ U : Rs (x) ∩ A = ∅}. (20)
P (Aα )dα = 0. The pair (R(A), R(A)) is referred to as a generalized crisp rough
0
set, and R and R : P(W ) → P(U ) are referred to as lower and
If α1 > 0, when α ∈ (0, α1 ], it is clear that upper generalized crisp approximation operators, respectively.
Aα = {x ∈ U : A(x) ≥ α1 } = U From the definition, many interesting properties of the lower
and upper approximation operators can be derived (see, e.g.,
then P (Aα ) = P (U ) = 1. Consequently [25], and [48]–[50]).
α1
P (Aα )dα = α1 P (U ) = α1 = α1 P (Aα 1 ). (16)
0 B. Generalized Fuzzy Rough Sets Determined
by Fuzzy Implicators
If α∞ < 1, then for any α ∈ (α∞ , 1], we have Aα = {x ∈ U :
A(x) ≥ α} = ∅. Hence, P (Aα ) = 0, and as a result Definition 5: Let U and W be two nonempty universes of dis-
1 course and R a fuzzy relation from U to W . The triple (U, W, R)
P (Aα )dα = 0. (17) is called a fuzzy approximation space. When U = W and R is a
α∞ fuzzy relation on U , we also call (U, R) a fuzzy approximation
If α∞ = 1, then the aforementioned equation obviously holds. space.
Definition 6: Let (U, W, R) be a fuzzy approximation space, I
Since the integral function is a measure and the series

x∈U A(x)P (x) is absolutely convergent, by employing (15)
an implicator on [0, 1], and N a negator on [0, 1]. For a fuzzy set
and (16), we can prove that A ∈ F(W ), the I-lower approximation of A w.r.t. (U, W, R),
∞ αk α∞ denoted as RI (A), is a fuzzy set of U whose membership func-
αj pj = lim P (Aα )dα = P (Aα )dα. (18) tion is defined by
k →∞ 0 0
j =1
RI (A)(x) = I(R(x, y), A(y)), x ∈ U. (21)
In terms of (17) and (18), we conclude that y ∈W
1 α∞ 1
The I-upper approximation of A w.r.t. (U, W, R), denoted as
P (Aα )dα = P (Aα )dα + P (Aα )dα
0 0 α∞
RI (A), is defined by
∞
RI (A) = ∼N RI (∼N A). (22)
= αj pj = A(x)P (x).
j =1 x∈U
The operators RI and RI from F(W ) to F(U ) are referred
to as I-upper and I-lower fuzzy rough approximation opera-
5) It follows immediately from 4). tors of (U, W, R), respectively, and the pair (RI (A), RI (A)) is
It should be noted that (11) is very important and is used called the I-fuzzy rough set of A w.r.t. (U, W, R). Specially,
throughout the rest of the paper. It also implies that Zadeh’s if I is an S-implicator based on a t-conorm S and an involu-
probability measures [57] of fuzzy events can be extended to tive negator N , and T and S are dual w.r.t. N , then we call
infinite countable universes of discourse. (RI (A), RI (A)) the S-fuzzy rough set of A w.r.t. (U, W, R).
V. ROUGH SETS ON TWO UNIVERSES OF DISCOURSE If I is an R-implicator based on a t-norm T , then we call
(RI (A), RI (A)) the R-fuzzy rough set of A w.r.t. (U, W, R).
In this section, we introduce concepts related to rough sets on For simplicity, in the discussion to follow, we always take
two universes of discourse in both the crisp and fuzzy environ- N = Ns .
ments, and examine their basic properties. Theorem 3: Let (U, W, R) be a fuzzy approximation space
and I a semicontinuous, hybrid monotonic implicator on
A. Generalized Crisp Rough Sets on Two
[0, 1]. Then, the I-fuzzy rough approximation operators RI
Universes of Discourse
and RI in Definition 6 have the following properties: for all
Definition 4: Let U and W be two nonempty universes of A, B ∈ F(W ), Aj ∈ F(W )(∀j ∈ J, J is an index set) and the
discourse. Suppose that R is an arbitrary crisp relation from U following hold.
to W . We define a set-valued function Rs : U → P(W ) by
FL1) RI (W ) = U.
Rs (x) = {y ∈ W : (x, y) ∈ R}, x∈U (19)
FU1) RI (∅) = ∅.
where Rs (x) is referred to as the successor neighborhood
of x w.r.t. R. Obviously, any set-valued function F from FL2) RI Aj = RI (Aj ).
U to W defines a binary relation from U to W by setting j ∈J j ∈J
FU0) Combining FL0) and (22), we have

FU2) RI Aj = RI (Aj ).
j ∈J j ∈J RI ( ) =∼ RI (1
α) =∼ RI (∼ α − α)
FL3) A ⊆ B =⇒ RI (A) ⊆ RI (B). =∼ 1
− α =∼∼ α
=α
.
FU3) A ⊆ B =⇒ RI (A) ⊆ RI (B). FL0) By taking α = 0 in FL0).
FU0) By taking α = 1 in FU0).
FL4) RI Aj ⊇ RI (Aj ).
j ∈J j ∈J VI. FUZZY BELIEF STRUCTURES AND FUZZY
BELIEF FUNCTIONS
FU4) RI Aj ⊆ RI (Aj ).
In this section, we first review the main results of the original
j ∈J j ∈J
Dempster–Shafer theory of evidence with its existing extension
Proof: The proofs of FL1)–FL4) are given in [42]. Properties in the fuzzy environment. We then extend Dempster–Shafer
FU1)–FU4) can be deduced from (22) and properties FL1)– theory of evidence to fuzzy sets in infinite universes of discourse
FL4), respectively. by employing arbitrary fuzzy implication operators.
Theorem 4: Let (U, W, R) be a fuzzy approximation space
and I a semicontinuous border implicator. If R is a serial fuzzy A. Belief Structures and Belief Functions in Crisp Environment
relation from U to W , then the I-fuzzy rough approximation in Finite Spaces
operators in Definition 6 satisfy the following properties. The original Dempster–Shafer theory of evidence is based on
a set function called the basic probability assignment on a finite
FLU0) RI (A) ⊆ RI (A)∀A ∈ F(W ).
universe of discourse.
FL0) RI ( ∀α ∈ [0, 1].
α) = α Definition 7 [32]: Let W be a nonempty finite set. A set
function m : P(W ) → [0, 1] is referred to as a basic probability
FU0) RI ( ∀α ∈ [0, 1].
α) = α
assignment on W if it satisfies the following axioms:
FL0) RI (∅) = ∅.
M1) m(∅) = 0 M2) m(X) = 1.
FU0) RI (W ) = U. X ⊆W
The value m(X) represents the degree of belief that a specific

Proof: FLU0) Let A ∈ F(W ), for any x ∈ U , since R is a
element of W belongs to set X, but not to any particular subset
serial fuzzy relation from U to W , there exists y0 ∈ W such that
of X. It is a relative level of confidence in X. It reflects the
R(x, y0 ) = 1. It should be noted that I is a border implicator.
probability that this information is properly and completely de-
Then, we have
scribed by the set X [12]. A set X ∈ P(W ) with m(X) > 0 is
RI (A)(x) + RI (∼ A)(x) referred to as a focal element of m. We denote by M the family
of all focal elements of m. The pair (M, m) is called a belief
= I(R(x, y), A(y)) + I(R(x, y), 1 − A(y)) structure [13], [46] or a body of evidence [16]. Obviously, m
y ∈W y ∈W defines a probability measure on P(W ), but not on W . Associ-
≤ I(R(x, y0 ), A(y0 )) + I(R(x, y0 ), 1 − A(y0 )) ated with each belief structure, a pair of belief and plausibility
functions can be derived [32].
= I(1, A(y0 )) + I(1, 1 − A(y0 )) Definition 8: Let (M, m) be a belief structure on W . A set
= A(y0 ) + 1 − A(y0 ) = 1. function Bel : P(W ) → [0, 1] is referred to as a belief function
on W if
It follows that
Bel(X) = m(A) ∀X ∈ P(W ). (23)
{A ∈P(W ):A ⊆X }
RI (A)(x) ≤ 1 − RI (∼ A)(x) = RI (A)(x).
A set function Pl : P(W ) → [0, 1] is referred to as a plausibility
Consequently, FLU0) holds. function on W if
FL0) Let α ∈ [0, 1]. Since I is a semicontinuous border
implicator, for any x ∈ U , we have Pl(X) = m(A) ∀X ∈ P(W ). (24)
{A ∈P(W ):A ∩X = ∅}

RI (
α)(x) = I(R(x, y), α) Bel(X) reflects the weight of evidence, which focuses on
y ∈W the subsets of X. Belief and plausibility functions based on the

same belief structure are connected by the dual property
=I R(x, y), α
y ∈W Pl(X) = 1 − Bel(∼ X). (25)
= I(1, α) = α = α
(x). Thus, Pl(X) reflects the weight of evidence, which does not
focus on ∼ X, the opposite of X. Alternatively, Pl(X) quan-
Thus, FL0) holds. tifies the amount of our belief that might be allocated to X.
Moreover, Bel(X) + Bel(∼ X) ≤ 1 and Bel(X) ≤ Pl(X) for and plausibility functions derived by (M, m), then there exists
all X ∈ P(W ). a crisp belief space ((U, P ), W, R) (i.e., there exists a finite set
The belief function Bel on a finite set W can be equivalently U , a probability measure P on U , and a serial binary relation R
defined as a set function from P(W ) to [0, 1] satisfying the from U to W ) such that
following properties [32].
Bel(X) = P (R(X)), X ∈ P(W )
MC1) Bel(∅) = 0.
MC2) Bel(W ) = 1. Pl(X) = P (R(X)), X ∈ P(W )
MC3) Bel satisfies subadditive property, i.e., for any n ∈ N
m(X) = P (j(X)), X ∈ P(W ) (27)
(N is the set of positive integer numbers) and Xi ∈
P(W ), i = 1, 2, . . . , n where j(X) = {u ∈ U : Rs (u) = X}.
n
|I |+1
Bel Xi ≥ (−1) Bel Xi . B. Belief Structures and Belief Functions
i=1 ∅= I ⊆{1,2,...,n } i∈I in the Fuzzy Environment
Likewise, Pl can be equivalently defined as a set function from Based on the concept of information granularity and the
P(W ) to [0, 1] satisfying the following properties. theory of possibility [55], [56], Zadeh first generalized the
AC1) Pl(∅) = 0. Dempster–Shafer theory to fuzzy sets. First of all, the belief
AC2) Pl(W ) = 1. structure should be generalized to the fuzzy environment.
AC3) Pl satisfies superadditive property, i.e., for any n ∈ N Definition 10: Let W be a nonempty set that may be infinite.
and Xi ∈ P(W ), i = 1, 2, . . . , n A set function m : F(W ) → [0, 1] is referred to as a fuzzy basic
n probability assignment if it satisfies the following axioms.
Pl Xi ≤ (−1)|I |+1 Pl Xi . FM1) m(∅) = 0.
i=1 ∅= I ⊆{1,2,...,n } i∈I
FM2) m(X) = 1.
The corresponding basic probability assignment m is linked
X ∈F(W )
with Bel and Pl as follows:
A fuzzy set X ∈ F(W ) with m(X) > 0 is referred to as a
m(A) = (−1)|A −B | Bel(B), A ∈ P(W )
B ⊆A
focal element of m. We denote by M the family of all focal ele-
ments of m. The pair (M, m) is called a fuzzy belief structure.
Pl(A) = 1 − Bel(∼ A), A ∈ P(W ). In the discussion to follow, all the focal elements are supposed
In the literature, a set function satisfying properties MC1)– to be normal, i.e., for any A ∈ M, there exists an x ∈ W such
MC3) is called a monotone Choquet capacity of infinite order, that A(x) = 1.
and it is said to be an alternating Choquet capacity of infinite We recall that a crisp belief function Bel : P(W ) → [0, 1]
order if it satisfies properties AC1)–AC3) [4]. Thus, MC1)– defined from a crisp belief structure can be represented as
MC3) and AC1)–AC3) are, respectively, the essential properties
Bel(X) = m(A) = m(A)I(A, X) (28)
of the belief and plausibility functions. A ⊆X A ∈P(W )
Definition 9: If U and W are two finite sets, R is a serial
binary relation from U to W , and P is a probability measure on where I(A, X) has to be understood as the inclusion degree of
U , then we call ((U, P ), W, R) a crisp belief space. A in X, which is 1 if A ⊆ X, and 0 otherwise. Based on a fuzzy
The original concepts of belief and plausibility functions in belief structure (M, m) on a finite universe of discourse W ,
the Dempster–Shafer theory of evidence come from the lower Zadeh defined the expected certainty, denoted by EC(X), and
and upper probabilities induced by a multivalued mapping [32]. the expected possibility, denoted by EΠ(X), as a generalization
The following theorem shows that any belief structure with its of Dempster–Shafer belief and plausibility functions: for all
inducing belief and plausibility functions can be represented as X ∈ F(W )
probabilities of lower and upper approximations derived from
an approximation space in rough set theory. EC(X) = inf(A ⇒ X) = 1 − EΠ(∼ X) (29)
A ∈M
Theorem 5 [43], [51]: Assume that ((U, P ), W, R) is a crisp
belief space, and R and R are the lower and upper approximation EΠ(X) = m(A) sup(X ∩ A) (30)
operators defined by (20). Denote A ∈M
Bel(X) = P (R(X)), X ∈ P(W ) where inf(A ⇒ X) measures the degree to which A is included
in X and sup(X ∩ A) measures the degree that X intersects
Pl(X) = P (R(X)), X ∈ P(W ). (26)
with A. It is easy to verify that the expected certainty and the
Then, Bel, Pl : P(W ) → [0, 1] are a dual pair of belief and expected possibility degenerate into the crisp belief and plau-
plausibility functions on W , and the basic probability as- sibility functions when the belief structure (M, m) and X are
signment is defined by m(X) = P (j(X)), X ∈ P(W ), where crisp. In order to define the belief degree of the elements of
j(X) = {u ∈ U : Rs (u) = X}. Conversely, if (M, m) is a be- F(W ), (28) is very useful because it is enough to substitute
lief structure on W and Bel, Pl : P(W ) → [0, 1] are the belief in it a suitable extension of the inclusion degree I(A, X) to
F(W ) × F(W ). So, the crucial point of the problem becomes is an extended necessity measure on W in the sense
the search for an extension of the inclusion relation to F(W ), of Abdel–Hamid and Morsi [1], i.e., it satisfies axioms
which allows us to express within the fuzzy framework the well- ENM1) and
ENM2).
known theory of belief measures [2]. ENM1) N ( Bj ) = N (Bj ), for every index family
Following Zadeh’s work, Ishizuka et al. [15], Ogawa and j ∈J j ∈J
Fu [23], Yager [46], and recently Biacino [2] have extended the {Bj : j ∈ J} ⊆ F(W ).
Dempster–Shafer theory to fuzzy sets by defining a measure of
inclusion I(A ⊆ X), the degree to which a set A is included in ENM2) N ( α) = α for all α ∈ [0, 1].
the set X by using the following formula, which is similar to Proof: 1) and 2) can be concluded immediately by the hybrid
Zadeh’s expected certainty EC(X) monotonicity of I.
3) Recall from [42] that I is a CP implicator iff for all (a, b) ∈
Bel(X) = m(A)I(A ⊆ X), X ∈ F(W ). (31) [0, 1] × [0, 1]
A ∈M
a ≤ b ⇐⇒ I(a, b) = 1.
Ishizuka et al. [15] and Yager [46], respectively, used the
Łukasiewicz implicator IL (x, y) = min{1, 1 − x + y} and the Then

Kleene–Dienes implicator IKD (x, y) = max{1 − x, y} (both I(A ⊆ B) = I(A(x), B(x)) = 1
are S-implicators) in fuzzy set theory to define the inclu- x∈W
sion measure. Analogous to conditional probability, Ogawa and
⇐⇒ I(A(x), B(x)) = 1, ∀x ∈ W
Fu [23] used relative sigma count to compute the degree of inclu-
sion. Although these extensions to the Dempster–Shafer theory ⇐⇒ A(x) ≤ B(x), ∀x ∈ W
of evidence arrive at frameworks within which both probabilis-
⇐⇒ A ⊆ B.
tic information and vague information can be managed, the
universes of discourse are all finite. Moreover, we are not sure 4) is obtained in [1] and [10].
whether or not these belief and plausibility functions satisfy In the discussion to follow, we always assume that I is a
axioms similar to MC3) and AC3) in the fuzzy environment. hybrid monotonic, semicontinuous, and border implicator on
Biacino [2] recently used the Kleene–Dienes implicator to cal- [0, 1], and we take N = Ns .
culate the degree of inclusion in an infinite universe of discourse. Lemma 1 [2]: Let (M, m) be a fuzzy belief structure on W .
As a result, a fuzzy belief function on the infinite universe of Then, the focal elements of m constitute a countable set.
discourse was defined. However, all these extensions have not Lemma 1 is quite important and instrumental to achieve the
shown whether or not the belief and plausibility functions can main results of this paper. It is possible for us to define belief
be presented as lower and upper probabilities similar to the case and plausibility functions on an arbitrary infinite universe of
of crisp sets. discourse.
By using a general type of implicator, we will define a gener- Definition 12: Let W be a nonempty universe of discourse
alized fuzzy belief structure and its induced dual pair of fuzzy that may be infinite, (M, m) a fuzzy belief structure on W , and
belief and plausibility functions in an infinite universe of dis- I an implicator on [0, 1]. A fuzzy set function Bel : F(W ) →
course. We first introduce a general type of inclusion degree in [0, 1] is referred to as a generalized fuzzy belief function on W
fuzzy sets. if for all X ∈ F(W )
Definition 11: Let W be a nonempty universe of discourse that
may be infinite and I an implicator on [0, 1], for A, B ∈ F(W ). Bel(X) = m(A)I(A ⊆ X)
We define A ∈F(W )

I(A ⊆ B) = I(A(x), B(x)) = m(A) I(A(x), X(x)). (34)
x∈W A ∈F(W ) x∈W

= (A ⇒I B)(x). (32) The fuzzy set function Pl : F(W ) → [0, 1] is referred to as a
x∈W generalized fuzzy plausibility function on W
Theorem 6: If I is a hybrid monotonic, semicontinuous, Pl(X) = 1 − Bel(∼ X), X ∈ F(W ). (35)
and border implicator on [0, 1], and A, B, C ∈ F(W ), then the
following hold. It is worthy of note that, by Lemma 1, the not null elements
1) B ⊆ C =⇒ I(A ⊆ B) ≤ I(A ⊆ C). that appear under the sign of sum are, for every fixed X ∈
2) B ⊆ C =⇒ I(C ⊆ A) ≤ I(B ⊆ A). F(W ), finite or countable, and since I(A ⊆ X) ≤ 1 for every
3) If I is a CP implicator, then A and X ∈ F(U ), the series that defines Bel(X) in (34) is
convergent. So, Definition 12 is reasonable.
I(A ⊆ B) = 1 ⇐⇒ A ⊆ B. By taking different implicators and belief structures, we can
obtain various classes of dual pairs of fuzzy belief and plausi-
4) If A is a normal fuzzy set, then the function NI,A :
bility functions.
F(W ) → [0, 1] defined by
Example 1: When I is an S-implicator based on a t-conorm S
NI,A (B) = I(A ⊆ B), B ∈ F(W ) (33) and T is a t-norm dual to S w.r.t. Ns , it can be checked that [46]
for X ∈ F(W ) It can be easily checked that j satisfies properties J1) and J2).

Bel(X) = m(A) S(1 − A(x), X(x)) (36) J1) A = B =⇒ j(A) ∩ j(B) = ∅.

A ∈F(W ) x∈W
J2) j(A) = U.

A ∈F(W )
Pl(X) = m(A) T (A(x), X(x)). (37)
A ∈F(W ) x∈W Since R is serial, we can observe that j(∅) = ∅. Consequently,
Example 2: When I = θT is an R-implicator determined by P (j(∅)) = P (∅) = 0 and

a t-norm T and its dual t-conorm S w.r.t. Ns , then it can be
checked that for X ∈ F(W ) P (j(A)) = P j(A) = P (U ) = 1.
A ∈F(W ) A ∈F(W )

Bel(X) = m(A) θT (A(x), X(x)) (38)
Define a fuzzy set function m : F(W ) → [0, 1] as follows:
A ∈F(W ) x∈W
m(A) = P (j(A)), A ∈ F(W ).
Pl(X) = m(A) σS (A(x), X(x)). (39)
A ∈F(W ) x∈W Obviously, m(∅) = 0 and

Example 3: When W is a finite universe of discourse, (M, m) m(A) = P (j(A))
is a crisp belief structure, and I is the Kleene–Dienes implicator, A ∈F(W ) A ∈F(W )
then [9], [11], [36] for X ∈ F(W )
=P j(A) = P (U ) = 1.
Bel(X) = m(A) X(x) (40) A ∈F(W )
{A :A ∈M} x∈W
Hence, m is a fuzzy basic probability assignment on W . And
Pl(X) = m(A) X(x). (41) for any X ∈ F(W ), by (11), J1), and J2), we have
{A :A ∈M} x∈W
Bel(X) = P (RI (X)) = RI (X)(x)P (x)
Example 4: When W is a finite universe of discourse and x∈U
(M, m) is a crisp belief structure, then
= I(R(x, y), X(y))P (x)
Bel(X) = m(A), X ∈ P(W ) (42) x∈U y ∈W
{A ∈M:A ⊆X }

= I(R(x, y), X(y))P (x)

Pl(X) = m(A), X ∈ P(W ). (43) A ∈F(W ) x∈j (A ) y ∈W
{A ∈M:A ∩X = ∅}

= I(A(y), X(y))P (j(A))
Definition 13: If U is a countable set, P is a probability A ∈F(W ) y ∈W
measure on U , W is a nonempty set that may be infinite, and
R is a serial fuzzy binary relation from U to W , then we call = m(A) I(A(y), X(y)) .
((U, P ), W, R) a fuzzy belief space. A ∈F(W ) y ∈W
Theorem 7 next shows that any fuzzy belief space can be asso-
Therefore, we have proved that Bel is a fuzzy belief function.
ciated with a fuzzy belief structure such that the probabilities of
By the duality, we can conclude that Pl is a fuzzy plausibility
the lower and upper fuzzy rough approximations induced from
function dual to Bel.
the fuzzy belief space produce, respectively, the corresponding
Theorem 7 provides us a potential tool for the reasoning and
fuzzy belief and plausibility functions derived from the fuzzy
knowledge acquisition in fuzzy information systems.
belief structure.
Example 5: Given a fuzzy information system (U, AT, F ),
Theorem 7: Let ((U, P ), W, R) be a fuzzy belief space in
where U = {u1 , u2 , . . . , un } is the object set, AT =
which U is a countable set and I an implicator. If RI and RI
{a1 , a2 , . . . , am } is the attribute set, F = {fa : a ∈ AT } is a
are the I-fuzzy rough approximation operators in Definition 6,
set of information functions such that fa : U → F(Va ), and Va
denote
is called the domain of attribute a. Let P be a probability mea-
Bel(X) = P (RI (X)), X ∈ F(W ) sure on U (if there is no probability measure on U , we can
take P (X) = |X|/|U | = |X|/n). Each attribute set B ⊆ AT
Pl(X) = P (RI (X)), X ∈ F(W ). (44) corresponds to a fuzzy binary relation RB ∈ F(U × U ) on
the object set. Such binary relation is often reflexive (see,
Then, Bel : F(W ) → [0, 1] and Pl : F(W ) → [0, 1] are a dual
e.g., [14], [40], and [44]), and therefore, each attribute set B
pair of fuzzy belief and plausibility functions.
can be associated with a belief structure (MB , mB ) on U , and
Proof: Let
mB (X) = P (jB (X)) for all X ∈ F(U ), where
j(A) = {x ∈ U : R(x, y) = A(y) ∀y ∈ W }, A ∈ F(W ). jB (X) = {x ∈ U : RB (x, y) = X(y), ∀y ∈ U }.
If we write RB (x, y) = RB (x)(y), then jB can be equivalently From R, we can obtain a mapping j : F(W ) → P(U ) as
written as follows:
jB (X) = {x ∈ U : RB (x) = X}. j(A) = {u ∈ U : R(u, w) = A(w) ∀w ∈ W }, A ∈ F(W ).
Obviously, RB (x) is a normal fuzzy subset of U for all x It is easy to see that j(A) = {ui } for A = Ai , and ∅ otherwise.
(RB (x)(x) = 1). Thus, each focal element of mB is nor- Consequently, m(A) = P (j(A)) > 0 for A ∈ M, and 0 other-
mal. Moreover, if B = B1 ∪ B2 , then RB = RB 1 ∩ RB 2 and wise. It should be noted that j(A) ∩ j(B) = ∅ for A = B and

mB i (X) = P (jB i (X)) for all X ∈ F(U ), i = 1, 2. It should A ∈F(W ) j(A) = U. Then, by Theorem 2, we can conclude that
be noted that RB (x) = RB 1 (x) ∩ RB 2 (x). Then for any X ∈ F(W )

mB (X) = mB 1 ⊕ mB 2 (X) P (RI (X)) = RI (X)(x)P (x)
= P ({x ∈ U : RB 1 (x) ∩ RB 2 (x) = X}). x∈U

Thus, the combination of two fuzzy basic probability assign- = I(R(x, y), X(y))P (x)
ments (which is one of the main issues in the study of the x∈U y ∈W
Dempster–Shafer theory of evidence) can be transformed into

= I(R(x, y), X(y))P (x)
the calculation of the intersection of the corresponding fuzzy
A ∈F(W ) x∈j (A ) y ∈W
binary relations. On the other hand, just as crisp belief and plau-
sibility functions can be used to study knowledge reduction in = I(A(y), X(y))P (j(A))
complete and incomplete information systems [19], [41], [45], A ∈F(W ) y ∈W
[58], we believe that the Dempster–Shafer theory of evidence

for fuzzy sets might be used to analyze knowledge reduction in = m(A) I(A(y), X(y))
fuzzy information systems and fuzzy decision systems. A ∈F(W ) y ∈W
Theorem 8 next shows that for any fuzzy belief structure,
there exists a fuzzy belief space such that fuzzy belief and = Bel(X).
plausibility functions induced from the fuzzy belief structure By the duality property, we have, on the other hand
can be represented by the probabilities of lower and upper fuzzy
rough approximations determined by the fuzzy belief space. Pl(X) = 1 − Bel(∼ X) = 1 − P (RI (∼ X))
Theorem 8: Let (M, m) be a fuzzy belief structure on W
= 1 − P (∼ RI (X)) = P (RI (X)).
and I an implicator on [0, 1]. If Bel : F(W ) → [0, 1] and Pl :
F(W ) → [0, 1] are the dual pair of fuzzy belief and plausibility Similar to Definitions 10 and 13, the definitions of belief struc-
functions defined in Definition 12, then there exists a countable ture and belief space in the crisp environment can be extended
set U , a serial fuzzy relation R from U to W , and a probability to infinite spaces (we omit the definitions here). By regarding a
measure P on U such that for all X ∈ F(W ) crisp set as special fuzzy set, according to Theorems 7 and 8,
we can arrive at Corollary 1, which is a generalization to infinite
Bel(X) = P (RI (X)) = RI (X)(x)P (x) (45) cases of Theorem 5.
x∈U
Corollary 1: Assume that ((U, P ), W, R) is a crisp belief

Pl(X) = P (RI (X)) = RI (X)(x)P (x). (46) space in which U is a countable set and W may be infinite,
x∈U and R and R are the lower and upper approximation operators

defined by (20). Denote
Proof: Since A ∈F(W ) m(A) = 1, by Lemma 1, we know
that the focal elements of m constitute a countable set. Without Bel(X) = P (R(X)), X ∈ P(W )
loss of generality, we assume that M has infinite countable
Pl(X) = P (R(X)), X ∈ P(W ). (47)
elements and we denote
Then, Bel, Pl : P(W ) → [0, 1] are a dual pair of belief and plau-
M = {Ai ∈ F(W ) : i ∈ N}
sibility functions on W , and the basic probability assignment

where i∈N m(Ai ) = 1. Let U = {ui : i ∈ N} be a set hav- is defined by m(X) = P (j(X)), X ∈ P(W ), where j(X) =
ing infinite countable elements. We define a set function P : {u ∈ U : Rs (u) = X}. Conversely, if (M, m) is a belief struc-
P(U ) → [0, 1] as follows: ture on W in which W may be infinite, and Bel, Pl : P(W ) →
[0, 1] the belief and plausibility functions derived by (M, m),
P ({ui }) = m(Ai ), i∈N then there exists a crisp belief space ((U, P ), W, R) (i.e., there

P (X) = P ({u}), X ∈ P(U ). exists a countable set U , a probability measure P on U , and a
u ∈X
serial binary relation R from U to W ) such that (47) holds and
the corresponding basic probability assignment is defined as
Obviously, P is a probability measure on U .
We further define a fuzzy relation R from U to W as follows: m(X) = P (j(X)), X ∈ P(W )
R(ui , w) = Ai (w), i ∈ N, w ∈ W. where j(X) = {u ∈ U : Rs (u) = X}.
Theorem 9: Let (M, m) be a fuzzy belief structure on W , I Proof: By Theorem 8, there exists a countable set U , a serial
a border implicator on [0, 1]. Then fuzzy relation R from U to W , and a probability measure P on
1) Bel(X) ≤ Pl(X) ∀X ∈ F(W ). U such that for all X ∈ F(W )
2) Bel(X) + Bel(∼ X) ≤ 1 ∀X ∈ F(W ).
Proof: 1) By Theorem 8, we can find a fuzzy belief space Bel(X) = P (RI (X)) = RI (X)(x)P (x)
((U, P ), W, R) such that Bel(X) = P (RI (X)) and Pl(X) = x∈U

P (RI (X)) for all X ∈ F(W ). Since R is a serial fuzzy re- Pl(X) = P (RI (X)) = RI (X)(x)P (x).
lation, by Theorem 4, we know that RI (X) ⊆ RI (X), then x∈U
P (RI (X)) ≤ P (RI (X)), i.e., Bel(X) ≤ Pl(X).
2) can be directly concluded from 1) and (35). Then, by Theorems 3 and 4, we have the following.
Definition 14: Let W be a nonempty set that may be infinite FMC1) Bel(∅) = P (RI (∅)) = P (∅) = 0.
and n ∈ N. A fuzzy set function F : F(W ) → [0, 1] is called a
fuzzy monotone Choquet capacity of order n on W if it satisfies FMC2) Bel(W ) = P (RI (W )) = P (U ) = 1.
the following axioms. FMC3) For any setsXi ∈ F(W ), i = 1, 2, . . . , n, n ∈ N,
FMC1) F (∅) = 0. we have
n n
FMC2) F (W ) = 1.
Bel Xi = P R I Xi
n i=1 i=1
FMC3) F Ai ≥ (−1)|J |+1 F Ai
n
i=1 ∅= J ⊆{1,2,...,n } i∈J
≥P RI (Xi ) (by FL4) in Theorem 3)
for all Ai ∈ F(W ), i = 1, 2, . . . , n. i=1
1
n
A fuzzy set function F : F(W ) → [0, 1] is called a fuzzy mono- = P RI (Xi ) dα (by Definition 3)
tone Choquet capacity of infinite order if F is a fuzzy monotone 0 i=1 α
Choquet capacity of order n for all n ∈ N.
n
1
A fuzzy set function F : F(W ) → [0, 1] is called a fuzzy = P (RI (Xi ))α dα (by Lemma 2)
alternating Choquet capacity of order n on W if it satisfies the 0 i=1
following axioms. 1
= (−1)|J |+1 P (RI (Xj ))α dα
FAC1) F (∅) = 0. 0 ∅= J ⊆{1,2,...,n } j ∈J
FAC2) F (W ) = 1.
(by property of crisp probability measure)
n 1

FAC3) F Ai ≤ (−1)|J |+1 F Ai = (−1)|J |+1
P RI (Xj ) dα
i=1 ∅= J ⊆{1,2,...,n } i∈J 0 α
∅= J ⊆{1,2,...,n } j ∈J
for all Ai ∈ F(W ), i = 1, 2, . . . , n. (by Lemma 2)

A fuzzy set function F : F(W ) → [0, 1] is called a fuzzy al- 1
ternating Choquet capacity of infinite order if F is a fuzzy = (−1)|J |+1 P RI Xj dα

∅= J ⊆{1,2,...,n } 0 j ∈J α
alternating Choquet capacity of order n for all n ∈ N.
The following lemma can be deduced directly from the defi- (by FL2) in Theorem 3)
nition of the α-level set of a fuzzy set.
Lemma 2: Let W be a nonempty set and Ai ∈ F(W ), i = = (−1)|J |+1 P RI Xj
1, 2, . . . , n. Then, for all α ∈ [0, 1], the following hold. ∅= J ⊆{1,2,...,n } j ∈J
n
n
1) Ai = (Ai )α . (by Definition 3)
i=1

α i=1

n
n
2) Ai = (Ai )α . = (−1)|J |+1 Bel Xj (by Theorem 8).
i=1 α i=1 ∅= J ⊆{1,2,...,n } j ∈J
Theorem 10 next shows that fuzzy belief and plausibility
functions determined by an implicator, respectively, satisfy the Thus, we have proved that Bel is a fuzzy monotone Choquet
subadditive and superadditive properties at any order. capacity of infinite order on W .
Theorem 10: Let W be a nonempty set that may be infinite and Similarly, we can prove that Pl is a fuzzy alternating Choquet
I a semicontinuous implicator on [0, 1]. If Bel, Pl : F(W ) → capacity of infinite order on W .
[0, 1] are the fuzzy belief and plausibility functions induced In [2, Corollary 4.4], it was proved that, in general, every
from a fuzzy belief structure (M, m), then Bel is a fuzzy mono- infinitely monotone map is a lower envelope, so by Theorem 10,
tone Choquet capacity of infinite order on W and Pl is a fuzzy the same can be said of every fuzzy belief function defined by a
alternating Choquet capacity of infinite order on W . system of fuzzy focal events and a fuzzy implicator.
VII. CONCLUSION [11] D. Dubois and H. Prade, “Evidence measures based on fuzzy information,”
Automatica, vol. 21, pp. 547–562, 1985.
We have defined a general type of fuzzy belief and plausibility [12] D. Dubois and H. Prade, “Properties of measures of information in ev-
functions in infinite spaces by constructing a fuzzy belief struc- idence and possibility theory,” Fuzzy Sets Syst., vol. 24, pp. 161–182,
1987.
ture and a measure of fuzzy inclusion determined by a fuzzy [13] D. Dubois and R. R. Yager, “Fuzzy set connectives as combinations of
implication operator. Many existing fuzzy belief and plausibil- belief structures,” Inf. Sci., vol. 66, pp. 245–275, 1992.
ity functions in the literature can be treated as special cases of [14] Q. H. Hu, D. R. Yu, Z. X. Xie, and J. F. Liu, “Fuzzy probabilistic approxi-
mation spaces and their information measures,” IEEE Trans. Fuzzy Syst.,
our definitions by taking specific fuzzy implication operators. vol. 14, no. 2, pp. 191–201, Apr. 2006.
We have shown that the fuzzy belief and plausibility functions [15] M. Ishizuka, K. S. Fu, and J. T. P. Yao, “Inference procedures and un-
defined by a fuzzy belief structure can be represented as the certainty for problem-reduction method,” Inf. Sci., vol. 28, pp. 179–206,
1982.
lower and upper probabilities in a countable set induced by [16] G. J. Klir, “A principle of uncertainty and information invariance,” Int. J.
the fuzzy belief space. We have also investigated properties of Gen. Syst., vol. 17, pp. 249–275, 1990.
the fuzzy belief and plausibility functions. The essential prop- [17] G. J. Klir and B. Yuan, Fuzzy Logic: Theory and Applications.
Englewood Cliffs, NJ: Prentice-Hall, 1995.
erties are that the fuzzy belief and plausibility functions are, [18] I. Kramosil, “Belief functions generated by fuzzy and randomized com-
respectively, the fuzzy monotone Choquet capacity and fuzzy patibility relations,” Fuzzy Sets Syst., vol. 135, pp. 341–366, 2003.
alternating Choquet capacity of infinite order. [19] P. J. Lingras and Y. Y. Yao, “Data mining using extensions of the rough
set model,” J. Amer. Soc. Inf. Sci., vol. 49, pp. 415–422, 1998.
Our results offer several advantages over previous work. First, [20] C. Lucas and B. N. Araabi, “Generalization of the Dempster–Shafer the-
semantics of the original Dempster–Shafer theory of evidence ory: A fuzzy-valued measure,” IEEE Trans. Fuzzy Syst., vol. 7, no. 3,
is still maintained. Of greater significance, the fuzzy belief and pp. 255–270, Jun. 1999.
[21] J.-S. Mi and W.-X. Zhang, “An axiomatic characterization of a fuzzy
plausibility functions on infinite universes of discourse can be generalization of rough sets,” Inf. Sci., vol. 160, pp. 235–249, 2004.
represented as the lower and upper fuzzy probabilities in a count- [22] N. N. Morsi and M. M. Yakout, “Axiomatics for fuzzy rough sets,” Fuzzy
able set. Second, we have provided a useful class of fuzzy belief Sets Syst., vol. 100, pp. 327–342, 1998.
[23] H. Ogawa and K. S. Fu, “An inexact inference for damage assessment of
and plausibility functions for real applications. One can select existing structures,” Int. J. Man-Mach. Stud., vol. 22, pp. 295–306, 1985.
different fuzzy implication operators with reference to various [24] Z. Pawlak, “Rough sets,” Int. J. Comput. Inf. Sci., vol. 11, pp. 341–356,
requirements to define fuzzy belief and plausibility functions in 1982.
[25] Z. Pawlak, Rough Sets: Theoretical Aspects of Reasoning about Data.
the analysis of uncertain information in complicated intelligent Boston, MA: Kluwer, 1991.
systems. Third, since the fuzzy belief and plausibility functions [26] Z. Pawlak and A. Skowron, “Rudiments of rough sets,” Inf. Sci., vol. 177,
can be interpreted as the lower and upper approximations in pp. 3–27, 2007.
[27] Z. Pawlak and A. Skowron, “Rough sets: Some extensions,” Inf. Sci.,
rough set theory, on one hand, rough set theory may be regarded vol. 177, pp. 28–40, 2007.
as the basis of the Dempster–Shafer theory of evidence. On the [28] A. M. Radzikowska and E. E. Kerre, “A comparative study of fuzzy rough
other hand, just as crisp belief and plausibility functions can be sets,” Fuzzy Sets Syst., vol. 126, pp. 137–155, 2002.
[29] C. Römer, A. Kandel, “Constraints on belief functions imposed by fuzzy
employed in knowledge reduction in complete and incomplete random variables,” IEEE Trans. Syst., Man, Cybern., vol. 25, no. 1,
information systems [41], [45], [58], the Dempster–Shafer the- pp. 86–99, Jan. 1995.
ory of evidence for fuzzy sets might provide a potentially useful [30] C. Römer, A. Kandel, “Applicability analysis of fuzzy inference by means
of generalized Dempster–Shafer theory,” IEEE Trans. Fuzzy Syst., vol. 3,
tool for reasoning and knowledge acquisition in fuzzy systems no. 4, pp. 448–453, Nov. 1995.
and fuzzy decision systems. [31] D. Ruan and E. E. Kerre, “Fuzzy implication operators and generalized
fuzzy method of cases,” Fuzzy Sets Syst., vol. 54, pp. 23–37, 1993.
[32] G. Shafer, A Mathematical Theory of Evidence. Princeton, NJ: Princeton
REFERENCES Univ. Press, 1976.
[33] A. Skowron, “The relationship between rough set theory and evidence
[1] A. A. Abdel-Hamid and N. N. Morsi, “On the relationship of extended theory,” Bulletin Polish Academy Sci.: Math., vol. 37, pp. 87–90, 1989.
necessity measures to implication operators on the unit interval,” Inf. Sci., [34] A. Skowron, “The rough sets theory and evidence theory,” Fundam. Inf.,
vol. 82, pp. 129–145, 1995. vol. 13, pp. 245–262, 1990.
[2] L. Biacino, “Fuzzy subsethood and belief functions of fuzzy events,” [35] A. Skowron and J. Grzymala Busse, “From rough set theory to evi-
Fuzzy Sets Syst., vol. 158, pp. 38–49, 2007. dence theory,” in Advance in the Dempster–Shafer Theory of Evidence,
[3] D. G. Chen, W. X. Yang, and F. C. Li, “Measures of general fuzzy rough R. R. Yager, M. Fedrizzi, and J. Kacprzyk, Eds. New York: Wiley, 1994,
sets on a probabilistic space,” Inf. Sci., vol. 178, pp. 3177–3187, 2008. pp. 193–236.
[4] G. Choquet, “Theory of capacities,” Annales de l’institut Fourier, vol. 5, [36] P. Smets, “The degree of belief in a fuzzy event,” Inf. Sci., vol. 25,
pp. 131–295, 1954. pp. 1–19, 1981.
[5] M. D. Cock, C. Cornelis, and E. E. Kerre, “Fuzzy rough sets: The forgotten [37] P. Smets, “The normative representation of quantified beliefs by belief
step,” IEEE Trans. Fuzzy Syst., vol. 15, no. 1, pp. 121–130, Feb. 2007. functions,” Artif. Intell., vol. 92, no. 1-/2, pp. 229–242, 1997.
[6] C. Cornelis, G. Deschrijver, and E. E. Kerre, “Implication in intuitionistic [38] P. Smets, “The combination of evidence in the transferable belief model,”
fuzzy and interval-valued fuzzy set theory: Construction, classification, IEEE Trans. Pattern Anal. Mach. Intell., vol. 12, no. 5, pp. 447–458, May
application,” Int. J. Approx. Reason., vol. 35, pp. 55–95, 2004. 1990.
[7] F. Cuzzolin, “A geometric approach to the theory of evidence,” IEEE [39] P. Smets and R. Kennes, “The transferable belief model,” Artif. Intell.,
Trans. Syst., Man, Cybern. C, Appl. Rev., vol. 38, no. 4, pp. 522–534, Jul. vol. 66, pp. 191–243, 1994.
2008. [40] X. Z. Wang, E. C. C. Tsang, S. Y. Zhao, D. G. Chen, and D. S. Yeung,
[8] A. P. Dempster, “Upper and lower probabilities induced by a multivalued “Learning fuzzy rules from fuzzy samples based on rough set technique,”
mapping,” Ann. Math. Statist., vol. 38, pp. 325–339, 1967. Fuzzy Sets Syst., vol. 177, pp. 4493–4514, 2007.
[9] T. Denoeux, “Modeling vague beliefs using fuzzy-valued belief struc- [41] W.-Z. Wu, “Attribute reduction based on evidence theory in incomplete
tures,” Fuzzy Sets Syst., vol. 116, pp. 167–199, 2000. decision systems,” Inf. Sci., vol. 178, no. 5, pp. 1355–1371, 2008.
[10] D. Dubois and H. Prade, “Fuzzy sets in approximate reasoning. Part 1: [42] W.-Z. Wu, Y. Leung, and J.-S. Mi, “On characterizations of (I, T )-fuzzy
Inference with possibility distributions,” Fuzzy Sets Syst., vol. 40, pp. 143– rough approximation operators,” Fuzzy Sets Syst., vol. 154, pp. 76–102,
202, 1991. 2005.
[43] W.-Z. Wu, Y. Leung, and W.-X. Zhang, “Connections between rough Yee Leung received the B.S.Sc. degree in geography
set theory and Dempster–Shafer theory of evidence,” Int. J. Gen. Syst., from The Chinese University of Hong Kong, Shatin,
vol. 31, pp. 405–430, 2002. Hong Kong, in 1972, the M.Sc. and Ph.D. degrees in
[44] W.-Z. Wu, W.-X. Zhang, and H.-Z. Li, “Knowledge acquisition in incom- geography, in 1974 and 1977, respectively, and the
plete fuzzy information systems via rough set approach,” Expert Syst., M.S. degree in engineering in 1977 from the Univer-
vol. 20, pp. 280–286, 2003. sity of Colorado, Boulder.
[45] W.-Z. Wu, M. Zhang, H.-Z. Li, and J.-S. Mi, “Knowledge reduction in He is currently a Professor of geography in the De-
random information systems via Dempster–Shafer theory of evidence,” partment of Geography and Resource Management,
Inf. Sci., vol. 174, pp. 143–164, 2005. The Chinese University of Hong Kong. His current
[46] R. R. Yager, “Generalized probabilities of fuzzy events from fuzzy belief research interests include the development and ap-
structures,” Inf. Sci., vol. 28, pp. 45–62, 1982. plications of intelligent spatial decision support sys-
[47] R. R. Yager, “On the normalization of fuzzy belief structure,” Int. J. tems, spatial data mining, fuzzy sets and logic, rough sets, concept lattices,
Approx. Reason., vol. 14, pp. 127–153, 1996. neural networks, and evolutionary computation. He has authored or coauthored
[48] Y. Y. Yao, “Generalized rough set model,” in Rough Sets in Knowl- five monographs and more than 150 articles in international journals and book
edge Discovery 1. Methodology and Applications, L. Polkowski and chapters. He serves in the editorial boards of several international journals.
A. Skowron, Eds. Heidelberg, Germany: Physica-Verlag, 1998, pp. 286–
318.
[49] Y. Y. Yao, “Relational interpretations of neighborhood operators and rough
set approximation operators,” Inf. Sci., vol. 111, pp. 239–259, 1998.
[50] Y. Y. Yao, “Constructive and algebraic methods of the theory of rough
sets,” J. Inf. Sci., vol. 109, pp. 21–47, 1998.
[51] Y. Y. Yao and P. J. Lingras, “Interpretations of belief functions in the
theory of rough sets,” Inf. Sci., vol. 104, pp. 81–106, 1998. Ju-Sheng Mi received the B.Sc. degree in mathe-
[52] J. Yen, “Generalizing the Dempster–Shafer theory to fuzzy sets,” IEEE matics from Hebei Normal University, Hebei, China,
Trans. Syst., Man, Cybern., vol. 20, no. 3, pp. 559–570, May/Jun. 1990. in 1986, the M.Sc. degree in mathematics from East
[53] J. Yen, “Computing generalized belief functions for continuous fuzzy China Normal University, Shanghai, China, in 1992,
sets,” Int. J. Approx. Reason., vol. 6, pp. 1–31, 1992. and the Ph.D. degree in applied mathematics from
[54] D. S. Yeung, D. G. Chen, E. C. C. Tsang, J. W. T. Lee, and X. Z. Wang, Xi’an Jiaotong University, Xi’an, China, in 2003.
“On the generalization of fuzzy rough sets,” IEEE Trans. Fuzzy Syst., He is currently a Professor in the College of Math-
vol. 13, no. 3, pp. 343–361, Jun. 2005. ematics and Information Science, Hebei Normal Uni-
[55] L. A. Zadeh, “Fuzzy sets as a basis for a theory of possibility,” Fuzzy Sets versity. His current research interests include approx-
Syst., vol. 1, pp. 3–28, 1978. imate reasoning, rough sets, concept lattices and ran-
[56] L. A. Zadeh, “Fuzzy sets and information granularity,” in Advances in dom sets. He has authored or coauthored more than
Fuzzy Set Theory and Applications, M. Gupta, R. Ragade, and R. R. Yager, 50 articles in international journals and book chapters.
Eds. Amsterdam, The Netherlands: North Holland, 1979, pp. 3–18.
[57] L. A. Zadeh, “Probability measures of fuzzy events,” J. Math. Anal. Appl.,
vol. 23, pp. 421–427, 1968.
[58] M. Zhang, L. D. Xu, W.-X. Zhang, and H.-Z. Li, “A rough set approach to
knowledge reduction based on inclusion degree and evidence reasoning
theory,” Expert Syst., vol. 20, pp. 298–304, 2003.
Wei-Zhi Wu received the B.Sc. degree in mathemat-

ics from Zhejiang Normal University, Jinhua, China,
in 1986, the M.Sc. degree in mathematics from East
China Normal University, Shanghai, China, in 1992,
and the Ph.D. degree in applied mathematics from
Xi’an Jiaotong University, Xi’an, China, in 2002.
He is currently a Professor of mathematics in
the School of Mathematics, Physics, and Informa-
tion Science, Zhejiang Ocean University, Zhoushan,
China. His current research interests include approxi-
mate reasoning, rough sets, random sets, formal con-
cept analysis, and granular computing. He has authored or coauthored two
monographs and more than 60 articles in international journals and book chap-
ters. He serves in the editorial boards of several international journals.
Fault Detection for Fuzzy Systems With Intermittent

Measurements
Yan Zhao, James Lam, Senior Member, IEEE, and Huijun Gao, Member, IEEE
Abstract—This paper investigates the problem of fault detection of the advancement of modeling and state estimation techniques
for Takagi–Sugeno (T–S) fuzzy systems with intermittent measure- [6], [7], [19], [20], model-based fault detection has been well
ments. The communication links between the plant and the fault de- developed [26]. To mention a few, fault detection problems have
tection filter are assumed to be imperfect (i.e., data packet dropouts
occur intermittently, which appear typically in a network environ- been investigated for sampled-data systems in [9] and [32], un-
ment), and a stochastic variable satisfying the Bernoulli random certain systems in [1], [11], and [33], systems with time delays
binary distribution is utilized to model the unreliable communica- in [10], and Markovian jump linear systems in [14].
tion links. The aim is to design a fuzzy fault detection filter such Most of the aforementioned results are concerned with linear
that, for all data missing conditions, the residual system is stochasti- models. But in reality, most physical systems are nonlinear, and
cally stable and preserves a guaranteed performance. The problem
is solved through a basis-dependent Lyapunov function method, thus, how to develop effective fault detection methods for non-
which is less conservative than the quadratic approach. The re- linear systems is an important and practical problem. However,
sults are also extended to T–S fuzzy systems with time-varying the difficulty in modeling nonlinearities makes fault detection
parameter uncertainties. All the results are formulated in the form a hard task. To solve this problem, some researchers model the
of linear matrix inequalities, which can be readily solved via stan- nonlinear plants as differential equations and solve the fault de-
dard numerical software. Two examples are provided to illustrate
the usefulness and applicability of the developed theoretical results. tection problem based on the conventional nonlinear system the-
ory [16], whose limitations often do not generalize the obtained
Index Terms—Basis-dependent Lyapunov functions, fault de- results. Other researchers take the advantage of artificial intel-
tection, intermittent measurements, Takagi–Sugeno (T–S) fuzzy
systems, uncertainties. ligence techniques [23], and use the conventional fuzzy models
to represent the nonlinear systems by applying the inference
engine. It is worth mentioning that there are no systematic and
I. INTRODUCTION consistent approaches for the stability and performance analysis
of those conventional fuzzy systems, and hence, the applicabil-
N CONTROL systems, due to the unexpected variations
I in external surroundings, normal wear in components, or
sudden changes in signals, there may appear different kinds
ity of those results is also limited.
In recent years, Takagi–Sugeno (T–S) fuzzy models are play-
ing more and more important roles in dealing with problems
of malfunction or imperfect behavior in normal operations, and
concerning nonlinear systems [3]. It has been proven that T–S
people call them faults. Since a fault can degrade a system’s per-
fuzzy systems with affine terms can smoothly approximate any
formance and even cause catastrophic accidents, it is of great
nonlinear functions to any specified accuracy within any com-
significance to detect it in time for the safety and reliability of
pact set, which provides a theoretical foundation for using T–S
control systems. The objective of fault detection is to detect the
fuzzy models to represent complex nonlinear systems approx-
fault signal accurately whenever it appears. Many researchers
imately. Meanwhile, T–S fuzzy models formulate the complex
have devoted themselves to investigating this problem, and a lot
nonlinear systems into a framework that interpolates some affine
of methods have been established, mainly including the model-
local models by a set of fuzzy membership functions. Based on
based fault detection approach [2], [4], the parameter estima-
this framework, a systematic analysis and design procedure for
tion approach [24], and the generalized likelihood method [30].
complex nonlinear systems can be possibly developed in view
Among these methods, the model-based one is very popular,
of the powerful control theories and techniques in linear sys-
which is to design a fault detection filter or observer generating a
tems. The T–S fuzzy model has attracted great interests from
residual including a threshold to detect the fault signal. In virtue
researchers, and a number of results have been reported in lit-
eratures, including stability analysis [13], [27], stabilizing and
Manuscript received April 15, 2008; revised August 26, 2008 and November
5, 2008; accepted December 24, 2008. First published February 10, 2009; H∞ control design [12], [21], [34], and state estimation [35].
current version published April 1, 2009. This work was supported in part by Since T–S fuzzy models have provided a convenient way to
the National Natural Science Foundation of China (60825303, 60834003), in study nonlinear systems, a feasible solution of the fault detec-
part by the 973 Project (2009CB320600), in part by the Research Fund for
the Doctoral Programme of Higher Education of China (20070213084), in part tion problem for nonlinear systems can be converted to that of
by the Heilongjiang Outstanding Youth Science Fund (JC200809), and in part fault detection for T–S fuzzy systems [17].
by Postdoctoral Science Foundation of China (200801282), and in part by the On the other hand, data packet dropout phenomena may often
University of Hong Kong Research and Conference Grants (CRCG) under Grant
200707176077. appear in many practical situations, i.e., some measurements
Y. Zhao and H. Gao are with the Space Control and Inertial Technology or control inputs may be lost during the transmission. This
Research Center, Harbin Institute of Technology, Harbin 150001, China (e- problem has attracted more and more attention as the insertion
mail: zhaoresponsible@gmail.com; hjgao@hit.edu.cn).
J. Lam is with the Department of Mechanical Engineering, University of of networked control systems (NCSs) in the control loops
Hong Kong, Hong Kong (e-mail: james.lam@hku.hk). becomes popular [15]. Compared with the traditional point-to-
Digital Object Identifier 10.1109/TFUZZ.2009.2014860 point communication bus, NCSs have several advantages such
1063-6706/$25.00 © 2009 IEEE
ZHAO et al.: FAULT DETECTION FOR FUZZY SYSTEMS WITH INTERMITTENT MEASUREMENTS 399
as low cost, reduced weight and power requirements, simple sented by a T–S fuzzy model, and the signal transmissions ex-
installation and maintenance, and high reliability. However, isting between the plant and the fault detection filter are inter-
in an NCS, since several components communicate over a mittent.
shared network, information flows are prone to the curse of
time sharing, and data loss always inevitably occurs from the A. Physical Plant
plant to the filter or controller. Since data packet dropout can
The nonlinear discrete-time system whose faults are to be
degrade a system’s performance and even cause instability,
detected is represented by the following T–S fuzzy model.
it has been regarded as an important issue in the analysis
Plant rule i: IF θ1 (k) is Mi1 and θ2 (k) is Mi2 and · · · and
and synthesis of network-based control systems, and some
θp (k) is Mip , THEN
researchers have begun to study various problems of control
systems, simultaneously considering this communication xk +1 = (Ai + ∆Ai (k)) xk + (Bi + ∆Bi (k)) uk
issue [28]. It is noted that most investigations concerning the
+ E1i wk + E2i fk
data loss phenomenon are focused on the stability analysis and
synthesis, and the plants are mostly linear. To the best of the yk = Ci xk + Di uk + F1i wk + F2i fk ,
our knowledge, there are no results about the fault detection
i = 1, . . . , r (1)
problem for nonlinear systems with intermittent measurements.
Motivated by the aforementioned observations, in this pa- where Mij is the fuzzy set, xk ∈ Rn p is the state vector; r
per, we investigate the problem of fault detection for T–S fuzzy is the number of IF–THEN rules, θ(k) = [θ1 (k), θ2 (k), . . . ,
systems with intermittent measurements. The measurements be- θp (k)] is the premise variable vector, uk ∈ Rm is the deter-
tween the plant and the fault detection filter are assumed to be ministic input vector, wk ∈ Rp is the exogenous disturbance
intermittent, and a stochastic variable is utilized to describe input that belongs to l2 [0, ∞), and fk ∈ Rq is the fault vector
the imperfect communication links. Attention is focused on the that is also deterministic. Without the loss of generality, we as-
fuzzy fault detection filter design such that the residual sys- sume that the l2 norms of uk and fk exist and are bounded.
tem is stochastically stable with the prescribed performance. A Ai , Bi , E1i , E2i , Ci , Di , F1i , and F2i are known constant ma-
basis-dependent Lyapunov function is utilized in the derivative trices with appropriate dimensions, ∆Ai (k) and ∆Bi (k) denote
process, which renders the results to be potentially less con- the uncertainties in the model and are of the form
servative. Furthermore, the results are extended to T–S fuzzy
∆Ai (k) = Ni Z(k)Qai ∆Bi (k) = Ni Z(k)Qbi ,
systems with time-varying uncertainties. All the results are for-
mulated in the form of linear matrix inequalities (LMIs). Two i = 1, . . . , r (2)
examples are illustrated to show the usefulness and applicability n p ×n z n z ×n p n z ×m
where Ni ∈ R , Qai ∈ R , and Qbi ∈ R are
of the obtained results.
known constant matrices, and Z(k) ∈ Rn z ×n z is an un-
The remainder of the paper is organized as follows. Section II
known time-varying matrix with Lebesgue measurable elements
formulates the problem under consideration. Section III presents
bounded by
the fault detection filter design for the nominal fuzzy system,
and the results are extended to the fuzzy system with time- Z T (k)Z(k) ≤ I. (3)
varying uncertainties in Section IV. Two examples are illustrated
Given a pair of (xk , uk ), the overall fuzzy system is inferred
in Section V to show the usefulness and applicability of the
as
proposed approaches, and the paper is concluded in Section VI.
The notation used throughout the paper is fairly stan-
r
xk +1 = hi (θ(k))[(Ai + ∆Ai (k))xk + (Bi + ∆Bi (k))uk
dard. The superscript “T ” stands for matrix transposition, Rn i=1
denotes the n-dimensional Euclidean space, 0 represents the
zero matrix with appropriate dimensions, the notation P > 0 + E1i wk + E2i fk ]
(≥ 0) means that P is real symmetric and positive definite
r
(semidefinite), l2 [0, ∞) is the space of square-integrable vector yk = hi (θ(k))[Ci xk + Di uk + F1i wk + F2i fk ] (4)
functions over [0, ∞), and · 2 stands for the usual l2 [0, ∞) i=1

norm. In symmetric block matrices or complex matrix expres- where
p hi (θ(k)) = ωi (θ(k))/ ri=1 ωi (θ(k)) and ωi (θ(k)) =
sions, we use an asterisk (∗) to represent a term that is induced by j =1 Mij (θj (k)), with Mij (θj (k)) representing the grade of
symmetry and diag{. . .} stands for a block-diagonal ma- membership of θj (k) in Mij . Then, it can be seen that
trix. In addition, E{x} and E{ x| y} will, respectively, mean

r
expectation of x and expectation of x conditional on y. Matri- ωi (θk ) ≥ 0, i = 1, 2, . . . , r, ωi (θk ) > 0
ces, if their dimensions are not explicitly stated, are assumed to i=1
be compatible for algebraic operations.
for all k.
II. PROBLEM FORMULATION B. Fault Detection Filter

Consider the fault detection problem for T–S fuzzy systems One key step of fault detection is the generation of a residual
with intermittent measurements. The physical plant is repre- signal, which must be sensitive to faults. This is often realized
by utilizing fault detection observers [25], [26], [33] or filters characteristics. The inferred statistics of the Bernoulli process
[10], [14], [20], [33]. Since disturbances often inevitably appear will then be used for designing the fault detection filter.
in many systems, the residual signal must also be capable of
distinguishing faults from exogenous signals. H∞ filter can not D. Fault Weighting System
only describe the estimated signal accurately but also suppress
the disturbance effectively. Thus, for the physical plant with In fault detection, a reference residual model is usually needed
disturbance in (1), we adopt the following fuzzy fault detection to describe the desired behavior of the residual vector rk . In this
filter form, whose role is to generate residual signal based on paper, the reference model is chosen as f¯ (z) = W (z) f (z)
the input yf k . [33], where W (z) is given a priori. The choice of Wz is to
Filter Rule i: IF θ1 (k) is Mi1 and θ2 (k) is Mi2 and · · · and impose frequency weighting on the spectrum of the fault signal
θp (k) is Mip , THEN for detection. Here, we choose a stable matrix W (z) to weight
the fault signal fk [33], whose state-space realization is
x̂k +1 = Af i x̂k + Bf i yf k
x̄k +1 = AW x̄k + BW fk
rk = Cf i x̂k + Df i yf k
f¯k = CW x̄k + DW fk (9)
i = 1, . . . , r. (5)
where x̄k ∈ Rn W and AW , BW , CW , and DW are priorly chosen.
Here, x̂k ∈ Rn f and rk ∈ Rq , and Af i , Bf i , Cf i , and Df i are
to be determined. Thus, the filter can be represented by the
following form: E. Residual Evaluation

r The residual evaluation function is to evaluate the generated
x̂k +1 = hi (θ(k)) (Af i x̂k + Bf i yf k ) residual. After the residual signal being constructed, a residual
i=1 evaluation value will be computed through a prescribed evalua-

r tion function, and it will be compared with a predefined thresh-
rk = hi (θ(k)) (Cf i x̂k + Df i yf k ) . (6) old. When the evaluation value is larger than the threshold, an
i=1 alarm of fault is generated. Here, we consider the following
evaluation function:
C. Communication Links
t2
1
In this paper, we assume that a communication medium exists rT = rkT rk T = t2 − t1 + 1.
between the physical plant and the fault detection filter, and T
k =t 1
the data packet dropout phenomenon happens intermittently.
Therefore, the measurement of the plant is no longer equivalent Choose a threshold Jth > 0, and for the detailed discussion of
to the input of the filter (i.e., yk = yf k ). A stochastic process is the threshold Jth , readers are referred to [1] and [5]. The residual
utilized to model the data loss phenomenon, i.e. evaluation function value and the threshold satisfy the following
relationship:
yf k = αk yk (7)
rT > Jth =⇒ with faults =⇒ alarm
where αk is a Bernoulli process. When the link fails (i.e., data
are lost), αk = 0, and when the transmission is perfect, αk = 1. rT ≤ Jth =⇒ no faults.
A natural assumption on αk can be made as
F. Residual System
Prob {αk = 1} = E {αk } = ᾱ Prob {αk = 0} = 1 − ᾱ
From (4), (5), and (9), the residual system can be obtained as
where ᾱ is assumed to be known. Based on this, we have

r
r

r ξk +1 = hi (θ(k))hj (θ(k))
x̂k +1 = hi (θ(k)) (Af i x̂k + αk Bf i yk ) i=1 j =1
i=1
× [Ãij ξk + B̃ij ϑk + α̃k Ā1ij ξk + α̃k B̄1ij ϑk ]

r
rk = hi (θ(k)) (Cf i x̂k + αk Df i yk ) . (8)
r
r
i=1 ek = hi (θ(k))hj (θ(k))
i=1 j =1
Remark 1: The description of imperfect communication links

existing between the plant and the fault detection filter follows × C̄ij ξk + D̄ij ϑk + α̃k C̄1ij ξk + α̃k D̄1ij ϑk (10)
that in the previous literature [28], [29]. The process of missing
data considered is assumed to satisfy the Bernoulli distributed where
process. The probability distribution of the process is estimated
ξk = [ x̄Tk xTk x̂Tk ]T ϑk = [ uTk wkT fkT ]T
based on experimental measurements of data transmitting from  
output of the plant to the input of the fault detection filter. This AW 0 0
can be achieved by sending a sequence of indexed data through ek = rk − f¯k , Ãij =  0 Ai + ∆Ai (k) 0 
the communication medium and measuring the data dropout 0 ᾱBf i Cj Af i
 
0 0 BW expectation is often used. The definition proposed here also
  obeys the criterion.
B̃ij =  Bi + ∆Bi (k) E1i E2i 
ᾱBf i Dj ᾱBf i F1j ᾱBf i F2j
  III. FUZZY FAULT DETECTION FILTER DESIGN
0 0 0
FOR NOMINAL SYSTEMS
Ā1ij = 0 0 0
0 Bf i Cj 0 In this section, the FFDIM problem is solved for the nominal
  fuzzy system in (12). The fault detection analysis problem is first
0 0 0 solved, and then, based on that, a full-rank fault detection filter
B̄1ij =  0 0 0  is designed (i.e., nf = np + nW ). The fuzzy basis-dependent
Bf i Dj Bf i F1j Bf i F2j technique is utilized, which potentially reduces the conservatism
C̄ij = [ −CW ᾱDf i Cj Cf i ] of the obtained results.
The nominal system of (4) takes the following form:
D̄ij = [ ᾱDf i Dj ᾱDf i F1j ᾱDf i F2j − DW ]
C̄1ij = [ 0 Df i Cj 0]
r
xk +1 = hi (θ(k)) [Ai xk + Bi uk + E1i wk + E2i fk ]
D̄1ij = [ Df i Dj Df i F1j Df i F2j ] i=1

r
yk = hi (θ(k)) [Ci xk + Di uk + F1i wk + F2i fk ] (12)
α̃k = αk − ᾱ, E{α̃k } = 0 and E{α̃k α̃k } = ᾱ(1 − ᾱ). i=1
The residual system presents the difference between the gen-
erated residual and the idealized reference residual signal. By and the nominal residual fuzzy system is given by
minimizing the H∞ norm of the difference, the effect of the dis-
turbance can be minimized and the sensitivity of the residual to
r
r
fault can be maximized [1], [18], [33]. Therefore, design of the ξk +1 = hi (θ(k))hj (θ(k))
fault detection filter can be converted as an H∞ model matching i=1 j =1
problem [33].
× Āij ξk + B̄ij ϑk + α̃k Ā1ij ξk + α̃k B̄1ij ϑk
Then, the problem to be addressed in this paper is expressed
as follows.
r
r
Problem: Fuzzy fault detection with intermittent measure- ek = hi (θ(k))hj (θ(k))

ments (FFDIMs): Consider the fuzzy system in (4), and suppose i=1 j =1

that the intermittent transmission parameter ᾱ is known. Given × C̄ij ξk + D̄ij ϑk + α̃k C̄1ij ξk + α̃k D̄1ij ϑk (13)
a scalar γ > 0, design a fuzzy fault detection filter in the form
of (5) such that: where
1) the residual system in (10) is stochastically stable; and
2) under zero initial conditions, the residual error ek satisfies  
AW 0 0
 
Āij =  0 Ai 0 
eE ≤ γϑ2 (11) 0 ᾱBf i Cj Af i
 
0 0 BW

∞  
where eE = E{ k =0 eTk ek } and ϑ2 = ( ∞ T
k =0 ϑk B̄ij =  Bi E1i E2i .
1
ϑk ) 2 . It is noted that the l2 norm of ϑk exists and is ᾱBf i Dj ᾱBf i F1j ᾱBf i F2j
bounded since its constituent variables are all l2 norm
bounded.
A. Fault Detection Analysis
Before proceeding further, we first introduce the following
definition. In this section, we assume that the fault detection filter matri-
Definition 1: The residual system in (10) is said to be stochas- ces in (5) are known, and the conditions are investigated under
tically stable in the mean square (ϑk ≡ 0) if there exists a finite which the residual system is stochastically stable and guarantees
V > 0independent of ξ0 , such that for any initial condition the performance defined in (11). The following theorem tells us
ξ0 E{ ∞ k =0 ξk ξk |ξ0 } < ξ0 V ξ0 .
T T
that the performance of the residual system can be guaranteed
Remark 2: Different statements of stochastic stability defini- if there exist some matrices satisfying certain LMIs.
tions are presented in [22], which serve for different systems. Theorem 1: Consider the fuzzy system in (12) and suppose the
They have the same criterion that the expectation values of all fault detection filter matrices Af i , Bf i , Cf i , and Df i in (6) are
the solutions of the system must be energy-bounded when the known. The residual system in (13) is stochastically stable with
operation time is infinite. The solutions are generally dependent a guaranteed performance γ if there exist n-dimensional matri-
on initial conditions or other elements, and thus, the condition ces Pl > 0 for l = 1, . . . , r, where n = nf + nW + np , and an
n-dimensional matrix G satisfying the following inequality: Define

  Ψ̄ij l = ĀTij Pl Āij + ᾱ (1 − ᾱ) ĀT1ij Pl Ā1ij − Pi
Πl 0 0 0 GT Āij GT B̄ij
 ∗ Π 0 f GT Ā1ij f GT B̄1ij 
r
 l 0  r r
  Ψ(k) = hl (θ(k + 1)) hi (θ(k))hj (θ(k))Ψ̄ij l .
 ∗ ∗ −I 0 C̄ij D̄ij 
 <0
 ∗ ∗ ∗ −I f D̄1ij 
l=1 i=1 j =1
 f C̄1ij 
  Then, the derivative process in (18) concludes J ≤ ξkT Ψ(k)ξk ,
 ∗ ∗ ∗ ∗ −Pi 0 
i.e.
∗ ∗ ∗ ∗ ∗ −γ 2 I
r
T
i, j, l = 1, . . . , r (14) E ξk +1 hl (θ(k + 1))Pl ξk +1 ξk

l=1
where f = ᾱ (1 − ᾱ) and Πl = Pl − G − GT .

r
Proof: Suppose there exist real symmetric positive definite − ξkT hi (θ(k))Pi ξk ≤ ξkT Ψ(k)ξk . (19)
matrices Pl for l = 1, . . . , r and a nonsingular matrix G satis- i=1
fying (14). Noting the inequality (Pl − G)T Pl−1 (Pl − G) ≥ 0
Taking mathematical expectation and summing up the terms on
implies Pl − G − GT ≥ −GT Pl−1 G, which together with (14)
both sides of (19) for k = 0, . . . , β, for any β > 1, we have
yields
  r r
Π̃i 0 0 0 GT Āij GT B̄ij T
E ξβ +1 hl (θ(β + 1))Pl ξβ +1 − ξ0 T
hi (θ(0))Pi ξ0
 ∗ Π̃ 0 f GT Ā1ij f GT B̄1ij  l=1 i=1
 i 0   
 
 ∗ ∗ −I 0 C̄ D̄  β
r r
r 
 ij ij  < 0 (15)
 ∗ ≤E hi(θ(k))hj (θ(k))ξkT Ψ̄ij l ξk
f D̄1ij 
hl (θ(β+1))
 ∗ ∗ −I f C̄1ij   
  k =0 l=1 i=1 j =1
 ∗ ∗ ∗ ∗ −Pi 0  
∗ ∗ ∗ ∗ ∗ −γ 2 I β
r
r
r
≤E hl (θ (β + 1)) hi (θ(k))hj (θ(k))

where Π̃l = −G T
Pl−1 G.
k =0 l=1 i=1 j =1
Pre- diag G−T , G−T , I, I, I, I and

and postmultiplying
diag G−1 , G−1 , I, I, I, I to (15) and by Schur complement, × λm ax Ψ̄ij l ξkT ξk
we have
T T
T
β
Āij f ĀT1ij Pl 0 Āij B̄ij C̄ij f C̄1ij
T T
+ T T
=E max λm ax Ψ̄ij l ξkT ξk .
B̄ij f B̄1ij 0 Pl f Ā1ij f B̄1ij D̄ij f D̄1ij k =0
i,j,l=1,...,r

C̄ij D̄ij Pi 0 From the aforementioned inequalities, it is not difficult to con-
× − < 0. (16)
f C̄1ij f D̄1ij ∗ γ2 I clude that, for i, j, l = 1, . . . , r and β, the following inequality
is true:
Now, we first prove the stochastic stability of the residual system
in (13). Define an index as E ξβT +1 Pl ξβ +1 − ξ0T Pi ξ0
β
r r
≤ max λm ax Ψ̄ij l E T
ξk ξk .
J = E ξkT+1 hl (θ(k+1))Pl ξk +1ξk − ξkT hi (θ(k))Pi ξk .
i,j,l=1,...,r
k =0
l=1 i=1
(17) When β → ∞, we obtain
When ϑk ≡ 0, along the nominal system in (13), J gives T
 E ξ∞ Pl ξ∞ − ξ0T Pi ξ0
 r r r r r ∞

J =E hl (θ(k + 1)) hi (θ(k))hj (θ(k)) ≤ max λm ax Ψ̄ij l E ξkT ξk .
 i,j,l=1,...,r
l=1 i=1 j =1 s=1 t=1
k =0
T T
×hs (θ(k))ht (θ(k)) Āij ξk + α̃k Ā1ij ξk Pl Considering nonzero initial condition and E ξ∞ Pl ξ∞ ≥ 0,
we have
r
∞

× Āst ξk + α̃k Ā1st ξk ξk − ξkT hi (θ(k))Pi ξk ! "−1
T
E ξk ξk ξ0 ≤ − max (λm ax (Ψ̄ij l )) ξ0T Pi ξ0
i=1 i,j,l=1,...,r
k =0

r
r
r
! ! "−1 "
≤ ξkT hl (θ(k + 1)) hi (θ(k))hj (θ(k))
≤ ξ0T − max (λm ax (Ψ̄ij l )) max (λm ax (Pi )) ξ0
l=1 i=1 j =1 i,j,l=1,...,r i=1,...,r

× ĀTij Pl Āij + ᾱ (1 − ᾱ) ĀT1ij Pl Ā1ij − Pi ξk . (18) = σξ0T ξ0
where where ηk = [ ξkT ϑTk ]T . By substituting (21) into (20), the

! "−1 following holds:
σ= − max (λm ax (Ψ̄ij l )) max (λm ax (Pi ))
i,j,l=1,...,r i=1,...,r

r
r
r
and ξ0 is the initial condition. From (1, 1) block in the left side J¯ ≤ hl (θ(k + 1)) hi (θ(k))hj (θ(k))
l=1 i=1 j =1
of (16), Ψ̄ij l < 0 is obtained, and thus, σ > 0. According to
$ % $ %
Definition 1, the residual system is stochastically stable in the ĀTij ĀT1ij
mean square. × ηkT Pl [ Āij B̄ij ] + f 2
Pl
T T
Next, we prove that the performance defined in (11) is guar- B̄ij B̄1ij
$ %
anteed. To this end, assume zero initial condition and ϑk = 0. T
C̄ij
An index is introduced as × [ Ā1ij B̄1ij ] + [ C̄ij D̄ij ]
T
D̄ij
r
$ %
J¯ = E ξkT+1 hl (θ(k + 1))Pl ξk +1 ξk T
C̄1ij
+f 2
[ C̄1ij D̄1ij ] ηk − ηkT
Pi 0
ηk
l=1
T
D̄1ij 0 γ2 I

r

− ξkT hi (θ(k))Pi ξk + E eTk ek ξk − γ 2 ϑTk ϑk . (20) which leads to J¯ ≤ 0 by consideration of (16), i.e.
i=1
r r
Along the nominal system in (13), we have
T
E ξk +1 hl (θ(k + 1))Pl ξk +1 ξk − ξkT hi (θ(k))Pi ξk

r l=1 i=1
T
E ξk +1 hl (θ(k + 1))Pl ξk +1 ξk + E eTk ek ξk − γ 2 ϑTk ϑk ≤ 0.
(22)
l=1
 Taking mathematical expectation on both sides of (22), we
 r r r r
r
obtain
=E hl (θ(k + 1)) hi (θ(k))hj (θ(k))
 r
l=1 i=1 j =1 s=1 t=1
T
#$ % $ %& E ξk +1 hl (θ(k + 1))Pl ξk +1
ĀTij ĀT1ij l=1
× hs (θ(k))ht (θ(k))ηkT + α̃k
T
B̄ij T
B̄1ij
r
− E ξkT hi (θ(k))Pi ξk + E{eTk ek } − γ 2 ϑTk ϑk ≤ 0.

i=1
×Pl ([ Āst B̄st ] + α̃k [ Ā1st B̄1st ]) ηk ξk
For k = 0, 1, 2, . . . , summing up both sides of the afore-
mentioned
r inequality, considering zero initial condition and

r
r
r
i=1 hi (θ(∞))Pi ξ∞ } > 0, we have
T
≤ hl (θ(k + 1)) hi (θ(k))hj (θ(k))ηkT E{ξ∞
∞ ∞
l=1 i=1 j =1
$ % $ % E ek ek − γ
T 2
ϑTk ϑk ≤ 0
ĀTij ĀT1ij
× Pl [ Āij B̄ij ] + f 2
Pl [ Ā1ij B̄1ij ] ηk k =0 k =0
T T
B̄ij B̄1ij which is equivalent to the inequality in (11), and thus, the proof

E eTk ek ξk is completed.
 Remark 3: Results of fault detection for fuzzy systems provide
 r r r r feasible solutions to the problem of fault detection for nonlinear
=E hi (θ(k))hj (θ(k)) systems [17], which are useful in practice since most physical

i=1 j =1 s=1 t=1 systems in the real world are nonlinear. Previous results are
#$ T % $ T %& mostly concerned with the perfect communication links. Ac-
C̄ij C̄1ij
× hs (θ(k))ht (θ(k))ηkT + α̃k tually, in practice, the transmission is often imperfect between
T
D̄ij T
D̄1ij the plant and the filter, i.e., data packet dropout may occur
intermittently, especially in systems based on the network com-

munication links. In this paper, data missing is considered in the
× ([ C̄st D̄st ] + α̃k [ C̄1st D̄1st ]) ηk ξk
fuzzy model, which makes the obtained results more general
and practical.

r
r
≤ ηkT hi (θ(k))hj (θ(k))
i=1 j =1
B. Fault Detection Filter Design
#$ T % $ T
% & In this section, the fault detection filter design problem will
C̄ij C̄1ij
× [ C̄ij D̄ij ] + f 2
[ C̄1ij D̄1ij ] ηk be investigated based on Theorem 1, i.e., a method will be
T T
D̄ij D̄1ij developed to determine the fault detection filter matrices in (5),
such that the residual system in (13) is stochastically stable and
(21)
the performance defined in (11) is guaranteed.
Theorem 2: Consider the fuzzy system in (12). For a given From (23), we know that
positive constant γ, if there exist n-dimensional matrices
P̃l < Ω + ΩT
P̃1l P̃2l
P̃l = > 0, which implies Ω and W are nonsingular. One can always find
P̃2lT P̃3l square and nonsingular matrices G3 and G4 such that W =
where n = nf + nW + np , matrices Ăf i , B̆f i , C̆f i , and D̆f i , GT4 G−1
3 G4 . Let
for any i, l = 1, . . . , r, and (np + nW )-dimensional matrices
G1 G2
U, X, and W satisfy the following inequality: G1 = U X = GT2 G−1 3 G4 V = G −1
3 G4 G =
G4 G3
 
Θl 0 0 0 Θ15ij Θ16ij (26)
 ∗ Θ 0 Θ25ij Θ26ij  and define a transposition matrix
 l 0 
 
 ∗ ∗ −I 0 Θ35ij Θ36ij  I 0
Φij l = 
 ∗
<0 T = . (27)
 ∗ ∗ −I Θ45ij Θ46ij   0 G−13 G4
 
 ∗ ∗ ∗ ∗ Θ55i 0  Without loss of generality, we can assume
∗ ∗ ∗ ∗ ∗ −γ I
2
P̃1i P̃2i
T = T T Pi T. (28)
i, j, l = 1, . . . , r (23) P̃2i P̃3i
where From (25) and (26), we know that
T
Θl = P̃l − Ω − ΩT Ω=
U X Ăf i B̆f i G4 0 Af i Bf i G−1
3 G4 0
WT WT = .
C̆f i D̆f i 0 I Cf i Df i 0 I

U T Ăi + ᾱB̆f i C̆j Ăf i f B̆f i C̆j 0 (29)
Θ15ij = Θ25ij = Substituting (26)–(29) into (23), we have
X T Ăi + ᾱB̆f i C̆j Ăf i f B̆f i C̆j 0

Θ45ij = [ f D̆f i C̆j 0 ] Θ35ij = [ ᾱD̆f i C̆j − CW C̆f i ] Θl = T T Pl − G − GT T = P̃l − Ω − ΩT
T $ %
−P̃1i −P̃2i U B̆1i + ᾱB̆f i D̆j GT1 Ăi + ᾱGT4 Bf i C̆j GT4 Af i G−13 G4
Θ55i = Θ = Θ15ij =
−P̃2iT
−P̃3i 16ij
X T B̆1i + ᾱB̆f i D̆j GT4 G−T T
GT4 Af i G−1
3 G2 Ăi + ᾱG4 Bf i C̆j 3 G4
' ( $ %
f B̆f i D̆j f GT4 Bf i C̆j 0
Θ26ij = Θ36ij = ᾱD̆f i D̆j − D̆W
f B̆f i D̆j Θ25ij = Θ45ij = [ f Df i C̆j 0 ]
f GT4 Bf i C̆j 0
AW 0
Θ46ij = f D̆f i D̆j Ăi =
0 Ai Θ35ij = [ ᾱDf i C̆j − CW Cf i G−1 3 G4 ]

0 0 BW −P1i −P2i
Θ55ij = T T T
B̆1i =
Bi E1i E2i
D̆f i = Df i C̆i = [ 0 Ci ] −P2i T
−P3i
$ %
D̆i = [ Di F1i F2i ] D̆W = [ 0 0 DW ] . (24) U T B̆1i + ᾱGT4 Bf i D̆j
Θ16ij =
then there exists a fuzzy fault detection filter in the form of GT4 G−T T
3 G2 B̆1i + ᾱG4 Bf i D̆j
$ %
(6), such that the residual system in (13) is stochastically stable f GT4 Bf i D̆j ' (
with the performance γ defined in (11). Moreover, if the afore- Θ26ij = Θ36ij = ᾱDf i D̆j − D̆W
mentioned conditions are satisfied, the matrices for the fault f GT4 Bf i D̆j
detection filter in (6) are given by Θ46ij = f Df i D̆j .
−T −1
Af i Bf i G4 0 Ăf i B̆f i V 0
= (25) Then, one can conclude that the inequality in (23) is equiv-
Cf i Df i 0 I C̆f i D̆f i 0 I
alent to (30), as shown at the bottom of this page, which
with G4 and V being nonsingular matrices and satisfying W = clearly guarantees the inequality in (14). Thus, the proof is
GT4 V . completed.
Proof: Suppose there exist real symmetric positive definite Remark 4: Without loss of generality, we assume the number
matrices P̃l for l = 1, . . . r, matrices U, X, and W satisfy (23). of rows in CW is nC W . Then, the two identity matrices in (3,3)
   
TT 0 0 0 0 0 Πl 0 0 0 GT Āij GT B̄ij T 0 0 0 0 0
 0 TT 0 0 0 0 ∗ Πl 0 0 f GT Ā1ij f G B̄1ij   0
T
T 0 0 0 0
   
 0 0 I 0 0 0 ∗ ∗ −I 0 C̄ij D̄ij   0 0 I 0 0 0
   <0 (30)
 0 0 0 I 0 0 ∗ ∗ ∗ −I f C̄1ij f D̄1ij   0 0 0 I 0 0
   
0 0 0 0 TT 0 ∗ ∗ ∗ ∗ −Pi 0 0 0 0 0 T 0
0 0 0 0 0 I ∗ ∗ ∗ ∗ ∗ −γ I2
0 0 0 0 0 I
and (4,4) blocks in the left side of (14) are nC W -dimensional, and Then, we are in a position to give the fault detection filter
the identity matrix in (6,6) block is (m + p + q)-dimensional, design for the fuzzy system with norm bounded uncertainties.
which are same as those of the identity matrices in (15) and (23). Theorem 3: Consider the fuzzy system in (4). For a given
Remark 5: After the variables Ăf i , B̆f i , C̆f i , and D̆f i are positive constant γ, if there exist n-dimensional matrices
obtained from the LMI in (23), we have to perform the decom-
P̃1l P̃2l
position on the matrix W to obtain the solution in (25). Since P̃l = >0
by most matrix decomposition methods G4 and V cannot be P̃2lT P̃3l
determined uniquely, the solution of (25) is not unique. where n = nf + nW + np , matrices Ăf i , B̆f i , C̆f i , and D̆f i ,
Remark 6: Theorem 2 presents the conditions under which the scalars εij l for any i, j, l = 1, . . . , r, and n-dimensional matrices
residual system is stochastically stable and satisfies the guaran- U, X, and W satisfying the following inequality:
teed performance γ. The result also covers the case of the per- $ %
fect communication links existing between the physical plant Φij l + εij l Q̃Ti Q̃i Ñi
<0 (31)
and the fault detection filter, i.e., there is no data packet dropout ÑiT −εij l I
and ᾱ = 1, with Φij l modified as
  where
Θl 0 Θ15ij Θ16ij
0 N̆i UT 0 0
 ∗ −I Θ35ij Θ36ij  Ñi = N̆i = 0
Φij l =   0 0 XT Ni Ni
∗ ∗ Θ55i 0
 
∗ ∗ ∗ −γ I2 [ 0 Qai ] 0 0
0 0
Q̃i = Q̆i =  ∗ 0 0 
where Θl , Θ55i , Ăi , B̆1i , C̆i , D̆i , and D̆W are defined in (24) 0 Q̆i
T ∗ ∗ [ Qbi 0 0]
U Ăi + B̆f i C̆j Ăf i
Θ15ij = and Φij l is defined in Theorem 2, then there exists a fuzzy
X T Ăi + B̆f i C̆j Ăf i
T fault detection filter in the form of (6), such that the residual
U B̆1i + B̆f i D̆j system is stochastically stable with the performance γ defined
Θ16ij =
X T B̆1i + B̆f i D̆j in (11). Moreover, if the aforementioned condition is satisfied,
the matrices for the fault detection filter in (6) are given by
Θ35ij = [ Df i C̆j − CW C̆f i ] −T −1
' ( Af i Bf i U 0 Ăf i B̆f i V 0
Θ36ij = Df i D̆j − D̆W . = (32)
Cf i Df i 0 I C̆f i D̆f i 0 I
Remark 7: It is noted that (23) is an LMI over both the with G4 and V being nonsingular matrices and satisfying W =
matrix variables and the scalar γ. Among those feasible so- GT4 V .
lutions, the best performance scalar γ can be found by solving Proof: Replacing Ai and Bi in (23) with Ai + Ni Z(k)Qai
an optimization problem in which γ is included as an opti- and Bi + Ni Z(k)Qbi , respectively, we have
mization variable. The minimum [in terms of the feasibility Φij l + Ñi Z̃(k)Q̃i + (Ñi Z̃(k)Q̃i )T < 0
of (23)] attenuation level of the fault detection filter can be
readily obtained by solving the following convex optimization where Z̃(k) is an appropriate dimensioned block-diagonal ma-
problem using LMI Toolbox: minimize γ subject to (23) over trix with entries Z(k). According to Lemma 1, the aforemen-
P̃1l , P̃2l , P̃3l , U, X, W, Ăf i , B̆f i , C̆f i , and D̆f i . tioned inequality holds if
Φij l + ε−1 T T
ij l Ñi Ñi + εij l Q̃i Q̃i < 0
IV. FUZZY FAULT DETECTION FILTER DESIGN
FOR UNCERTAIN SYSTEMS which, by Schur complement, is equivalent to the inequality in
(31). The proof is completed.
In this section, the results obtained previously for nominal
Remark 8: The conditions derived here are based on the basis-
systems will be extended to fuzzy systems with uncertainties
dependent Lyapunov function method, which can potentially
described in (2), i.e., the fuzzy fault detection filter is designed
reduce the conservatism of the results. But the computation cost
for the uncertain fuzzy system in (4), such that the residual
will be increased at the same time, especially when the number
system in (10) is stochastically stable with the performance
of fuzzy rules of the plant is large. One way to solve this problem
defined in (11).
is to try to reduce the rules in modeling of the physical plant.
Before proceeding further, we first give the following lemma
However, when the number of fuzzy rules for some complex
that is needed for our subsequent derivation.
nonlinear systems is large and cannot be reduced in fuzzy mod-
Lemma 1 [31]: Given matrices Φ = ΦT , N, Q, and R =
T eling, we can adopt the quadratic Lyapunov function approach
R > 0 of appropriate dimensions
in solving the fault detection problem. The quadratic approach
Φ + N F Q + QT F T N T < 0 can be found in [3] and [27].
In this paper, we utilize the flexible and powerful LMI tool to
for all F satisfying F T F ≤ R, if and only if there exists a scalar
solve the fault detection problem, which is preferred by many
ε > 0, such that
researchers [1], [33]. The Schur complement and congruent
Φ + ε−1 N N T + εQT RQ < 0. transformation are used to convert the H∞ norm constraints
 
into LMIs feasibility conditions, which are actually convex op- 0.0294 −0.0048 −0.0658
timization problems. Solutions can be determined by solving Ăf 2 =  −0.0027 0.0011 0.0273  D̆f 2 = 0.0368
those optimization problems via MATLAB toolbox. −0.0039 0.0222 0.7146
 
V. ILLUSTRATIVE EXAMPLES 0.0143
B̆f 2 =  −0.0011  C̆f 2 = [ −0.4995 0.0372 0.0041 ]
In this section, two examples are presented to illustrate the −0.0104
usefulness and applicability of the fault detection filter design  
approaches developed previously. 0.2959 −0.0250 −0.0736
W =  −0.0248 0.0033 0.0327 
A. Example 1 −0.0778 0.0330 0.8847
In this example, we will illustrate the applicability of the fault and the guaranteed performance defined in (11) is γ ∗ = 0.1042.
detection filter design method for the nominal fuzzy system. Applying a full-rank factorization on W , we get G4 and V
Consider a tunnel diode circuit system whose model is estab-  
0.1249 −0.0399 −0.9914
lished in [8]. With a sampling time T = 0.02 s, the discrete-time G4 =  0.9893 −0.0711 0.1275 
model is obtained as −0.0755 −0.9967 0.0306

2  
xk +1 = hi (x1k ) (Ai xk + Ei wk ) 0.1151 −0.0360 −0.8876
i=1 V =  0.2845 −0.0207 0.0377 
−0.0000 −0.0004 0.0000

2
yk = hi (x1k ) (Ci xk + Fi wk ) (33) and thus, the fuzzy fault detection filter matrices can be calcu-
i=1 lated by (25)
 
where the state variables are chosen as x1 (t) = vC (t) and 0.8034 −0.2985 −3.1752
x2 (t) = iL (t), and vC (t) and iL (t) are the capacitor voltage Af 1 =  −0.0266 0.1147 0.6210  Df 1 = 0.0369
and inductance current, respectively. The parameter matrices 0.0005 0.0011 0.0089
are given by  
−0.0042
A1 =
0.9887 0.9024
E1 =
0.0093 Bf 1 =  0.0171  Cf 1 = [ −0.0810 −1.7323 4.1194 ]
−0.0180 0.8100 0.0181 0.0004
 
0.9033 0.8617 0.0091 0.7961 −0.2950 0.3848
A2 = E2 =
−0.0172 0.8103 0.0181 Af 2 =  −0.0224 0.1104 0.1722  Df 2 = 0.0368
Ci = [ 1 0] Fi = 1. 0.0005 0.0009 0.0245
 
We assume that there are faults on the capacitor voltage, and the −0.0085
fault matrices are given by Bf 2 =  0.0156  Cf 2 = [ −0.0779 −1.7237 3.4360 ] .
0.0003
0.9887 0.9033
G1 = G2 = . Fig. 1 shows the residual response rk and the response of the
−0.0180 −0.0172
residual evaluation function · T varying as time k when wk =
The aim is to design a fuzzy fault detection filter such that the 0, where the fault is supposed to be
residual system in the form of (13) is stochastically stable and
the performance defined in (11) is guaranteed. 1, 300 ≤ k ≤ 600
fk = (34)
The fault weighting system is in the form of (9) and the 0, else.
matrices are chosen as follows: From the figure, we can see that the designed filter can detect
AW = 0.1 BW = 0.25 CW = 0.5 DW = 0. the fault effectively when it occurs.
Then, we assume the disturbance
We first consider the perfect communication case, i.e., there
rand [0, 1] , 200 ≤ k ≤ 700
is no data packet dropout between the physical plant and the wk = (35)
0, else.
fault detection filter, and thus, ᾱ = 1. By solving the LMI in
Theorem 2, the matrix variables are obtained as Fig. 2 shows the residual response and the residual evaluation
  function response with the disturbance wk , respectively, which
0.0301 −0.0043 −0.0628
indicate that the residual can not only reflect the fault in time, but
Ăf 1 =  −0.0028 0.0010 0.0273  D̆f 1 = 0.0369
also recognize the fault without confusing it with the disturbance
−0.0038 0.0210 0.7217
  wk .
0.0164 Next, we consider the fault detection problem with imperfect
B̆f 1 =  −0.0014  C̆f 1 = [ −0.5023 0.0372 0.0066 ] communication links. The parameter is assumed to be ᾱ = 0.8,
0.0063 which means that there are 20% data lost during the transmission
Fig. 1. Residual response and evaluation function for the nominal system with zero u k and w k (ᾱ = 1).
Fig. 2. Residual response and evaluation function for the nominal system with w k (ᾱ = 1).
from the physical plant to the fault detection filter. By applying sponse, which clearly indicate that the residual can also detect
Theorem 2, the fuzzy fault detection filter matrices are given by the fault without confusing it with the disturbance.
  It is worth noting that the obtained minimum-guaranteed per-
0.7880 −0.4444 10.9624 formance γ ∗ will change as the different values of ᾱ, which is
Af 1 =  −0.0734 0.1514 −8.8659  Df 1 = 0.0240 shown in Table I. From the table, we know that as the values
−0.0007 −0.0010 0.6085 of ᾱ become larger, the higher γ ∗ can be obtained. This is true
 
0.0007 since the larger ᾱ means the less missing measurements; there-
Bf 1 =  −0.0046  Cf 1 = [ 0.1412 1.7282 11.2867 ] fore, the better disturbance attenuation performance γ ∗ can be
0.0002 obtained.
 
0.7986 −0.4315 8.3850
Af 2 =  −0.0754 0.1482 −5.3386  Df 2 = 0.0240 B. Example 2
−0.0005 −0.0008 0.3067 In this example, we will design a fuzzy fault detection filter
  for the fuzzy system with uncertainties. We still consider the
0.0022
fuzzy system in Example 1. For simulation, we assume there
Bf 2 =  −0.0049  Cf 2 = [ 0.1560 1.7534 7.8455 ]
are some uncertainties in the form of (2), and the parameters are
0.0003
given by
and the guaranteed performance is γ ∗ = 0.1216. By assuming
the same fault as in (34) and zeros disturbance wk , Fig. 3 shows 0.25
Ni = Qai = [ 0.1 0] Qbi = 0.1.
the residual response of rk and the evaluation function response, 0.25
which clearly tell us that the fault can be detected effectively
when it appears. First, we consider the perfect communication condition (ᾱ =
Then, assume the same disturbance as that in (35). Fig. 4 1). By solving the inequalities in Theorem 3, the minimum
shows the residual response and the evaluation function re- performance value is obtained as γ ∗ = 0.1043, and the fuzzy
Fig. 3. Residual response and evaluation function for the nominal system with zero u k and w k (ᾱ = 0.8).
Fig. 4. Residual response and evaluation function for the nominal system with w k (ᾱ = 0.8).
TABLE I Next, we consider the robust case with intermittent measure-

MINIMUM GUARANTEED PERFORMANCE γ FOR DIFFERENT VALUES OF ᾱ
ments. Without loss of generality, we suppose ᾱ = 0.8. The min-
imum performance defined in (11) is obtained as γ ∗ = 0.1241,
and the fuzzy fault detection filter matrices are obtained as
 
0.2766 −0.3567 −5.0519
detection filter matrices are Af 1 =  −0.2921 0.6742 −6.2744  Df 1 = 0.0203
  −0.0018 0.0009 0.6778
0.7912 −0.2836 1.5684  
Af 1 =  −0.1096 0.1483 −0.7187  Df 1 = 0.0377 −0.0017
−0.0013 −0.0008 −0.0020 Cf 1 = [ 1.4222 1.0240 10.7865 ] Bf 1 =  −0.0029 
  0.0002
−0.0017  
Bf 1 =  −0.0166  Cf 1 = [ 0.3017 1.6957 4.1770 ] 0.2961 −0.3280 −1.9103
0.0003 Af 2 =  −0.2847 0.6502 −6.2402  Df 2 = 0.0195
  −0.0013 −0.0003 0.4207
0.7660 −0.2853 −0.9327  
Af 2 =  −0.1011 0.1451 −0.9949  Df 2 = 0.0376 −0.0011
−0.0013 −0.0005 0.0236 Bf 2 =  −0.0037  Cf 2 = [ 1.4503 0.9920 8.0067 ] .
  0.0002
0.0014
Bf 2 =  −0.0154  Cf 2 = [ 0.2969 1.6880 3.5792 ] . Fig. 5 shows the residual response and the response of the resid-
0.0003 ual evaluation function. It can be seen that the designed fuzzy
Fig. 5. Residual response and evaluation function for the uncertain system with w k (ᾱ = 0.8).
fault detection filter is also effective for the uncertain fuzzy [10] B. Jiang, M. Staroswiecki, and V. Cocquempot, “H ∞ fault detection filter
system with missing measurements. design for linear discrete-time systems with multiple time delays,” Int. J.
Syst. Sci., vol. 34, no. 5, pp. 365–373, 2003.
[11] M. Bask, A. Johansson, and T. Norlander, “Dynamic threshold generators
VI. CONCLUDING REMARKS for robust fault detection in linear systems with parameter uncertainty,”
Automatica, vol. 42, pp. 1095–1106, 2006.
In this paper, the problem of fault detection for T–S fuzzy [12] J. Lam and S. S. Zhou, “Dynamic output feedback H ∞ control of discrete-
time fuzzy systems: A fuzzy-basis-dependent Lyapunov function ap-
systems with intermittent measurements has been investigated. proach,” Int. J. Syst. Sci., vol. 38, no. 1, pp. 25–37, 2007.
The communication links between the plant and the fault detec- [13] C. Lin, Q. G. Wang, and T. H. Lee, “Stability and stabilization of a class of
tion filter are assumed to be imperfect, and a stochastic variable fuzzy time-delay descriptor systems,” IEEE Trans. Fuzzy Syst., vol. 14,
no. 4, pp. 542–551, Aug. 2006.
satisfying the Bernoulli random binary distribution is utilized [14] Z. Mao, B. Jiang, and P. Shi, “H ∞ fault detection filter design for net-
to model the unreliable communication links. A fuzzy fault worked control systems modelled by discrete Markovian jump systems,”
detection filter has been designed such that, for all data miss- IET Control Theory Appl., vol. 1, no. 5, pp. 1336–1343, 2007.
[15] Z. Mao and B. Jiang, “Fault identification and fault-tolerant control for a
ing conditions, the residual system is stochastically stable and class of network control systems,” Int. J. Innov. Comput., Inf. Control,
preserves a guaranteed performance. The results have been ex- vol. 3, no. 5, pp. 1121–1130, 2007.
tended to the T–S fuzzy systems with time-varying parameter [16] R. Mattone and A. D. Luca, “Relaxed fault detection and isolation: An
application to a nonlinear case study,” Automatica, vol. 42, no. 1, pp. 109–
uncertainties. All the results are formulated in the form of LMIs. 116, 2006.
Two examples have been provided to illustrate the usefulness [17] S. K. Nguang, P. Shi, and S. Ding, “Fault detection for uncertain fuzzy
and applicability of the results. systems: An LMI approach,” IEEE Trans. Fuzzy Syst., vol. 15, no. 6,
pp. 1251–1262, Dec. 2007.
[18] D. C. Oh and J. H. Kim, “A simple frequency weighted model reduction
REFERENCES using structurally balanced truncation: Existence of solutions,” Int. J.
Control, vol. 75, pp. 1190–1195, 2002.
[1] A. Casavola, D. Famularo, and G. Franzè, “A robust deconvolution scheme [19] P. Shi, “Filtering on sampled-data systems with parametric uncertainty,”
for fault detection and isolation of uncertain linear systems: An LMI IEEE Trans. Autom. Control, vol. 43, no. 7, pp. 1022–1027, Jul. 1998.
approach,” Automatica, vol. 41, no. 8, pp. 1463–1472, 2005. [20] P. Shi, E. K. Boukas, and R. K. Agarwal, “Kalman filtering for continuous-
[2] A. Fekih, H. Xu, and F. Chowdhury, “Neural networks based system iden- time uncertain systems with Markovian jumping parameters,” IEEE
tification techniques for model based fault detection of nonlinear systems,” Trans. Autom. Control, vol. 44, no. 8, pp. 1592–1597, Aug. 1999.
Int. J. Innov.. Comput., Inf. Control, vol. 3, no. 5, pp. 1073–1085, 2007. [21] P. Shi and S. K. Nguang, “H ∞ output feedback control of fuzzy system
[3] G. Feng, “A survey on analysis and design of model-based fuzzy control models under sampled measurements,” Comput. Math. Appl., vol. 46,
systems,” IEEE Trans. Fuzzy Syst., vol. 14, no. 5, pp. 676–697, Oct. 2006. no. 5–6, pp. 705–717, 2003.
[4] P. Frank and X. Ding, “Frequency domain approach to optimally robust [22] P. Shi, Y. Xia, G. P. Liu, and D. Rees, “On designing of sliding-mode
residual generation and evaluation for model-based fault diagnosis,” Au- control for stochastic jump systems,” IEEE Trans. Autom. Control, vol. 51,
tomatica, vol. 30, pp. 789–804, 1994. no. 1, pp. 97–103, Jan. 2006.
[5] P. Frank and X. Ding, “Survey of robust residual generation and evaluation [23] I. Skrjanc, S. Blazic, and O. Agamennoni, “Identification of dynamical
methods in observer-based fault detection systems,” J. Process Control, systems with a robust interval fuzzy model,” Automatica, vol. 41, pp. 327–
vol. 7, pp. 403–424, 1997. 332, 2005.
[6] H. Gao and C. Wang, “Delay-dependent robust H ∞ and L 2 –L ∞ filtering [24] A. Stoorvogel, H. Niemann, A. Saberi, and P. Sannuti, “Optimal fault
for a class of uncertain nonlinear time-delay systems,” IEEE Trans. Autom. signal estimation,” Int. J. Robust Nonlinear Control, vol. 12, pp. 697–
Control, vol. 48, no. 9, pp. 1661–1666, Sep. 2003. 727, 2002.
[7] H. Gao and C. Wang, “Robust L 2 – L ∞ filtering for uncertain systems [25] H. Wang and J. Lam, “Robust fault detection for uncertain discrete-time
with multiple time-varying state delays,” IEEE Trans. Circuits Syst. I, systems,” J. Guid. Control Dyn., vol. 25, no. 2, pp. 291–301, 2002.
Fundam. Theory Appl., vol. 50, no. 4, pp. 594–599, Apr. 2003. [26] H. Wang, J. Wang, and J. Lam, “Robust fault detection observer design:
[8] W. Assawinchaichote and S. K. Nguang, “H ∞ filtering for fuzzy dy- Iterative LMI approaches,” J. Dyn. Syst. Meas. Control, vol. 129, no. 1,
namic systems with D stability constraints,” IEEE Trans. Circuits Syst. I, pp. 77–82, 2007.
Fundam. Theory Appl., vol. 50, no. 11, pp. 1503–1508, Nov. 2003. [27] H. O. Wang, K. Tanaka, and M. F. Griffin, “An approach to fuzzy control
[9] I. Izadi, T. Chen, and Q. Zhao, “Norm invariant discretization for sampled- of nonlinear systems: Stability and design issues,” IEEE Trans. Fuzzy
data fault detection,” Automatica, vol. 41, no. 9, pp. 1633–1637, 2005. Syst., vol. 4, no. 2, pp. 14–23, Feb. 1996.
[28] Z. Wang, D. W. C. Ho, and X. Liu, “Variance-constrained filtering for Huijun Gao (M’06) was born in Heilongjiang,
uncertain stochastic systems with missing measurements,” IEEE Trans. China, in 1976. He received the M.S. degree in elec-
Autom. Control, vol. 48, no. 7, pp. 1254–1258, Jul. 2003. trical engineering from Shenyang University of Tech-
[29] Z. Wang, F. Yang, D. W. C. Ho, and X. Liu, “Robust finite-horizon filtering nology, Shenyang, China, in 2001, and the Ph.D. de-
for stochastic systems with missing measurements,” IEEE Signal Process. gree in control science and engineering from Harbin
Lett., vol. 12, no. 6, pp. 437–440, Jun. 2005. Institute of Technology, Harbin, China, in 2005.
[30] A. S. Willsky and H. L. Jones, “A generalized likelihood ratio approach From November 2003 to August 2004, he was a
to the detection and estimation of jumps in linear sytems,” IEEE Trans. Research Associate in the Department of Mechanical
Autom. Control, vol. AC-21, no. 1, pp. 108–112, Feb. 1976. Engineering, University of Hong Kong, Hong Kong.
[31] L. Xie, “Output feedback H ∞ control of systems with parameter uncer- In November 2004, he joined Harbin Institute of
tainty,” Int. J. Control, vol. 63, pp. 741–750, 1996. Technology, where he is currently a Professor. From
[32] P. Zhang, S. X. Ding, G. Z. Wang, and D. H. Zhou, “A frequency domain October 2005 to October 2007, he was a Postdoctoral Researcher in the Depart-
approach to fault detection in sampled-data systems,” Automatica, vol. 39, ment of Electrical and Computer Engineering, University of Alberta, Edmonton,
no. 7, pp. 1303–1307, 2003. AB, Canada. He is an Associate Editor of the Journal of Intelligent and Robotics
[33] M. Zhong, S. X. Ding, J. Lam, and H. Wang, “An LMI approach Systems, Circuits, System and Signal Processing, etc. His current research in-
to design robust fault detection filter for uncertain LTI systems,” Au- terests include network-based control, robust control, and time-delay systems
tomatica, vol. 39, no. 3, pp. 543–550, 2003. and their industrial applications.
[34] S. Zhou, G. Feng, J. Lam, and S. Xu, “Robust H ∞ control for discrete- Prof. Gao is an Associate Editor of the IEEE TRANSACTIONS ON SYSTEMS,
time fuzzy systems via basis-dependent Lyapunov functions,” Inf. Sci., MAN AND CYBERNETICS PART B: CYBERNETICS and the IEEE TRANSACTIONS
vol. 174, no. 3/4, pp. 197–217, 2005. ON INDUSTRIAL ELECTRONICS. He was an outstanding reviewer for the IEEE
[35] S. Xu and J. Lam, “Exponential H ∞ filter design for uncertain Takagi– TRANSACTIONS ON AUTOMATIC CONTROL and Automica in 2008 and 2007, re-
Sugeno fuzzy systems with time delay,” Eng. Appl. Artif. Intell., vol. 17, spectively, and an appreciated reviewer for the IEEE TRANSACTIONS ON SIGNAL
pp. 645–659, 2004. PROCESSING in 2006. He was the recipient of the University of Alberta Dorothy
J. Killam Memorial Postdoctoral Fellow Prize in 2005, the National Outstand-
ing Youth Science Fund in 2008, and the National Outstanding Doctoral Thesis
Award in 2007. He was the corecipient of the National Natural Science Award
Yan Zhao received the B.S. degree in chemical engi- of China in 2008.
neering and equipment control and the M.S. degree
in mechanical engineering from the Inner Mongolia
University of Technology, Hohhot, China, in 2002
and 2005, respectively. She is currently working to-
ward the Ph.D. degree in control science and engi-
neering at Harbin Institute of Technology, Harbin,
China.
Her current research interests include fuzzy con-
trol systems, robust control, and networked control
systems.
James Lam (S’86–M’86–SM’99) received the B.Sc.

degree (with first class) in mechanical engineer-
ing from the University of Manchester, Manchester,
U.K., and the M.Phil. and Ph.D. degrees in con-
trol engineering from the University of Cambridge,
Cambridge, U.K. He received the Ashbury Scholar-
ship, the A.H. Gibson Prize, and the H. Wright Baker
Prize for his academic performance.
He is currently a Professor in the Department of
Mechanical Engineering, University of Hong Kong,
Hong Kong. He is an Associate Editor of the Asian
Journal of Control, International Journal of Systems Science, Journal of Sound
and Vibration, International Journal of Applied Mathematics and Computer Sci-
ence, the IEEE TRANSACTIONS ON SIGNAL PROCESSING, Journal of the Franklin
Institute, Dynamics of Continuous, Discrete and Impulsive Systems (Series B:
Applications and Algorithms), and Automatica. He is also a member of the
Editorial Board of the Institution of Engineering and Technology (IET) Control
Theory and Applications, Open Electrical and Electronic Engineering Journal,
Research Letters in Signal Processing, International Journal of Systems, Con-
trol and Communications, and Journal of Electrical and Computer Engineering.
His current research interests include reduced-order modeling, delay systems,
descriptor systems, stochastic systems, multidimensional systems, robust con-
trol, and filtering. He was an Editor-in-Chief of the Institution of Electrical
Engineers (IEE) Proceedings Control Theory and Applications.
Prof. Lam is a Chartered Mathematician and a Chartered Scientist. He is a
Fellow of the Institute of Mathematics and Its Applications, and the IET. He is
a Scholar and a Fellow of the Croucher Foundation.
Robust Output Feedback Stabilization for Uncertain

Discrete-Time Fuzzy Markovian Jump Systems
With Time-Varying Delays
Yashun Zhang, Shengyuan Xu, and Baoyong Zhang
Abstract—This paper provides a delay-dependent approach to sis for T–S fuzzy systems have been extensively studied; see,
the design of fuzzy dynamic output feedback controllers for un- e.g., [7], [12]–[14], [19], [33], [35], [36], and the references
certain discrete-time fuzzy Markovian jump systems with interval therein. Therefore, during the past few years, fuzzy Markovian
time-varying delays. First, by a fuzzy-basis-dependent and mode-
dependent Lyapunov functional, a stochastic stability condition is jump systems (FMJSs) have received much attention. For ex-
derived by using the Finsler’s lemma. Second, in terms of linear ample, on the basis of a new fuzzy model with two levels of
matrix inequalities (LMIs), a delay-dependent sufficient condition structure, a fuzzy-model-based technique for Markovian jump
is presented, under which there exists a fuzzy output feedback nonlinear systems was introduced in [1]. The design method of
controller such that the resulting closed-loop system is robustly robust H∞ output feedback controllers for FMJSs was proposed
stochastically stable. A desired controller can be constructed when
these LMIs are feasible. Finally, the effectiveness of the proposed in [20]. Recently, the problems of stabilization for continuous-
design method is demonstrated by a simulation example. and discrete-time FMJSs were investigated in [29] and [30], re-
spectively. Very recently, the problems of stability analysis and
Index Terms—Delay-dependent stabilization, fuzzy systems, in-
terval time-varying delays, Markovian jump systems, output feed- controller design for FMJSs were studied in [11] by introduc-
back controllers. ing some slack variables to separate Lyapunov matrices from
system matrices. It should be pointed out that in [1], [11], [29],
and [30] the controllers were designed under the assumption
I. INTRODUCTION that all state variables are available.
ANY practical complex systems may experience abrupt On the other hand, time delays are inherent in a lot of phys-
M changes such as random component failures or repairs,
sudden environmental disturbances, and changes in the intercon-
ical systems. The existence of time delays is often the main
source of instability or poor performance of a control system.
nections of subsystems [15]. These systems can be represented Hence, a great number of results on time-delay systems have
as Markovian jump systems, which involve both time-evolving been presented in [2], [6], [17], [18], [21], [28], and [37], and
and event-driven mechanisms. In Markovian jump systems, each the references therein. Recently, much attention has also been
event governed by a Markov process corresponds to the jump in devoted to the problems of stability analysis and stabilization for
finite operation modes of practical systems. Hence, considerable Markovian jump linear systems with time delays. For instance,
attention has been paid to the study of Markovian jump systems some sufficient conditions for stability analysis were proposed
over the past decades. Many significant results on estimation in [3], [26], and [34] for the continuous case, where stabilizing
and control problems for Markovian jump linear systems have controllers were also designed. The corresponding results for
been reported in [9], [10], [15], [24], and [32], and the references the discrete case can be found in [4], [5], [8], [25], and [31].
therein. It is shown that the delay-dependent results in [4], [8], [31],
However, few results on the problems of stability and sta- and [34] are less conservative than the delay-independent ones
bilization for Markovian jump nonlinear systems have been in [3], [5], [25], and [26], especially when the delays are small.
given because of the difficulties inherent in the analysis of In this paper, we consider the problem of robust output feed-
nonlinear dynamics. Recently, the technique based on Takagi– back stabilization for a class of discrete-time FMJSs with inter-
Sugeno (T–S) fuzzy model has been proved to be a power- val time-varying delays and parameter uncertainties. The pur-
ful method for the control problem of complex nonlinear sys- pose is to provide a delay-dependent condition for the exis-
tems. The problems of stability analysis and controller synthe- tence of a full-order fuzzy dynamic output feedback controller
such that the resulting closed-loop system is robustly stochas-
Manuscript received June 4, 2008; revised October 31, 2008; accepted tically stable. We employ a fuzzy-basis-dependent and mode-
December 23, 2008. Current version published April 1, 2009. This work was dependent Lyapunov functional, which contains full informa-
supported in part by the National Science Foundation for Distinguished Young
Scholars of P. R. China under Grant 60625303, in part by the Specialized
tion on the lower and upper bounds of interval time-varying
Research Fund for the Doctoral Program of Higher Education under Grant delays. By using the Finsler’s lemma, a delay-dependent suffi-
20060288021, in part by the Natural Science Foundation of Jiangsu Province cient condition for the solvability of this problem is obtained in
under Grant BK2008047, and in part by the National Natural Science Founda-
tion of P.R. China under Grant 60850005.
terms of certain linear matrix inequalities (LMIs). By solving
The authors are with the School of Automation, Nanjing University of Science these LMIs, a desired fuzzy dynamic output feedback controller
and Technology, Nanjing 210094, China (e-mail: yashunzhang@gmail.com; can be constructed. A simulation example is given to show the
syxu02@yahoo.com.cn; baoyongzhang@yahoo.com.cn).
Digital Object Identifier 10.1109/TFUZZ.2009.2013203
effectiveness of the proposed design method.
1063-6706/$25.00 © 2009 IEEE
Notations: For real symmetric matrices X and Y , the notation

w
X ≥ Y and X > Y mean that the matrix X − Y is positive y(k) = hi (s(k)) [Ci,l x(k) + Cdi,l x(k − τ (k))] (6)
semidefinite and positive definite, respectively. I denotes the i=1
identity matrix with appropriate dimensions. The superscript where

“T ” represents the transpose. “∗” is used as an ellipsis for terms
g
i (s(k))
that are induced by symmetry. Matrices, if explicitly stated, are hi (s(k)) = w i (s(k)) = µij (sj (k))
assumed to have compatible dimensions for algebra operations. j =1 j (s(k)) j =1
s(k) = [ s1 (k) s2 (k) ··· sg (k) ]

II. PROBLEM FORMULATION
in which, hi (s(k)) is the fuzzy basis function, and µij (sj (k)) is
Consider the following uncertain discrete-time FMJS with
interval time-varying delays: of sj (k) in µij . Then, it can be seen that
the grade of membership
i (s(k)) ≥ 0 and w j =1 j (s(k)) > 0 for all k. Therefore, we
Plant Rule i: IF s1 (k) is µi1 and s2 (k) is µi2 and . . . and w
sg (k) is µig , THEN have that hi (s(k)) ≥ 0 and j =1 hj (s(k)) = 1 for all k. For
the sake of simplicity, we shall use the following notations:
x (k + 1) = [Ai (rk ) + ∆Ai (rk , k)] x (k)
hki = hi (s(k)) hki +1 = hi (s(k + 1))
+ [Adi (rk ) + ∆Adi (rk , k)] x (k − τ (k))
hki −τ = hi (s(k − τ (k))).
+ [Bi (rk ) + ∆Bi (rk , k)] u (k) (1)
Throughout this study, we shall use the following concept of
y (k) = Ci (rk ) x (k) + Cdi (rk ) x (k − τ (k)) (2) stochastic stability:
x (k) = φ (k) , k = −τ2 , −τ2 + 1, . . . , 0 (3) Definition 1: The uncertain FMJS in (1)–(3), with u(k) ≡ 0,
is said to be robustly stochastically stable, if for any initial
where i ∈ S = {1, 2, . . . , w}, and w is the number of IF- condition (φ, r(0)), the following is satisfied
THEN rules; µij is the fuzzy set; x (k) ∈ Rn is the state; N
u (k) ∈ Rp is the control input; y (k) ∈ Rq is the measured
lim E x (k, φ, r(0))x(k, φ, r(0)) < ∞
T
output; s1 (k) , s2 (k) , . . . , sg (k) are the premise variables N →∞
k =1
that do not depend on the input variables u (k) explicitly;
{rk } is a discrete-time Markov process taking values in a fi- where x(k, φ, r(0)) denotes the solution to the system at time k
nite set T = {1, 2, . . . , m} with transition probability matrix under the initial conditions φ and r(0).
Now, we consider the following full-order fuzzy dynamic
Π = {πlf } given by πlf = P r (rk +1 = f |rk = l), where πlf is
output feedback controller for any rk = l, l ∈ T:
the transition probability from mode l at time k to mode f at Control Rule i: IF s1 (k) is µi1 and s2 (k) is µi2 and . . . and
time k + 1, and for all l, f ∈ T, πlf ≥ 0, and m f =1 πlf = 1.
sg (k) is µig , THEN
In (1) and (2), τ (k) is a positive integer denoting the interval
time-varying delay and satisfies 0 < τ1 ≤ τ (k) ≤ τ2 , where τ1 x̂(k + 1) = Aci,l x̂(k) + Bci,l y(k)
and τ2 are positive integers. Ai (rk ), Adi (rk ), Bi (rk ), Ci (rk ),
and Cdi (rk ) are known real-valued matrix functions of rk . u(k) = Cci,l x̂(k)
∆Ai (rk , k), ∆Adi (rk , k), and ∆Bi (rk , k) are unknown matrix where x̂(k) ∈ Rn is the controller state, Aci,l , Bci,l , and Cci,l
functions of rk and are assumed to be of the following form: are matrices to be determined later. Then, we obtain the overall
fuzzy controller, for any rk = l, l ∈ T, as follows:
[ ∆Ai (rk , k) ∆Adi (rk , k) ∆Bi (rk , k) ]

w
= Ei (rk )F (rk , k) [ H1i (rk ) H2i (rk ) H3i (rk ) ] (4) x̂(k + 1) = hki [Aci,l x̂(k) + Bci,l y(k)] (7)
i=1
where Ei (rk ), H1i (rk ), H2i (rk ), and H3i (rk ) are known real-
valued matrices, and F (rk , k) is the unknown time-varying ma-
w
trix satisfying F T (rk , k)F (rk , k) ≤ I for all rk ∈ T. u(k) = hki Cci,l x̂(k). (8)
i=1
For notational simplicity, in the sequel, any matrix as Ω(rk )
will be denoted by Ωl when rk = l, l ∈ T. For example, Ai (rk ) Using the overall fuzzy output feedback controller, we have
is denoted by Ai,l , and ∆Ai (rk , k) is denoted by ∆Ai,l (k), and the following closed-loop system for any rk = l, l ∈ T:
so on.
w
w

Given a pair (x(k), u(k)), we obtain the following final output ξ(k + 1) = hki hkj Āij,l ξ(k) + Ādij,l ξ(k − τ (k))
of the fuzzy system in (1) and (2) for any rk = l, l ∈ T: i=1 j =1
(9)

w
x(k + 1) = hi (s(k)){[Ai,l + ∆Ai,l (k)]x(k) where
i=1
ξ(k) = [ xT (k) x̂T (k) ]T
+ [Adi,l + ∆Adi,l (k)]x(k − τ (k))
Ai,l + ∆Ai,l (k) [Bi,l + ∆Bi,l (k)] Ccj,l
Āij,l =
+ [Bi,l + ∆Bi,l (k)]u(k)} (5) Bcj,l Ci,l Acj,l
ZHANG et al.: ROBUST OUTPUT FEEDBACK STABILIZATION FOR UNCERTAIN DISCRETE-TIME FUZZY MARKOVIAN JUMP SYSTEMS 413

Adi,l + ∆Adi,l (k) 0 Bi,l Ccj,l Adi,l 0n ×5n
Ādij,l = . W2ij,l =
Bcj,l Cdi,l 0 Acj,l Bcj,l Cdi,l 0n ×5n
The purpose is to design a fuzzy dynamic output feedback T
Ēi,l = [ Ei,l 0n ×n ]T
controller in the form of (7) and (8) for the uncertain FMJS with
time delays in (1)–(3) such that the resulting closed-loop system H̄ij,l = [ 0n ×2n H1i,l H3i,l Ccj,l H2i,l 0n ×5n ]T
in (9) is robustly stochastically stable. σ = τ 2 − τ1
 
Φ11 Φ12 0 0 0
III. MAIN RESULTS
 ∗ Φ22 Φ23 S1i −N1i 
 
We first give the following lemma that will be used in the Φij tv ,l =  ∗ ∗ Φ33 S2i −N2i 
 
proof of our main results. ∗ ∗ ∗ −Q2 0
Lemma 1 (Finsler’s lemma) [22]: Consider θ ∈ Rn , H = ∗ ∗ ∗ ∗ −Q3
H ∈ Rn ×n , and A ∈ Rm ×n such that rank(A) = r < n. The
T

m
following conditions are equivalent: Φ11 = πlf Pt,f + τ2 R1 + σR2
1) θT Hθ < 0, ∀θ such that Aθ = 0, θ
= 0; f =1
2) ∃M ∈ Rn ×m , H + MA + AT MT < 0.
Φ12 = −τ2 R1 − σR2
Now, we propose a delay-dependent stochastic stability con-
dition for the closed-loop system in (9). Φ22 = −Pi,l + (σ + 1)Q1i + Q2 + Q3 + M1i
Theorem 1: Given integers τ2 and τ1 with τ2 > τ1 > 0, T
+ M1i + τ2 X11 + σY11 + τ2 R1 + σR2
then for any τ (k) satisfying 0 < τ1 ≤ τ (k) ≤ τ2 , the FMJS
in (9) is robustly stochastically stable if there exist matrices Φ23 = −M1i + M2i
T
+ N1i − S1i + τ2 X12 + σY12
M1i , M2i , N1i , N2i , S1i , S2i , Pi,l > 0, Q1i > 0, i ∈ S, l ∈ T,
Φ33 = −Q1v − M2i − M2i
T T
+ N2i + N2i − S2i
X11 X12
Q2 > 0, Q3 > 0, R1 > 0, R2 > 0, X = ≥ 0,
∗ X22 − S2i
T
+ τ2 X22 + σY22 .
Y11 Y12
Y = ≥ 0, Z, and scalars εij,l > 0, i, j ∈ S, l ∈ T Proof: We first denote
∗ Y22
such that (10) at the bottom of this page, and the following
w
w
matrix inequalities hold for each l ∈ T: Pl (hk ) = hki Pi,l Q1 (hk ) = hki Q1i
  i=1 i=1
Giitv ,l εii,l Z Ēi,l H̄ii,l
w
w
 ∗ −εii,l I 0  < 0, i, t, v ∈ S (11) M1 (h ) =k
hki M1i k
M2 (h ) = hki M2i
∗ ∗ −εii,l I i=1 i=1
 
w
w
X11 X12 M1i k
N1 (h ) = hki N1i k
N2 (h ) = hki N2i
Ψ1 =  ∗ X22 M2i  > 0, i ∈ S (12)
i=1 i=1
∗ ∗ R1
 
w
w
Y11 Y12 S1i S1 (hk ) = hki S1i S2 (hk ) = hki S2i
Ψ2 =  ∗ Y22 S2i  > 0, i ∈ S (13) i=1 i=1
∗ ∗ R2 η(k) = ξ(k + 1) − ξ(k).
 
X11 + Y11 X12 + Y12 N1i Now, we define ξk = (ξ(k), . . . , ξ(k − τ (k))) and choose
Ψ3 =  ∗ X22 + Y22 N2i  >0, i ∈ S (14) the following fuzzy-basis-dependent stochastic Lyapunov
∗ ∗ R 1 + R2 functional:
where
4
V (ξk , rk , k) = Vi (ξk , rk , k) (15)
T
Gij tv ,l = Φij tv ,l + ZWij,l + Wij,l ZT , i=1
Wij,l = [ W1ij,l W2ij,l ] where

w

−In ×n 0 Ai,l V1 (ξk , rk , k) = ξ (k) T
hki Pi (rk ) ξ(k)
W1ij,l =
0 −In ×n Bcj,l Ci,l i=1
 
Gij tv ,l + Gj itv ,l εij,l Z Ēi,l εj i,l Z Ēj,l H̄ij,l H̄j i,l
 ∗ −εij,l I 0 0 0 
 
 ∗ ∗ −εj i,l I 0 0  < 0, 1 ≤ i < j ≤ w, t, v ∈ S. (10)
 
∗ ∗ ∗ −εij,l I 0
∗ ∗ ∗ ∗ −εj i,l I
−τ 1
k −1
k −τ 1 −1 k −τ (k )−1

V2 (ξk , rk , k) = ξ T (i)Q1 (hi )ξ(i) ϕT2 (k)Y ϕ2 (k) = ϕT2 (k)Y ϕ2 (k)
j = −τ 2 i= k + j i=k −τ 2 i=k −τ 2
k −1
k −1
k −τ 1 −1

V3 (ξk , rk , k) = ξ T (i)Q2 ξ(i) + ξ T (i)Q3 ξ(i) + ϕT2 (k)Y ϕ2 (k). (21)
i= k −τ 1 i= k −τ 2 i=k −τ (k )
−1
k −1
On the other hand, it can be verified that
V4 (ξk , rk , k) = η T (i)R1 η(i)
k −1

j = −τ 2 i= k + j
ψ1 (k) = ξ(k) − ξ(k − τ (k)) − η(i) = 0 (22)
−τ
1 −1 k −1
i=k −τ (k )
+ η T (i)R2 η(i).
k −τ (k )−1
j = −τ 2 i= k + j

ψ2 (k) = ξ(k − τ (k)) − ξ(k − τ2 ) − η(i) = 0
Then, define i=k −τ 2
ϕ1 (k) = [ ξ T (k + 1) ξ T (k) ξ T (k − τ (k)) (23)

T k −τ 1 −1
ξ (k − τ1 )
T
ξ (k − τ2 ) ]
T

ψ3 (k) = ξ(k − τ1 ) − ξ(k − τ (k)) − η(i) = 0.
ϕ2 (k) = [ ξ T (k) ξ T (k − τ (k)) ]T i=k −τ (k )
ϕ3 (k, i) = [ ξ T (k) ξ T (k − τ (k)) η T (i) ]T (24)

and for each rk = l, l ∈ T, simple computation yields From (16)–(24), we have, for each rk = l, l ∈ T
E [V1 (ξk +1 , rk +1 )|ξk , rk = l] − V1 (ξk , l) E [V (ξk +1 , rk +1 )|ξk , rk = l] − V (ξk , l)
 
m
4
= ξ T (k + 1)  πlf Pf (hk +1 ) ξ(k + 1) = {E [Vi (ξk +1 , rk +1 )|ξk , rk = l] − Vi (ξk , l)}

f =1 i=1
− ξ T (k)Pl (hk )ξ(k) (16) + 2ϕT2 (k) [ M1T (hk ) M2T (hk ) ]T ψ1 (k)
E [V2 (ξk +1 , rk +1 )|ξk , rk = l] − V2 (ξk , l) + 2ϕT2 (k) [ N1T (hk ) N2T (hk ) ]T ψ2 (k)
≤ (σ + 1)ξ T (k)Q1 (hk )ξ(k) + 2ϕT2 (k) [ S1T (hk ) S2T (hk ) ]T ψ3 (k)
k −1
− ξ T (k − τ (k))Q1 (hk −τ )ξ(k − τ (k)) (17)
≤ ϕT1 (k)Φl (k)ϕ1 (k) − ϕT3 (k, i)Ψ1 ϕ3 (k, i)
E [V3 (ξk +1 , rk +1 )|ξk , rk = l] − V3 (ξk , l) i=k −τ (k )
= ξ (k) (Q2 + Q3 ) ξ(k) − ξ (k − τ1 )Q2 ξ(k − τ1 )

T T
k −τ
1 −1
− ξ T (k − τ2 )Q3 ξ(k − τ2 ) (18) − ϕT3 (k, i)Ψ2 ϕ3 (k, i)

i=k −τ (k )
E [V4 (ξk +1 , rk +1 )|ξk , rk = l] − V4 (ξk , l)
k −τ (k )−1
T

= η (k)(τ2 R1 + σR2 )η(k) − ϕT3 (k, i)Ψ3 ϕ3 (k, i) (25)
k −τ i=k −τ 2
k −1
1 −1
− η T (i)R1 η(i) − η T (i)R2 η(i) where

i= k −τ (k ) i= k −τ (k )  
Φ11 (k) Φ12 0 0 0
k −τ (k )−1
 
 ∗ Φ22 (k) Φ23 (k) S1 (hk ) −N1 (hk ) 
− η T (i)(R1 + R2 )η(i). (19)  
 
i= k −τ 2 Φl (k) =  ∗ ∗ Φ33 (k) S2 (hk ) −N2 (hk ) 
 
Now, we consider the following general case: τ (k)
= τ2 and  
 ∗ ∗ ∗ −Q2 0 
τ (k)
= τ1 . Observe that
∗ ∗ ∗ ∗ −Q3
k −1 k −τ (k )−1

m
ϕT2 (k)Xϕ2 (k) = ϕT2 (k)Xϕ2 (k) Φ11 (k) = πlf Pf (hk +1 ) + τ2 R1 + σR2
i=k −τ 2 i= k −τ 2 f =1
k −1
Φ22 (k) = −Pl (hk ) + (σ + 1)Q1 (hk ) + Q2 + Q3 + M1 (hk )
+ ϕT2 (k)Xϕ2 (k) (20)
i= k −τ (k ) + M1T (hk ) + τ2 X11 + σY11 + τ2 R1 + σR2
Φ23 (k) = −M1 (hk ) + M2T (hk ) + N1 (hk ) − S1 (hk ) where

w
w

+ τ2 X12 + σY12 Wl (k) = hki hkj Wij,l + Ēi,l Fl (k)H̄ij,l
T
.
Φ33 (k) = −Q1 (hk −τ ) − M2 (hk ) − M2T (hk ) + N2 (hk ) i=1 j =1
+ N2T (hk ) − S2 (hk ) − S2T (hk ) + τ2 X22 + σY22 . In addition, from (9), we can deduce, for each l ∈ T
Then, for each l ∈ T Wl (k)ϕ1 (k) = 0.

w
w
w
w Applying Lemma 1 gives
Φl (k) = hki hkj hkt +1 hkv −τ Φij tv ,l .
i=1 j =1 t=1 v =1
ϕT1 (k)Φl (k)ϕ1 (k) < −ϕT1 (k)I1 ϕ1 (k) (29)
Now, we suppose that the conditions in (10)–(14) are satisfied. for each l ∈ T. Therefore, it follows from (12)–(14), (25), and
By using the Schur complements, it follows from (10) and (11) (29) that, for each l ∈ T
that there exists a sufficiently small scalar > 0 satisfying the E [V (ξk +1 , rk +1 )|ξk , rk = l] − V (ξk , l) < −ξ T (k)ξ(k).
following inequalities for each l ∈ T and t, v ∈ S (30)
T
Gij tv ,l + Gj itv ,l + εij,l Z Ēi,l Ēi,l ZT In the case when τ (k) = τ2 or τ (k) = τ1 , we can follow a
similar line as earlier to have (30). Hence, from (30) and the
T
+ εj i,l Z Ēj,l Ēj,l Z T + ε−1 T
ij,l H̄ij,l H̄ij,l result in [4], it can be deduced that the uncertain FMJS in (9) is
robustly stochastically stable.
+ ε−1
j i,l H̄j i,l H̄j i,l < −2I1 ,
T
1≤i<j≤w (26) Remark 1: In the proof of Theorem 1, the Lyapunov functional
T
Giitv ,l + εii,l Z Ēi,l Ēi,l ZT (15) is not only mode-dependent but also fuzzy-basis-dependent
owing to the use of the matrices Pl (hk ) and Q1 (hk ). More-
+ ε−1
ii,l H̄ii,l H̄ii,l < −I1 ,
T
i∈S (27) over, this functional also depends on both the upper and lower
bounds of the interval time-varying delay by introducing some
where additional terms into V4 (ξk , rk , k). Hence, this Lyapunov func-
 
02n ×2n 0 0 tional contains information on the Markovian jump parameters,
I1 =  ∗ I2n ×2n 0 . the fuzzy-basis functions, and the interval time-varying delay.
∗ ∗ 06n ×6n This implies that the condition in Theorem 1 is less conservative
than those derived by using mode-independent or fuzzy-basis-
From (26) and (27), it can be seen that, for each l ∈ T independent Lyapunov functionals.
w
w w Remark 2: By using the Finsler’s lemma, we derive the anal-
k +1 k −τ
ht hv hki hki Giitv ,l + εii,l Z Ēi,l Ēi,l
T
ZT ysis conditions (10)–(14) in an enlarged space. This method
t=1 v =1 i=1 guarantees that the Lyapunov matrices are separated from the
−1
w w system matrices. Therefore, it is easy to realize the controller
+ ε−1 H̄ H̄ T
ii,l ii,l ii,l + I1 + hki hkj synthesis by utilizing Theorem 1.
i=1 j = i+1 We are now in a position to give the main result on the design
of the robust fuzzy output feedback controller for the FMJS in
× Gij tv ,l + Gj itv ,l + εij,l Z Ēi,l Ēi,l
T
ZT
(1)–(3).
T
+ εj i,l Z Ēj,l Ēj,l Z T + ε−1 T
ij,l H̄ij,l H̄ij,l
Theorem 2: Given integers τ2 and τ1 with τ2 > τ1 > 0, then
there exists a fuzzy output feedback controller in the form
−1 T
of (7) and (8) such that the resulting closed-loop system in
+ εj i,l H̄j i,l H̄j i,l + 2I1 (9) is robustly stochastically stable for any τ (k) satisfying
0 < τ1 ≤ τ (k) ≤ τ2 , if there exist matrices M̄1i , M̄2i , N̄1i , N̄2i ,

w
w
S̄1i , S̄2i , K1i,l , K2i,l , K3i,l , P̄i,l > 0, Q̄1i> 0, i ∈ S, l ∈ T,
= Φl (k) + hki hkj ZWij,l + Wij,l
T
ZT
X̄11 X̄12
i=1 j =1 Q̄2 > 0, Q̄3 > 0, R̄1 > 0, R̄2 > 0, X̄ = ≥ 0,
∗ X̄22
T
+ εij,l Z Ēi,l Ēi,l ZT Ȳ11 Ȳ12
Ȳ = ≥ 0, Z1 , L1 , U , and scalars εij,l > 0, i, j ∈
∗ Ȳ22
+ ε−1 T
ij,l H̄ij,l H̄ij,l + I1 < 0. S, l ∈ T such that (31) at the bottom of the next page and the
By Proposition 2.2 in [16], it is easy to see that following LMIs hold for each l ∈ T
 
T Ḡiitv ,l εii,l Ξi,l Υii,l Γ1i,l 0
Z Ēi,l Fl (k)H̄ij,l + H̄ij,l FlT (k)Ēi,l
T
ZT
 ∗ −εii,l I 0 0 εii,l I 
 
≤ εij,l Z Ēi,l Ēi,l
T
Z T + ε−1 T
ij,l H̄ij,l H̄ij,l
 ∗ ∗ −εii,l I 0 0  < 0, i, t, v ∈ S
 
∗ ∗ ∗ −I 0
for each l ∈ T and i, j ∈ S. Therefore, for each l ∈ T ∗ ∗ ∗ ∗ −I
Φl (k) + ZWl (k) + WlT (k)Z T < −I1 (28) (32)
 
X̄11 X̄12 M̄1i Σ1ij,l = [ 0n ×n ATdi,l Z1 + Cdi,l
T T
K1j,l 0n ×8n ]T
 ∗ X̄22 M̄2i  > 0, i ∈ S (33)
∗ ∗ R̄1 Σ2i,l = [ 0n ×n T
Ei,l Z1 0n ×8n ]T
 
Ȳ11 Ȳ12 S̄1i Σ3 = [ 0n ×4n L1 0n ×5n ]T
 ∗ Ȳ22 S̄2i  > 0, i ∈ S (34) Γ2ij,l = [ Σ1ij,l + Σ1j i,l Σ̂1ij,l Σ̂2ij,l
∗ ∗ R̄2
  Σ2i,l Σ2j,l Σ3 Σ̂3 Σ̂4ij,l ]
X̄11 + Ȳ11 X̄12 + Ȳ12 N̄1i T
 ∗ X̄22 + Ȳ22 N̄2i  > 0, i∈S (35) Σ̂1ij,l = [ 0n ×n (Ci,l − Cj,l )T (K1j,l − K1i,l )T 0n ×8n ]
∗ ∗ R̄1 + R̄2 T
Σ̂2ij,l = [ 0p×n (Bi,l − Bj,l )T Z1 0p×8n ]
where Σ̂3 = [ 0n ×2n L1 0n ×7n ]T
Ḡij tv ,l = Φ̄ij tv ,l + W̄ij,l Σ̂4ij,l = [ 0p×2n K2j,l − K2i,l 0p×7n ]T

W̄1 W̄2ij,l W̄3ij,l Γ3ij,l = [ εij,l In ×n 0n ×n ]
W̄ij,l =
∗ 0 0
Γ4ij,l = [ 0n ×n εj i,l In ×n ]
−L1 − LT1 −I − U T
W̄1 =
∗ −Z1 − Z1T σ = τ 2 − τ1 .

Ai,l L1 + Bi,l K2j,l Ai,l Furthermore, a desired fuzzy output feedback controller is given
W̄2ij,l =
K3j,l Z1T Ai,l + K1j,l Ci,l in the form of (7) and (8) with the parameters as follows:

Adi,l L1 Adi,l 0n ×4n
W̄3ij,l = Aci,l = Z2−T K3i,l − Z1T Ai,l L1 − K1i,l Ci,l L1
0 Z1T Adi,l + K1j,l Cdi,l 0n ×4n
  − Z1T Bi,l K2i,l L−1
2 (36)
Φ̄11 Φ̄12 0 0 0
 ∗ Φ̄22 Φ̄23 S̄1i −N̄1i  Bci,l = Z2−T K1i,l , Cci,l = K2i,l L−1
2 , i ∈ S, l ∈ T (37)
 
Φ̄ij tv ,l =  ∗ ∗ Φ̄33 S̄2i −N̄2i 
 
∗ ∗ ∗ −Q̄2 0 where Z2 and L2 are some nonsingular matrices satisfying
∗ ∗ ∗ ∗ −Q̄3 Z2T L2 = U − Z1T L1 .

m Proof: Similar to [23], we choose the relaxed variable Z
Φ̄11 = πlf P̄t,f + τ2 R̄1 + σ R̄2 introduced in Theorem 1 as
f =1
Z̄ T Z1 Z3
Φ̄12 = −τ2 R̄1 − σ R̄2 Z=
08n ×2n
Z̄ =
Z2 Z4
.
Φ̄22 = −P̄i,l + (σ + 1)Q̄1i + Q̄2 + Q̄3 + M̄1i + M̄1i
T
Define
+ τ2 X̄11 + σ Ȳ11 + τ2 R̄1 + σ R̄2

Φ̄23 = −M̄1i + M̄2i + N̄1i − S̄1i + τ2 X̄12 + σ Ȳ12 L1 L3 L1 I
Z̄ −1 =
T
L=
L2 L4 L2 0
Φ̄33 = −Q̄1v − M̄2i − M̄2i
T T
+ N̄2i + N̄2i − S̄2i − S̄2i
T
+ τ2 X̄22 + σ Ȳ22 and for each l ∈ T and i ∈ S, denote

T
Ξi,l = [ Ei,l 0n ×9n ]T U = Z1T L1 + Z2T L2 Θ = diag (L, L, L, L, L)
Υij,l = [ 0n ×2n H1i,l L1 + H3i,l K2j,l H1i,l K1i,l = Z2T Bci,l K2i,l = Cci,l L2
T
H2i,l L1 H2i,l 0n ×4n ] K3i,l = Z1T Ai,l L1 + K1i,l Ci,l L1 + Z1T Bi,l K2i,l
Γ1i,l = [ Σ1ii,l Σ2i,l Σ3 ] + Z2T Aci,l L2
 
Ḡij tv ,l + Ḡj itv ,l εij,l Ξi,l εj i,l Ξj,l Υij,l Υj i,l Γ2ij,l 0
 ∗ −εij,l I 0 0 0 0 Γ3ij,l 
 
 ∗ ∗ −εj i,l I 0 0 0 Γ4ij,l 
 
 ∗ ∗ ∗ −εij,l I 0 0 0  < 0, 1 ≤ i < j ≤ w, t, v ∈ S (31)
 
 ∗ ∗ ∗ ∗ −εj i,l I 0 0 
 
∗ ∗ ∗ ∗ ∗ −I 0
∗ ∗ ∗ ∗ ∗ ∗ −I
P̄i,l = LT Pi,l L Q̄1i = LT Q1i L Q̄2 = LT Q2 L TABLE I

MODES OF THE PARAMETERS a 1 , a 2 , AND b
Q̄3 = LT Q3 L M̄1i = LT M1i L M̄2i = LT M2i L
N̄1i = LT N1i L N̄2i = LT N2i L S̄1i = LT S1i L
S̄2i = LT S2i L R̄1 = LT R1 L R̄2 = LT R2 L
X̄11 = LT X11 L X̄12 = LT X12 L X̄22 = LT X22 L
Ȳ11 = LT Y11 L Ȳ12 = LT Y12 L Ȳ22 = LT Y22 L. Then, it can be seen that, for each l ∈ T and t, v ∈ S, (41) holds
Then, for each l ∈ T, pre- and postmultiplying (11) by if the following inequality holds for each l ∈ T and t, v ∈ S
diag(ΘT , I2n ×2n ) and its transpose, respectively, we obtain  
Ḡij tv ,l + Ḡj itv ,l εij,l Ξi,l εj i,l Ξj,l Υij,l Υj i,l
   ∗ −εij,l I 0 0 0 
Ḡiitv ,l εii,l Ξi,l Υii,l  
   ∗ ∗ −εj i,l I 0 0 
 ∗ −εii,l I 0  + Λ1 ΛT2 + Λ2 ΛT1 < 0,  
∗ ∗ ∗ −εij,l I 0
∗ ∗ −εii,l I ∗ ∗ ∗ ∗ −εj i,l I
i, t, v ∈ S (38) + [ Λ3 Λ 4 ] [ Λ3 Λ4 ]T < 0, 1≤i<j≤w (42)
where which, by using the Schur complements, implies (31). Pre- and
Σ1ii,l Σ2i,l Σ3 010n ×n
Λ1 = Λ2 = postmultiplying (12)–(14) by diag(LT , LT , LT ) and its trans-
02n ×n 02n ×n 02n ×n Σ4i,l pose, respectively, we obtain (33)–(35). Therefore, we derive
Σ4i,l = [ εii,l In ×n 0n ×n ]T . the conditions in (31)–(35).
Remark 3: Theorem 2 provides a delay-dependent condition
Note that (38) holds for each l ∈ T if the following inequality for the existence of robust fuzzy dynamic output feedback con-
holds for each l ∈ T and i, t, v ∈ S trollers for uncertain discrete-time FMJSs. It is worth pointing
 
Ḡiitv ,l εii,l Ξi,l Υii,l T out that (31)–(35) are in the form of LMIs that are easy to be
  Λ1 solved. Thus, our main result in Theorem 2 is effective for robust
 ∗ −ε ii,l I 0  + [ Λ 1 Λ 2 ] < 0.
ΛT2 output feedback stabilization of uncertain discrete-time FMJSs
∗ ∗ −εii,l I with time-varying delays.
(39)
IV. SIMULATION EXAMPLE
Applying the Schur complements to (39), we obtain (32).
Next, we observe that the following equation holds for each In this section, we provide a simulation example to illustrate
l ∈ T and i, j ∈ S, the design approach of the fuzzy output feedback controller
T developed in the previous section. Consider the following un-
Z1 Ai,l L1 + K1j,l Ci,l L1 + Z1T Bi,l K2j,l + Z2T Acj,l L2
certain discrete-time Markovian jump nonlinear system with

+ Z1T Aj,l L1 + K1i,l Cj,l L1 + Z1T Bj,l K2i,l + Z2T Aci,l L2 interval time-varying delays:
= K3i,l + K3j,l + Z1T (Bi,l − Bj,l ) (K2j,l − K2i,l ) x1 (k + 1) = −x21 (k) + a1 x2 (k) + a2 x1 (k)x1 (k − τ (k))
+ (K1j,l − K1i,l ) (Ci,l − Cj,l ) L1 . (40) + 0.02c1 (k)x1 (k)x2 (k) + u1 (k)
Now, for each l ∈ T, pre- and postmultiplying (10) by + 0.01c2 (k)x2 (k − τ (k))
diag(ΘT , I4n ×4n ) and its transpose, respectively, and using − 0.2x1 (k)x2 (k − τ (k)) (43)
(40), we have, for t, v ∈ S
  x2 (k + 1) = 0.1x1 (k) + bx2 (k) − 0.1x1 (k)x2 (k − τ (k))
Ḡij tv ,l + Ḡj itv ,l εij,l Ξi,l εj i,l Ξj,l Υij,l Υj i,l
 ∗ −εij,l I 0 0 0  + 0.02c1 (k)x2 (k) + 0.5u2 (k) (44)
 
 ∗ ∗ −εj i,l I 0 0 
  y(k) = 0.6x1 (k) (45)
∗ ∗ ∗ −εij,l I 0
∗ ∗ ∗ ∗ −εj i,l I where the parameters a1 , a2 , and b have two modes as shown
in Table I; τ (k) is an interval time-varying delay satisfying
+ Λ3 ΛT4 + Λ4 ΛT3 < 0, 1 ≤ i < j ≤ w, (41)
0 < τ1 ≤ τ (k) ≤ τ2 ; c1 (k) and c2 (k) are uncertain parameters
where satisfying c1 (k) ∈ [−1, 1] and c2 (k) ∈ [−1, 1]. The nonlinear
Σ1ij,l + Σ1j i,l Σ̂1ij,l Σ̂2ij,l Σ2i,l Σ2j,l system switches between the two modes, and the transition
Λ3 =
04n ×n 04n ×n 04n ×p 04n ×n 04n ×n probability matrix is given by

Σ3 Σ̂3 Σ̂4ij,l 010n ×n 010n ×n 0.75 0.25
Λ4 = Π= .
04n ×n 04n ×n 04n ×p Σ5ij,l Σ̂5ij,l 0.30 0.70
Σ5ij,l = [ εij,l In ×n 0n ×3n ]T
Similar to [27], we assume that x1 (k) ∈ [−0.5, 0.5] and
Σ̂5ij,l = [ 0n ×n εij,l In ×n 0n ×2n ]T . set the membership functions as µ11 (x1 (k)) = 12 (1 − 2x1 (k)),
µ21 (x1 (k)) = 12 (1 + 2x1 (k)). Then, we represent the Marko-

vian jump nonlinear time-delay system in (43)–(45) as the fol-
lowing T–S model:
Plant rule i: IF x1 (k) is µi1 (x1 (k)), THEN
x(k + 1) = [Ai,l + ∆Ai,l (k)]x(k) + [Adi,l + ∆Adi,l (k)]

× x(k − τ (k)) + [Bi,l + ∆Bi,l (k)]u(k)
y(k) = Ci x(k), i = 1, 2, l = 1, 2
where

0.5 0.3 0.5 0.4
A1,1 = A1,2 =
0.1 1.0 0.1 1.06

−0.5 0.3 −0.5 0.4
A2,1 = A2,2 =
0.1 1.0 0.1 1.06

−0.05 0.1 −0.07 0.1
Ad1,1 = Ad1,2 =
0 0.05 0 0.05

0.05 −0.1 0.07 −0.1
Ad2,1 = Ad2,2 =
0 −0.05 0 −0.05

1.0 0
B1,1 = B2,1 = B1,2 = B2,2 = Fig. 1. Operation mode, time-varying delay, and state responses of the open-
0 0.5 loop system.
C1,1 = C2,1 = C1,2 = C2,2 = [ 0.6 0].
The uncertain parameters ∆Ai,l (k), ∆Adi,l (k), and ∆Bi,l (k),
for i = 1, 2 and l = 1, 2 can be represented in the form of (4)
with

−0.05 0.1 0.05 0.1
E11 = E12 = E21 = E22 =
0.1 0 0.1 0

0 0.2 0 0
H11 = H12 = H21 = H22 =
0 1 0 0.1

0 0
H31 = H32 = .
0 0
The aim is to develop a fuzzy output feedback controller

such that the resulting closed-loop system is robustly stochas-
tically stable. The comparison between the conditions in (31)–
(35) and those obtained by replacing the Lyapunov functional
in (15) with a mode-independent and fuzzy-basis-independent
Lyapunov functional will be given by finding the maximum al-
lowable size of τ2 with τ1 = 2 for the aforesaid robust control
problem. By using the sufficient conditions obtained by mode-
independent and fuzzy-basis-independent Lyapunov function-
als, we obtain that the maximum allowable size of τ2 with
τ1 = 2 is 3. However, using Theorem 2, we obtain that the cor-
responding maximum value is 4. This means that the conditions Fig. 2. Operation mode, time-varying delay, and control results of the closed-
loop system.
in Theorem 2 are less conservative than those derived by us-
ing mode-independent and fuzzy-basis-independent Lyapunov
functionals for this example. Then, choosing τ1 = 2, τ2 = 4 and

solving the LMIs in (31)–(35), we obtain the parameters of the −0.0473 −0.0126 1.6787
Ac1,2 = Bc1,2 =
fuzzy output feedback controller as follows: 0.8994 −0.4884 −6.7662

0.0472 −0.0740 0.9509 0.6172 −0.1875 −5.5554
Ac1,1 = Bc1,1 = Ac2,1 = Bc2,1 =
0.8470 −0.4661 −6.5443 0.6597 −0.3377 −5.1108

0.5431 −0.1490 −4.9470 [10] C. E. de Souza and M. D. Fragoso, “H ∞ control for linear systems with
Ac2,2 = Bc2,2 = Markovian jumping parameters,” Control Theory Adv. Technol., vol. 9,
0.6978 −0.3407 −5.3518 no. 2, pp. 457–466, 1993.
[11] J. Dong and G. Yang, “An LMI-based approach for state feedback con-
−0.0524 0.0897 −0.0561 0.1115 troller design of Markovian jump nonlinear systems,” in Proc. 16th IEEE
Cc1,1 = Cc1,2 =
0.0242 0.4374 0.0016 0.4498 Int. Conf. Control Appl., Singapore, Oct.2007, pp. 1516–1521.
[12] J. Dong and G. Yang, “State feedback control of continuous-time T–S
0.0961 0.0516 0.0972 0.0663 fuzzy systems via switched fuzzy controllers,” Inf. Sci., vol. 178, no. 6,
Cc2,1 = Cc2,2 = . pp. 1680–1695, 2008.
0.0451 0.3995 0.0287 0.4056
[13] G. Feng, “Controller synthesis of fuzzy dynamic systems based on piece-
wise Lyapunov functions,” IEEE Trans. Fuzzy Syst., vol. 11, no. 5,
Now, we set the initial conditions as r(0) = 1 and φ(k) = pp. 605–612, Oct. 2003.
[0.2, −0.2]T , k = −4, −3, . . . , 0. We further assume that [14] G. Feng, “Stability analysis of discrete-time fuzzy dynamic systems based
c1 (k) = cos(k) and c2 (k) = sin(k). on piecewise Lyapunov functions,” IEEE Trans. Fuzzy Syst., vol. 12, no. 1,
pp. 22–28, Feb. 2004.
The simulations of the operation mode, the time-varying de- [15] Y. Ji and H. J. Chizeck, “Controllability, stabilizability, and continuous-
lay, and the state responses of the open-loop system are given in time Markovian jump linear quadratic control,” IEEE Trans. Autom.
Fig. 1(a), Fig. 1(b), and Fig. 1(c), respectively. Now, we apply Control., vol. 35, no. 7, pp. 777–788, Jul. 1990.
[16] X. Li and C. E. de Souza, “Criteria for robust stability and stabilization
the designed fuzzy dynamic output feedback controller in the of uncertain linear systems with state-delay,” Automatica, vol. 33, no. 9,
form of (7) and (8) to the nonlinear system in (43)–(45). The pp. 1657–1662, 1997.
state responses of the resulting closed-loop system are shown [17] C. Lin, Q. Wang, and T. Lee, “Stabilization of uncertain fuzzy time-delay
systems via variable structure control approach,” IEEE Trans. Fuzzy Syst.,
in Fig. 2(c) under the operation mode in Fig. 2(a) and the time- vol. 13, no. 6, pp. 787–798, Dec. 2005.
varying delay in Fig. 2(b). The control law is also given in Fig. [18] C. Lin, Q. Wang, and T. Lee, “Stability and stabilization of a class of fuzzy
2(d). These results show that the designed fuzzy output feed- time-delay descriptor systems,” IEEE Trans. Fuzzy Syst., vol. 14, no. 4,
pp. 542–551, Aug. 2006.
back controller can effectively stabilize the uncertain Markovian [19] C. Lin, Q. Wang, T. Lee, and Y. He, LMI Approach to Analysis and Control
jump nonlinear time-delay system in (43)–(45). of Takagi–Sugeno Fuzzy Systems with Time Delay. Berlin, Germany:
Springer-Verlag, 2007.
[20] S. K. Nguang, W. Assawinchaichote, P. Shi, and Y. Shi, “Robust H ∞
V. CONCLUSION control design for uncertain fuzzy systems with Markovian jumps: An
LMI approach,” in Proc. Amer. Control Conf., Portland, OR, Jun. 2005,
We have studied the problem of delay-dependent robust out- pp. 1805–1810.
put feedback stabilization for a class of uncertain discrete-time [21] Y. Niu and D. W. Ho, “Robust observer design for Itô stochastic time-delay
FMJSs with interval time-varying delays and parameter uncer- systems via sliding mode control,” Syst. Control Lett., vol. 55, no. 10,
pp. 781–793, 2006.
tainties. By using the Finsler’s lemma, delay-dependent suffi- [22] I. R. Petersen and C. V. Hollot, “A Riccati equation approach to the
cient conditions for the robust stabilization problem have been stabilization of uncertain linear systems,” Automatica, vol. 22, no. 4,
given in terms of LMIs. A desired fuzzy dynamic output feed- pp. 397–411, 1986.
[23] J. Qiu, G. Feng, and J. Yang, “Delay-dependent output feedback guar-
back controller can be constructed by solving these LMIs. A anteed cost control for uncertain discrete-time switched delay sys-
simulation example has also demonstrated the effectiveness of tems,” in Proc. IEEE Int. Conf. Control Autom., May, 2007, pp. 847–
the design method. 852.
[24] P. Shi and E. K. Boukas, “H ∞ -control for Markovian jumping linear
systems with parametric uncertainty,” J. Optim. Theory Appl., vol. 95,
REFERENCES no. 1, pp. 75–99, 1997.
[25] P. Shi, E. K. Boukas, and R. K. Agarwal, “Control of Markovian jump
[1] N. S. D. Arrifano and V. A. Oliveira, “Robust H ∞ fuzzy control approach discrete-time systems with norm bounded uncertainty and unknown de-
for a class of Markovian jump nonlinear systems,” IEEE Trans. Fuzzy lay,” IEEE Trans. Autom. Control, vol. 44, no. 11, pp. 2139–2144, Nov.
Syst., vol. 14, no. 6, pp. 738–754, Dec. 2006. 1999.
[2] M. Basin, E. Sanchez, and R. Martinez-Zuniga, “Optimal linear filtering [26] P. Shi, M. S. Mahmoud, J. Yi, and A. Ismail, “Worst case control of
for systems with multiple state and observation delays,” Int. J. Innovative uncertain jumping systems with multi-state and input delay information,”
Comput. Inform. Control, vol. 3, no. 5, pp. 1309–1320, 2007. Inf. Sci., vol. 176, no. 2, pp. 186–200, 2006.
[3] K. Benjelloun and E. K. Boukas, “Mean square stochastic stability of [27] K. Tanaka and H. O. Wang, Fuzzy Control Systems Design and Analysis:
linear time-delay system with Markovian jumping parameters,” IEEE A Linear Matrix Inequality Approach. New York: Wiley, 2001.
Trans. Autom. Control., vol. 43, no. 10, pp. 1456–1460, Oct. 1998. [28] S. Tong, W. Wang, and L. Qu, “Decentralized robust control for uncer-
[4] E. K. Boukas and Z. K. Liu, “Robust H ∞ control of discrete-time Marko- tain T–S fuzzy large-scale systems with time-delay,” Int. J. Innovative
vian jump linear systems with mode-dependent time-delays,” IEEE Trans. Comput., Inform. Control, vol. 3, no. 3, pp. 657–672, 2007.
Autom. Control., vol. 46, no. 12, pp. 1918–1924, Dec. 2001. [29] H. Wu and K. Cai, “Mode-independent robust stabilization for uncer-
[5] Y.-Y. Cao and J. Lam, “Stochastic stabilizability and H ∞ control for tain Markovian jump nonlinear systems via fuzzy control,” IEEE Trans.
discrete-time jump linear systems with time delay,” J. Franklin Inst., Syst., Man, Cybern., Part B, Cybern., vol. 36, no. 3, pp. 509–519, Jun.
vol. 336, no. 8, pp. 1263–1281, 1999. 2006.
[6] B. Chen, J. Lam, and S. Xu, “Memory state feedback guaranteed cost [30] H. Wu and K. Cai, “Robust fuzzy control for uncertain discrete-time
control for neutral delay systems,” Int. J. Innovative Comput. Inform. nonlinear Markovian jump systems without mode observations,” Inf.
Control, vol. 2, no. 2, pp. 293–303, 2006. Sci., vol. 177, no. 6, pp. 1509–1522, 2007.
[7] B. Chen, X.-P. Liu, S.-C. Tong, and C. Lin, “Observer-based stabilization [31] J. Xiong and J. Lam, “Stabilization of discrete-time Markovian jump
of T–S fuzzy systems with input delay,” IEEE Trans. Fuzzy Syst., vol. 16, linear systems via time-delayed controllers,” Automatica, vol. 42, no. 5,
no. 3, pp. 652–663, Jun. 2008. pp. 747–753, 2006.
[8] W. Chen, Z. Guan, and P. Yu, “Delay-dependent stability and H ∞ control [32] J. Xiong, J. Lam, H. Gao, and D. W. Ho, “On robust stabilization of Marko-
of uncertain discrete-time Markovian jump systems with mode-dependent vian jump systems with uncertain switching probabilities,” Automatica,
time delays,” Syst. Control Lett., vol. 52, no. 5, pp. 361–376, 2004. vol. 41, no. 5, pp. 897–903, 2005.
[9] O. L. V. Costa and R. P. Marques, “Mixed H 2 /H ∞ -control of discrete- [33] S. Xu and J. Lam, “Robust H ∞ control for uncertain discrete-time-delay
time Markovian jump linear systems,” IEEE Trans. Autom. Control., fuzzy systems via output feedback controllers,” IEEE Trans. Fuzzy Syst.,
vol. 43, no. 1, pp. 95–100, Jan. 1998. vol. 13, no. 1, pp. 82–93, Feb. 2005.
[34] S. Xu, J. Lam, and X. Mao, “Delay-dependent H ∞ control and filtering Baoyong Zhang received the B.Sc. degree in mathe-
for uncertain Markovian jump systems with time-varying delays,” IEEE matics in 2003 and the M.Sc. degree in control theory
Trans. Circuits Syst. I, Reg. Papers, vol. 54, no. 9, pp. 2070–2077, Sep. in 2006, both from Qufu Normal University, Qufu,
2007. China. He is currently working toward the Ph.D. de-
[35] G. Yang and J. Dong, “Control synthesis of singularly perturbed fuzzy gree at the School of Automation, Nanjing University
systems,” IEEE Trans. Fuzzy Syst., vol. 16, no. 3, pp. 615–629, Jun. of Science and Technology, Nanjing, China.
2008. His current research interests include robust con-
[36] S. Zhou, J. Lam, and W. Zheng, “Control design for fuzzy systems based trol and filtering, nonlinear systems, time-delay sys-
on relaxed nonquadratic stability and H ∞ performance conditions,” IEEE tems, and stochastic systems.
Trans. Fuzzy Syst., vol. 15, no. 2, pp. 188–199, Apr. 2007.
[37] S. Zhou and T. Li, “Robust stabilization for delayed discrete-time fuzzy
systems via basis-dependent Lyapunov–Krasovskii function,” Fuzzy Sets
Syst., vol. 151, no. 1, pp. 139–153, 2005.
Yashun Zhang received the B.E. degree in automa-

tion in 2003 and the M.E. degree in control theory
and application in 2006, both from Hefei University
of Science and Technology, Hefei, China. He is cur-
rently working toward the Ph.D. degree at the School
of Automation, Nanjing University of Science and
Technology, Nanjing, China.
His current research interests include fuzzy con-
trol, sliding mode control, and nonlinear control.
Shengyuan Xu received the B.Sc. degree from

Hangzhou Normal University, Hangzhou, China, in
1990, the M.Sc. degree from Qufu Normal Univer-
sity, Qufu, China in 1996, and the Ph.D. degree
from Nanjing University of Science and Technology,
Nanjing, China, 1999.
From 1999 to 2000, he was a Research Asso-
ciate in the Department of Mechanical Engineering,
University of Hong Kong, Hong Kong. From Decem-
ber 2000 to November 2001, and December 2001 to
September 2002, he was a Postdoctoral Researcher
in CESAME at the Université Catholique de Louvain, Belgium, and the Depart-
ment of Electrical and Computer Engineering, University of Alberta, Canada,
respectively. From September 2002 to September 2003, and September 2003
to September 2004, he was a William Mong Young Researcher and an Hon-
orary Associate Professor, respectively, both in the Department of Mechanical
Engineering, University of Hong Kong. Since November 2002, he has been a
Professor with the Department of Automation, Nanjing University of Science
and Technology. His current research interests include robust filtering and con-
trol, singular systems, time-delay systems, nonlinear systems, neural networks
and multidimensional systems.
Prof. Xu is a member of the Editorial Boards of the Multidimensional Sys-
tems and Signal Processing and an Associate Editor of the Circuits Systems
and Signal Processing. He was a recipient of the National Excellent Doctoral
Dissertation Award in the year 2002 from the Ministry of Education of China.
In the year 2006, he obtained a grant from the National Science Foundation
for Distinguished Young Scholars of China. He was awarded a Cheung Kong
Professorship in the year 2008 from the Ministry of Education of China.
Representation of Uncertain Multichannel Digital

Signal Spaces and Study of Pattern Recognition
Based on Metrics and Difference Values on Fuzzy
n-Cell Number Spaces
Guixiang Wang, Peng Shi, Senior Member, IEEE, and Paul Messenger
Abstract—In this paper, we discuss the problem of character- erties of probability functions. Since then, both the numbers and
ization for uncertain multichannel digital signal spaces, propose the problems in relation to them (see, for example, [3]–[6], [11],
using fuzzy n-cell number space to represent uncertain n-channel [16], and [19]–[21]) have been widely studied. With the develop-
digital signal space, and put forward a method of constructing such
fuzzy n-cell numbers. We introduce two new metrics and concepts ment of theories and applications of fuzzy numbers, this concept
of certain types of difference values on fuzzy n-cell number space becomes more and more important. Kaleva [7], however, used a
and study their properties. Further, based on the metrics or differ- special type of n-dimensional fuzzy number, whose sets of cuts
ence values appropriately defined, we put forward an algorithmic are all hyperrectangles. In 2002, we carefully studied the special
version of pattern recognition in an imprecise or uncertain environ- type of n-dimensional fuzzy number and called it fuzzy n-cell
ment, and we also give practical examples to show the application
and rationality of the proposed techniques. number in [14] and [15]. It has been demonstrated that the fuzzy
n-cell number is used much more conveniently than general
Index Terms—Difference values, fuzzy n-cell numbers, metrics, n-dimensional fuzzy numbers in theoretical investigations and
n-dimensional fuzzy vectors, pattern recognition, uncertain mul-
tichannel digital signals. in some fields of application in [14], [15], and [17]. On the other
hand, n-dimensional fuzzy vector is also an important concept,
which is the Cartesian product of n 1-D fuzzy numbers. In 1985,
I. INTRODUCTION Kaufmann and Gupta [8] had already studied fuzzy vectors.
T IS known that in a precise or certain environment, mul- Soon after, Miyakawa et al. [9], Nakamura [10], and Ramik and
I tichannel digital signals can be represented by elements of
multidimensional Euclidean space, i.e., crisp multidimensional
Nakamura [12] also studied the problems of theories and appli-
cations in relation to fuzzy vectors. In 1997, Butnariu [1] studied
vectors. If, however, we wish to study multichannel digital sig- methods of solving optimization problems and linear equations
nals in an imprecise or uncertain environment, then the signals in the space of fuzzy vectors. Recently, Wang et al. [14] showed
themselves are imprecise or have no certain bound, and it be- that fuzzy n-cell numbers and n-dimensional fuzzy vectors can
comes unwise to use crisp multidimensional vectors to represent represent each other and obtained the representations of the joint
them. In this paper, we recommend using fuzzy n-cell numbers membership function and the edge membership functions of a
to represent imprecise or uncertain multichannel digital signals fuzzy n-cell number.
and put forward a method of constructing such fuzzy n-cell In a previous paper [15], we defined a metric DL on the fuzzy
numbers. n-cell number space and studied its properties. We again stud-
The concept of general fuzzy numbers was introduced by ied this type of metric in [14] with regard to two fuzzy n-cell
Chang and Zadeh [2] in 1972 with the consideration of the prop- numbers in the form of n-dimensional fuzzy vectors. Although
metric DL can be more conveniently used in applications and
theoretical investigations, it has some shortcomings, i.e., it has a
Manuscript received July 16, 2008; revised October 15, 2008; accepted
November 29, 2008. First published January 9, 2009; current version published tendency to be rougher, and cannot really characterize the degree
April 1, 2009. This work was supported in part by the Natural Science Founda- of difference of two fuzzy n-cell numbers in some applications
tions of China under Grant 60772006, by the Natural Science Foundations of (see Example 3.1 in Section III of this paper). In this paper, in
Zhejiang Province, China, under Grant Y7080044, and by the Engineering and
Physical Sciences Research Council, U.K., under Grant EP/F0219195. order to discuss the problem of pattern recognition in an impre-
G. Wang is with the Institute of Operational Research and Cyber- cise or uncertain environment based on degree of difference, we
netics, Hangzhou Dianzi University, Hangzhou 310018, China (e-mail: define two new metrics and some concepts of difference values
g.x.wang@hdu.edu.cn).
P. Shi is with the Department of Computing and Mathematical Sciences, Uni- on fuzzy n-cell number space, which may better characterize
versity of Glamorgan, Pontypridd CF37 1DL, U.K. He is also with the Institute the degree of difference of two fuzzy n-cell numbers in some
for Logistics and Supply Chain Management, School of Science and Engineer- applications, and study their properties.
ing, Victoria University, Melbourne, Vic. 8001, Australia, and also with the
School of Mathematics and Statistics, University of South Australia, Adelaide, It is well known that pattern recognition is an important field
S.A. 5095, Australia (e-mail: pshi@glam.ac.uk). of research. In this aspect, research achievements are many (for
P. Messenger is with the Department of Computing and Mathematical example, see [13]). In this paper, as applications of the metrics
Sciences, University of Glamorgan, Pontypridd CF37 1DL, U.K. (e-mail:
pmesseng@glam.ac.uk). and difference values (defined by us), we study the problem of
Digital Object Identifier 10.1109/TFUZZ.2008.2012352 pattern recognition in an imprecise or uncertain environment,
1063-6706/$25.00 © 2009 IEEE
put forward an algorithmic version of pattern recognition based some uncertain attributes based on statistical data. About the
on the metrics or difference values (defined by us) of fuzzy algorithmic version of such fuzzy n-cell numbers, we can see
n-cell numbers, and also give examples to show the application the first or second step of the algorithmic version in Section V.
and rationality of the method. Example 2.1: It is well known that different kinds of terrain
This paper is organized as follows. In Section II, we give or landcover possess different reflections of the electromagnetic
an example to show how to set up fuzzy n-cell numbers to spectrum. Based on this principle, one can set up a method to
represent imprecise or uncertain multichannel digital signals. In recognize the category of landcover, a challenging remote sens-
Section III, we define two new metrics and study their properties. ing classification problem, using spectral and terrain features for
In Section IV, we introduce concepts of difference values of two vegetation classification in some zone. In remote sensing classi-
fuzzy n-cell numbers and examine their properties. In Section V, fication, the colligation of all species covering a zone of 4500 m2
an algorithmic version of pattern recognition is given based on can be boiled down to an element of remote sensing space. We
the metrics or the difference values defined by us and examples use “Korean pine accounts for the main part” to denote forest
are also given to show the application and rationality of the that mainly contains Korean pines. Because in different “Korean
method. Finally, in Section VI, we give a brief conclusion of pine accounts for the main part” areas, there are many different
this paper. factors, such as the difference of the density of Korean pines,
of the species and quantity of other plants, of the physiognomy,
etc., the values of reflections of the electromagnetic spectrum
II. REPRESENTATIONS OF UNCERTAIN MULTICHANNEL are also different. Therefore, “Korean pine accounts for the main
DIGITAL SIGNALS part” should not be a certain crisp value but a fuzzy set without
A fuzzy set of the Euclidean space Rn is a function u: Rn → certain bound. So, using a fuzzy number to represent the spec-
[0, 1]. For fuzzy set u, we denote [u]r = {x ∈ Rn : u(x) ≥ r} tral sensitivity level of the “Korean pine accounts for the main
for r ∈ [0, 1] and [u]0 = {x ∈ Rn : u(x) > 0} (the closure of part” is more suitable than using a crisp number. Suppose that
{x ∈ Rn : u(x) > 0}). If u is a normal and fuzzy convex fuzzy we use four wave bands: MSS-4, MSS-5, MSS-6, and MSS-7.
set of Rn , u(x) is upper semicontinuous, and [u]0 is com- We take ten samples and acquire the following data for some
pact, then we call u an n-dimensional fuzzy number and de- zone of “Korean pine accounts for the main part”:
note the n-dimensional fuzzy number space by E n . If u ∈ E,
and for each r ∈ [0, 1], [u]r is a hyperrectangle, i.e., there ex- MSS-4 MSS-5 MSS-6 MSS-7
ist ui (r), ui(r) ∈ R with ui (r) ≤ ui (r), (i = 1, 2, . . . , n) such sample 1 15.01 13.30 40.50 19.37
that [u]r = ni=1 [ui (r), ui (r)], then we call u a fuzzy n-cell sample 2 15.60 12.56 38.81 16.35
number and denote the fuzzy n-cell number space by L(E n ). An sample 3 15.82 12.79 37.70 18.16
n-dimensional fuzzy vector is an ordered class (u1 , u2 , . . . , un ), sample 4 14.90 11.70 35.50 14.75
where ui ∈ E (i.e., E 1 ), i = 1, 2, . . . , n. We have shown in [14] sample 5 16.10 13.80 42.10 20.75
that fuzzy n-cell numbers and n-dimensional fuzzy vectors can sample 6 13.80 11.94 32.10 15.54
represent each other, and as the representation is unique, L(E n ) sample 7 15.90 10.98 30.87 14.29
and the n-dimensional fuzzy vector space (i.e., the Cartesian sample 8 16.82 13.67 37.64 18.62
n sample 9 15.50 12.58 36.10 18.02

product E × E × · · · × E) may be regarded as identical. sample 10 15.38 12.48 34.08 17.45.
When exploring and discussing some quantity, properties, or
We can directly work out the following means µi (i =
laws of movement of phenomena/objects in the physical world,
1, 2, 3, 4) and standard deviations σi (i = 1, 2, 3, 4) from the
it is essential for us to establish the description space of them.
data:
For instance, when the quantity in question is only the one with
a single factor, we can take it as a dot in real number field R, MSS-4 MSS-5 MSS-6 MSS-7
i.e., the space of quantities corresponding to single factor can be µi : µ1 = 15.46 µ2 = 12.58 µ3 = 36.54 µ4 = 17.33
described by 1-D Euclidean space R. Similarly, we can describe σi : σ1 = 1.22 σ2 = 0.88 σ3 = 3.55 σ4 = 2.08.
the quantities with n factors, using n-dimensional Euclidean
space Rn . However, in the physical world, many phenomena From the means and the standard deviations, with
are imprecise or uncertain (such as having no certain bound).
ui (xi )
When the quantity discussed by us possesses some imprecise

or uncertain attributes, it is unsuitable that we still use Rn to  xi − (µi − 2σi )

 , if xi ∈ [µi − 2σi , µi ] ∩ (0, +∞)
represent the space of the quantities (see Remark 2.1). It is  2σi
our opinion that using the fuzzy n-cell number space discussed = (µi + 2σi ) − xi

 , if xi ∈ (µi , µi + 2σi ] ∩ (0, +∞)
in [14] and [15] to describe the quantities with some uncertain 
 2σi
factors and discuss these quantities in this n-dimensional fuzzy 0, if xi ∈
/ [µi −2σi , µi +2σi ] ∩ (0, +∞)
vector space is a more suitable method to reveal the objective i = 1, 2, 3, 4
laws of things in physical world (see Remark 2.1).
In the following example, we demonstrate how we construct we can define four triangular model 1-D fuzzy numbers
a fuzzy n-cell number to represent a quantity that possesses u1 , u2 , u3 , and u4 , which correspond to MSS-4, MSS-5, MSS-6,
WANG et al.: REPRESENTATION OF UNCERTAIN MULTICHANNEL DIGITAL SIGNAL SPACES AND STUDY 423


and MSS-7, respectively  (x3 − 36.54)2
v3 (x3 ) = exp − , if x3 ∈ (0, +∞)
 x − 13.02 
25.21


1
, if x1 ∈ [13.02, 15.46] 0, if x3 ∈
/ (0, +∞)

 2.44
 

u1 (x1 ) = 17.9 − x1 ,  (x4 − 17.33)2
 if x1 ∈ (15.46, 17.9] v4 (x4 ) = exp − , if x4 ∈ (0, +∞)


 2.44 8.65

 0, 
if x1 ∈
/ [13.02, 17.9] 0, if x4 ∈
/ (0, +∞)
 x − 10.82


2
, if x2 ∈ [10.82, 12.58] and obtain the membership function of the fuzzy four-cell

 1.76
 number v = (v1 , v2, v3 , v4 ) determined by v1 , v2 , v3 , and v4
u2 (x2 ) = 14.34 − x2 , if x2 ∈ (12.58, 14.34] as v(x1 , x2 , x3 , x4 ) = min{v1 (x1 ), v2 (x2 ), v3 (x3 ), v4 (x4 )},

 (x1 , x2 , x3 , x4 ) ∈ R4 . Then, the fuzzy four-cell number v can


1.76
 0, if x2 ∈
/ [10.82, 14.34] also be used to represent the “Korean pine accounts for the main
 x − 29.44 part.”


3
, if x3 ∈ [29.44, 36.54] Remark 2.1: Of course, if the quantity to describe is pre-

 7.1
 cise and certain, we should use a crisp multidimensional vec-
u3 (x3 ) = 43.64 − x3 , if x3 ∈ (36.54, 43.64] tor to represent it. However, if the quantity to describe is




7.1 imprecise and uncertain, such as “Korean pine accounts for
 0, if x3 ∈
/ [29.44, 43.64] the main part,” then using a fuzzy n-cell number to rep-
 x − 13.17 resent it is better than using a crisp n-dimensional vector.


4
, if x4 ∈ [13.17, 17.33] If we narrowly use a crisp multidimensional vector, such as

 4.16
 (15.46, 12.58, 36.54, 17.33) (i.e., the mean vector), to repre-
u4 (x4 ) = 21.49 − x4 , if x4 ∈ (17.33, 21.49] sent “Korean pine accounts for the main part,” then it can-




4.16
not clearly tell us the relationship of “Korean pine accounts
 0, if x4 ∈
/ [13.17, 21.49]. for the main part” and the zone whose value of reflection of
electromagnetic spectrum is (15.16, 12.80, 37.50, 16.79) since
(15.16, 12.80, 37.50, 16.79) = (15.46, 12.58, 36.54, 17.33). If
By [14, Th. 3.1 and Th. 3.2], we know that u1 , u2 , u3 , and we use fuzzy n-cell number v = (v1 , v2, v3 , v4 ) to represent
u4 determine a fuzzy four-cell number u = (u1 , u2, u3 , u4 ) and it, then we can almost affirm that the zone whose value
the membership function of u is is (15.16, 12.80, 37.50, 16.79) is a part of “Korean pine ac-
counts for the main part” since v(15.16, 12.80, 37.50, 16.79) =
u(x1 , x2 , x3 , x4 ) = min{u1 (x1 ), u2 (x2 ), u3 (x3 ), u4 (x4 )} min(0.94, 0.94, 0.93, 0.93) = 0.93, i.e., the degree of the zone
that is “Korean pine accounts for the main part” is 0.93.
(x1 , x2 , x3 , x4 ) ∈ R4 .
Then, u can be used to represent “Korean pine accounts for the III. METRICS ON FUZZY n-CELL NUMBER SPACE
main part.” Diamond and Kloeden [3] studied the metric dp ( · , · ) [note
Likewise, from the means and the standard deviations, ac- that in this paper, we rewrite dp ( · , · ) as Dp ( · , · )] on general
cording to n-dimensional fuzzy number space E n , which is defined by
1
Dp (u, v) = ( 0 [d([u]r , [v]r )]p dr)1/p for any u, v ∈ E n , and

point out that the metric Dp is complete.
v (x )= exp − (xi −µi ) , if x ∈ (0, +∞)
2
i i i
We studied the metrics D and DL on L(E n ) in [15], but
2σi2 i = 1, 2, 3, 4 the two metrics seem to be “rough” in certain applications (see

0, if xi ∈
/ (0, +∞) Example 3.1). In the following, other metrics are defined on
L(E n ), which better reveal the difference between two different
uncertain quantities (see Example 3.1). Their properties are also
we can also define four Gaussian model 1-D fuzzy num- discussed such that they may be used appropriately.
bers v1 , v2 , v3 , and v4 , which correspond to MSS-4, MSS-5, We denote LC(Rn )= {A : there exist ai≤ bi , i = 1,
MSS-6, and MSS-7, respectively 2, . . . , n such that A = ni=1 [ai , bi ]}, where ni=1 [ai , bi ] is
the Cartesian product [a1 , b1 ] × [a2 , b2 ] × · · · × [an , bn ].

Theorem 3.1: We define mappings
 (x1 − 15.46)2
v1 (x1 ) = exp − , if x1 ∈ (0, +∞)
2.98 dˆα : LC(Rn ) × LC(Rn ) → [0, +∞)

0, if x1 ∈
/ (0, +∞)


 (x2 − 12.58)2 and
v2 (x2 ) = exp − , if x2 ∈ (0, +∞)
1.56

0, if x2 ∈
/ (0, +∞) d˜α : LC(Rn ) × LC(Rn ) → [0, +∞)
by we see that 4)–6) of the theorem hold for dˆα . For d˜α , we can

n similarly prove that 4)–6) of the theorem also hold.
dˆα (A, B) = αi max{|ai − bi |, |ai − bi |} Theorem 3.2: We define mappings
i=1
D̂α ,p : L(E n ) × L(E n ) → [0, +∞)
and

n
|ai − bi | + |ai − bi | and
d˜α (A, B) = αi
i=1
2 D̃α ,p : L(E n ) × L(E n ) → [0, +∞)
by
for any A = ni=1 [ai , ai ] ∈ LC(Rn ) and B = ni=1 [bi , bi ]
∈ LC(Rn ), where α = (α1 , α2 , . . . , αn ) satisfies
1/p
1
n
i=1α = 1 and α ≥ 0, i = 1, 2, . . . , n. Then, for any D̂α ,p (u, v) = [r · dˆα ([u]r , [v]r )]p dr
i i
A = ni=1 [ai , ai ], B = ni=1 [bi , bi ], C = ni=1 [ci , ci ] in 0
LC(Rn ) and each k ∈ R, dˆα and d˜α satisfy and

1) dˆα (A, B) = dˆα (B, A) and d˜α (A, B) = d˜α (B, A);
1 1/p
2) dˆα (A, B) ≥ 0 and d˜α (A, B) ≥ 0; D̃α ,p (u, v) = [r · d˜α ([u]r , [v]r )]p dr
0
3) dˆα (A, B) = 0 ⇔ A = B ⇔ d˜α (A, B) = 0; i.e.
4) dˆα (A, B) ≤ dˆα (A, C) + dˆα (C, B) and d˜α (A, B) ≤
1
n
d˜α (A, C) + d˜α (C, B); D̂α ,p (u, v) = r αi max{|ui (r)−vi (r)|, |ui (r)
5) dˆα (A + C, B + C) = dˆα (A, B) and d˜α (A + C, B + C) 0 i=1
= d˜α (A, B); p 1/p
6) dˆα (kA, kB) = |k | dˆα (A, B) and d˜α (kA, kB) = |k |
− vi (r)|} dr
d˜α (A, B).
Proof: We show only proofs 4)–6) (the other proofs are easy).
From and
n
1 1
n
dˆα (A, B) = αi max{|ai − bi |, |ai − bi |} D̃α ,p (u, v) = r αi (|ui (r) − vi (r)|
i=1 2 0 i=1

n p 1/p
≤ αi max{|ai −ci |+|ci −bi |,
+ |ui (r) − vi (r)|) dr
i=1
|ai −ci | + |ci − bi |} for any (u, v) ∈ L(E n ) × L(E n ), where p ≥ 1, and α =
n

n (α1 , α2 , . . . , αn ) satisfies i=1 αi = 1 and αi > 0, i =
≤ αi (max{|ai − ci |, |ai − ci |} 1, 2, . . . , n. Then, for any u, v, w ∈ L(E n ) and each k ∈
i=1 R, D̂α ,p and D̃α ,p satisfy
+ max{|ci − bi |, |ci − bi |}) 1) D̂α ,p (u, v) = D̂α ,p (v, u) and D̃α ,p (u, v) = D̃α ,p (v, u);
2) D̂α ,p (u, v) ≥ 0 and D̃α ,p (u, v) ≥ 0;
= dˆα (A, C) + dˆα (C, B)
3) D̂α ,p (u, v) = 0 ⇔ u = v ⇔ D̃α ,p (u, v) = 0;

n
4) D̂α ,p (u, v) ≤ D̂α ,p (u, w) + D̂α ,p (w, v) and D̃α ,p (u, v)
dˆα (A + C, B + C) = αi max{|ai + ci − bi + ci |,
≤ D̃α ,p (u, w) + D̃α ,p (w, v);
i=1
5) D̂α ,p (u + w, v + w) = D̂α ,p (u, v) and D̃α ,p (u + w, v +
|ai + ci − bi + ci |} w) = D̃α ,p (u, v);

n
6) D̂α ,p (ku, kv) = |k|D̂α ,p (u, v) and D̃α ,p (ku, kv) =
= αi max{|ai + ci − bi − ci |, |k|D̃α ,p (u, v).
i=1
Proof: It is obvious that 1) and 2) of the theorem hold.
|ai + ci − bi − ci |} By the definition of D̂α ,p , it is obvious that u = v ⇒
D̂α ,p (u, v) = 0. Otherwise, let D̂α ,p (u, v) = 0. Then, we have
= dˆα (A, B) 1 n
i=1 αi max{|ui (r) − vi (r)|, |ui (r) − vi (r)|}] dr =
p
0 [r

n
0. By taking note of αi > 0, we see that
dˆα (kA, kB) = αi max{|kai −kbi |, |kai − kbi |}
r( ni=1 αi max{|ui (r)− vi (r)|, |ui (r) − vi (r)|}) = 0 holds
i=1
for r almost everywhere on [0, 1]. Further, we have that

n n
i=1 αi max{|ui (r) − vi (r)|, |ui (r) − vi (r)|} = 0 holds for
= αi max{|kai −kbi |, |kai − kbi |}
r almost everywhere on [0, 1], so we can see that ui (r) = vi (r)
i=1
and ui (r) = vi (r) hold for r almost everywhere on [0, 1] for
= |k|dˆα (A, B) i = 1, 2, . . . , n. Therefore, we obtain that [u]r = [v]r holds for

r almost everywhere on [0, 1], so we know that u = v holds  x − 2,
 if x ∈ [2, 3]
by [18, Lemma 2.1]. Thus, D̂α ,p (u, v) = 0 ⇔ u = v holds. vi (x) = 4 − x, if x ∈ (3, 4] i = 1, 2
Likewise, we can prove that D̃α ,p (u, v) = 0 ⇔ u = v, so 3) of 

0, if x ∈
/ [2, 4]
the theorem holds. From Theorem 3.1, we have 

1 n 
 x, if x ∈ [0, 1]
D̂α ,p (u, v) = r αi max{|ui (r) w1 (x) = 2 − x, if x ∈ (1, 2]
0 i=1


0, if x ∈
/ [0, 2]
p 1/p 
− vi (r)|, |ui (r) − vi (r)|} dr  x − 2,
 if x ∈ [2, 3]
w2 (x) = 4 − x, if x ∈ (3, 4]
1/p 

1 0, if x ∈
/ [2, 4].
≤ (r[dˆα ([u]r , [w]r ) + dˆα ([w]r , [v]r )])p dr
0
Then, we know that ui (r) = r, ui (r) = 2 − r, vi (r) =
1/p
1 2 + r, vi (r) = 4 − r (i = 1, 2), w1 (r) = r, w1 (r) = 2 − r,
≤ [r · dˆα ([u]r , [w]r )]p dr w2 (r) = 2 + r, and w2 (r) = 4 − r for r ∈ [0, 1]. From the def-
0 initions of DL , we have DL (u, v) = 2 = DL (u, w), i.e., DL
1/p cannot tell us the difference of DL (u, v) and DL (u, w), so
1
+ [r · dˆα ([w] , [v] )] dr
r r p we say that DL seems to be “rough” (similar proof for D).
0 However, as a matter of fact, DL (u, v) and DL (u, w) should
have some difference. Taking α = (1/2, 1/2), from the defi-
= D̂α ,p (u, w) + D̂α ,p (w, v). nitions of D̃α ,p , we can obtain D̃α ,p (u, v) = 2/(1 + p)1/p >
1/(1 + p)1/p = D̃α ,p (u, w), which accords with the fact.
The proofs of D̃α ,p (u, v) ≤ D̃α ,p (u, w) + D̃α ,p (w, v), 5)
If we restrain the metric Dp (i.e., dp ( · , · ) defined by
and 6) can be similarly proved.
Remark 3.1: From Theorems 3.1 and 3.2, we know that dˆα , d˜α Diamond in [3], see paragraph 1 of this section) on general
and D̂α ,p , D̃α ,p are metrics on LC(Rn ) and L(E n ), respec- n-dimensional fuzzy number space E n into L(E n ), then it also
tively, and satisfy translation invariance and absolute homo- becomes a metric on L(E n ). In the following, we give the rela-
geneity. Also, from the factor r of the integrands in the defini- tionships of the metrics D̂α ,p , D̃α ,p , and Dp .
tions of D̂α ,p (u, v) and D̃α ,p (u, v), we can see that the bigger Theorem 3.3: Metrics D̂α ,p , D̃α ,p , and Dp satisfy
the degrees of the points are, which belong to the fuzzy n-cell 1) (1/2)D̂α ,p ≤ D̃α ,p ≤ D̂α ,p ≤ Dp ≤ D, i.e., (1/2)D̂α ,p
numbers u and v, the greater the effects on the metric of u and (u, v) ≤ D̃α ,p (u, v) ≤ D̂α ,p (u, v) ≤ Dp (u, v) ≤ D(u, v)
v. This is true in reality. for any u,v ∈ L(E n ) (D is discussed
in [15]);
Example 3.1: Let u, v, and w be the two-cell numbers defined 2) D̂α ,p ≤ 1/(p + 1)1/p DL ≤ 1/(p + 1)1/p D, i.e.,

by u = (u1 , u2 ), v = (v1 , v2 ), and w = (w1 , w2 ), where D̂α ,p (u, v) ≤ (1/(p + 1)1/p )DL (u, v) ≤ 1/(p + 1)1/p
 D(u, v) for any u, v ∈ L(E n ).

 x, if x ∈ [0, 1]
Proof: For any u, v ∈ L(E n ) and r ∈ [0, 1], by the definitions
ui (x) = 2 − x, if x ∈ (1, 2] i = 1 and 2. of dˆα and d˜α , we have the following equation shown at the


0, if x ∈/ [0, 2] bottom of this page.
1
n
1ˆ
dα ([u]r , [v]r ) = αi max{|ui (r) − vi (r)|, |ui (r) − vi (r)|}
2 2 i=1
1 |ui (r) − vi (r)| + |ui (r) − vi (r)| + ||ui (r) − vi (r)| − |ui (r) − vi (r)||
n
= αi
2 i=1 2

n
|ui (r) − vi (r)| + |ui (r) − vi (r)| + |ui (r) − vi (r)| + |ui (r) − vi (r)|
≤ αi
i=1
4

n
|ui (r) − vi (r)| + |ui (r) − vi (r)|
≤ αi (= d˜α ([u]r , [v]r ))
i=1
2

n
2 max{|ui (r) − vi (r)|, |ui (r) − vi (r)|}
≤ αi
i=1
2
= dˆα ([u]r , [v]r ).
From this, we can directly obtain (1/2)D̂α ,p (u, v) ≤ and [15, Th. 4.5], we can see that 2) of the theorem
D̃α ,p (u, v) ≤ D̂α ,p (u, v). holds.
√ Remark 3.2: From 1) of Theorem 3.3, we see that
By [15, Th. 4.4], we know dL ≤ d ≤ ndL , where
d (A, B) = max1≤i≤n {|ai − bi |, |ai − bi |} for any A = (1/2)D̂α ,p ≤ D̃α ,p ≤ D̂α ,p , i.e., D̂α ,p and D̃α ,p are equivalent,
Ln n so we know that D̂α ,p and D̃α ,p induce equivalent topologies
i=1 [ai , ai ] ∈ LC(R ) and B = i=1 [bi , bi ] ∈ LC(R ).
n n
Therefore, for any u, v ∈ L(E ), we have

n on L(E n ) by the knowledge of topological space.

1 n
D̂α ,p (u, v) = r αi max{|ui (r) − vi (r)|, IV. DIFFERENCE VALUES ON FUZZY n-CELL NUMBER SPACE
0 i=1 In Section III, we discussed metrics on L(E n ). But some-
p 1/p times, these have some shortcomings demonstrating the differ-
|ui (r) − vi (r)|} dr ence of two objects. For example, we consider that the degree
of difference of 1 and 2 is bigger than the degree of difference

n
1
of 1010 and 1010 + 1 though their metrics (Euclidean metric)
≤ r αi max {|uj (r) − vj (r)|, both measure 1. A mapping from the Cartesian product X × X
0 1≤j ≤n
i=1
of a set X into R needs to satisfy stronger conditions so that it
p 1/p
can become a metric, and this brings limitations in some appli-
|uj (r) − vj (r)|} dr cations. The measure used to characterize the differences does

p 1/p not need to satisfy all metric conditions, for example, when we
1
n
set up a method of pattern recognition based on the principle of
r r
= rdL ([u] , [v] ) αi dr
0 minimal difference (i.e., the principle of maximal likelihood),
i=1
1/p the measure used to characterize the differences does not need
1
p
to satisfy all metric conditions. To conveniently set up methods
≤ [d([u]r , [v]r )] dr (=Dp (u, v)) of pattern recognition using fuzzy n-cell numbers, we intro-
0
duce the concepts of difference values on L(E n ) and study the
p 1/p properties.
1
≤ r
sup d([u] , [v] ) dr r
Let u ∈ L(E n ) and α = (α1 , α2 , . . . , αn ) ∈ Rn satisfy
n
0 r ∈[0,1]
i=1 αi = 1 and αi ≥ 0 (i = 1, 2, . . . , n). We denote
n 1
1/p Mα (u) = i=1 αi 0 r[ui (r) + ui (r)]dr, M̄ (u) = Mα (u) as
1
= [D(u, v)] dr p α = ((1/n), (1/n), . . . , (1/n)), and M (u) = M1 (u) as u ∈ E.
0 Definition 4.1: Let u, v ∈ L(E n ) with Mα (u), Mα (v) ≥ 0
and Mα (u) + Mα (v) = 0. We denote
≤ D(u, v).
n 1
Thus, we also obtain D̂α ,p (u, v) ≤ Dp (u, v) ≤ D(u, v) and the i=1 αi 0 2r|ui (r) − vi (r)|dr
Lα ,a (u, v) =
proof of 1) of the theorem is complete. [Mα (u) + Mα (v)]a
From

1 n and
D̂α ,p (u, v) = r αi max{|ui (r) − vi (r)|, n 1
0 i=1 αi 0 2r|ui (r) − vi (r)|dr
i=1 Rα ,a (u, v) =
p 1/p [Mα (u) + Mα (v)]a
|ui (r) − vi (r)|} dr and call Lα ,a (u, v) and Rα ,a (u, v) a left difference value and a
right difference value of u and v (with respect to the weight α
1/p
1 and parameter a), respectively. And we denote
p
≤ r
[rdL ([u] , [v] )] dr r
0 1
∆α ,a (u, v) = [Lα ,a (u, v) + Rα ,a (u, v)]
p 1/p 2
1
≤ r sup dL ([u] , [v] ) dr r r i.e.
0 r ∈[0,1] n 1
i=1 αi 0 r|(ui (r)−vi (r)|+|ui (r)−vi (r)|]dr
1/p ∆α ,a (u, v)= 1 a
1 n
= [rDL (u, v)] dr p i=1 αi 0 r[ui (r)+vi (r)+ui (r)+vi (r)]dr
0
1/p and call ∆α ,a (u, v) a difference value of u and v (with respect to
1 α and parameter a), where α = (α1 , α2 , . . . , αn ) ∈
the weight
≤ D(u, v) rp dr n
Rn with i=1 αi = 1 and αi ≥ 0 (i = 1, 2, . . . , n), and
0
a ∈ (0, +∞). Specially, we denote ∆a (u, v) = ∆α ,a (u, v) as
1 α = ((1/n), (1/n), . . . , (1/n)), and ∆a (u, v) = ∆1,a (u, v) as
= D(u, v)
(p + 1)1/p u, v ∈ E.
Remark 4.1: 1, 2). From the definitions of D̃α ,p and ∆α ,a , we can ob-
1) Generally speaking, we consider that the degree of the tain D̃α ,p (u, v), D̃α ,p (u , v ), ∆α ,a (u, v), and ∆α ,a (u , v ),
difference of two numbers is related not only to the metric as shown at the bottom of the next page, so we have
of them but also to the sizes of them. As the metrics are the D̃α ,p (u, v) = D̃α ,p (u , v ) and ∆α ,a (u, v) > ∆α ,a (u , v ).
same, the bigger the sizes of the two numbers, the smaller Property 4.1: Let u, v ∈ L(E n ) with Mα (u), Mα (v) ≥
the degree of their difference. The denominator [Mα (u) + 0, and Mα (u) + Mα (v) = 0, α = (α1 , α2 , . . . , αn ) ∈ Rn with
Mα (v)]a in the definition of ∆α ,a just plays the action (see n
i=1 αi = 1 and αi > 0 (i = 1, 2, . . . , n), and a ∈ (0, +∞).
Example 4.1), and the exponent a in [Mα (u) + Mα (v)]a Then
can be properly chosen accordingly
n to the case in question. 1) ∆α ,a (u, v) ≥ 0;
2) Taking the note that i= 1
αi [(M (ui ) + M (vi )) 2) ∆α ,a (u, v) = 0 if and only if u = v;
/ (M (ui ) + M (vi ))]a = 1 holds as α = ((1/n), (1/n) 3) ∆α ,a (u, v) = ∆α ,a (v, u);
n
, . . . , (1/n)) and a = 1, and i= 1
αi [(M (ui ) + M 4) ∆α ,a (u, w) ≥ ∆α ,a (v, w) for any w ∈ L(E n ) and u ≤
a
(vi ))/(Mα (u) + Mα (v))] = 1 does not necessarily v ≤ w;
hold as α = ((1/n), (1/n), . . . , (1/n)) or a = 1, from 5) ∆α ,a (u + w, v + w) ≤ ∆α ,a (u, v) for any w ∈ L(E n )
and w ≥ (0̂, 0̂, . . . , 0̂);
∆α ,a (u, v)
6) ∆α ,a (ku, kv) = k1−a ∆α ,a (u, v) for any k > 0.
n 1
i=1 αi 0 r[|ui (r) − vi (r)| + |ui (r) − vi (r)|]dr
Proof: It is obvious that the conclusions 1) and 3) hold. The
= 1 a proof of conclusion 2) can also be completed by imitating the
n
i=1 αi 0 r[ui (r) + vi (r) + ui (r) + vi (r)]dr proof of 3) of Theorem 3.2 by using [18, Lemma 2.1].
1 a From u ≤ v ≤ w, we know that ui (r) ≤ vi (r) ≤ wi (r) and
n
0 r[ui (r)+vi (r)+ui (r)+vi (r)]dr ui (r) ≤ vi (r) ≤ wi (r) (i = 1, 2, . . . , n), so |ui (r) − wi (r)| ≥
= αi n 1
i=1 i=1 αi 0 r[ui (r)+vi (r)+ui (r)+vi (r)]dr |vi (r) − wi (r)| and |ui (r) − wi (r)| ≥ |vi (r) − wi (r)|. There-
1 fore, we have
r[|ui (r) − vi (r)| + |ui (r) − vi (r)|]dr
× 0 a

n 1
1
0 r[ui (r) + vi (r) + ui (r) + vi (r)]dr αi r[|ui (r) − wi (r)| + |ui (r) − wi (r)|]dr

n
a i=1 0
M (ui ) + M (vi )
= αi ∆a (ui , vi )
n 1
i=1
Mα (u) + Mα (v) ≥ αi r[|vi (r) − wi (r)| + |vi (r) − wi (r)|]dr.
i=1 0
we can directly see that ∆1 (u, v) is a convex combination
of ∆1 (ui , vi ), i = 1, 2, . . . , n, but ∆α ,a (u, v) is not neces- On the other hand, from ui (r) ≤ vi (r) ≤ wi (r) and
sarily a convex combination of ∆a (ui , vi ), i = 1, 2, . . . , n ui (r) ≤ vi (r) ≤ wi (r) (i = 1, 2, . . . , n), we can also
1
as α = ((1/n), (1/n), . . . , (1/n)) or a = 1. see that Mα (u) = ni=1 αi 0 r[ui (r) + ui (r)]dr ≤
Example 4.1: Let u, v, u , and v be the two-cell numbers n 1
i=1 αi 0 r[vi (r) + vi (r)]dr = Mα (v), so we can ob-
defined by u = (u1 , u2 ), v = (v1 , v2 ), u = (u1 , u2 ), and v = tain 0 ≤ Mα (u) ≤ Mα (v) ≤ Mα (w). Thus, we have
(v1 , v2 ), where
 ∆α ,a (u, w)

 xi , if xi ∈ [0, 1]
n 1
ui (xi ) = 2 − xi , if xi ∈ (1, 2] i=1 αi 0 r[|ui (r) − wi (r)| + |ui (r) − wi (r)|]dr

 =
0, if xi ∈ / [0, 2] (Mα (u) + Mα (w))a
 n 1
 −10 + xi , if xi ∈ [10, 11]
 i=1 αi 0 r[|vi (r) − wi (r)| + |vi (r) − wi (r)|]dr
≥
vi (xi ) = 12 − xi , if xi ∈ (11, 12] (Mα (v) + Mα (w))a


0, if xi ∈/ [10, 12] = ∆α ,a (v, w)

 −100 + xi , if xi ∈ [100, 101]
 so conclusion 4) holds.

ui (xi ) = 102 − xi , if xi ∈ (101, 102] From w ≥ 0̂, we know that Mα (w) ≥ 0, so we have


0, if xi ∈/ [100, 102] ∆α ,a (u + w, v + w), as shown at the bottom of the next page,
 i.e., conclusion 5) holds.
 −110 + xi , if xi ∈ [10, 11] For any k > 0, we have

vi (xi ) = 112 − xi , if xi ∈ (11, 12]

0, if xi ∈/ [10, 12]
∆α ,a (ku, kv)
and i = 1, 2. Then, we know that ui (r) = r, ui (r) = 2 − r, n 1
vi (r) = 10 + r, vi (r) = 12 − r, ui (r) = 100 + r, ui (r) = i=1 αi 0 r[|(ku)i (r)−(kv)i (r)|+|(ku)i (r)−(kv)i (r)|]dr
= n 1 a
102 − r, vi (r) = 110 + r, and vi (r) = 112 − r (r ∈ [0, 1], i = i=1 αi 0 r[(ku)i (r)+(kv)i (r)+(ku)i (r)+(kv)i (r)]dr
n 1
Remark 4.2:
k i=1 αi 0 r[|ui (r) − vi (r)| + |ui (r) − vi (r)|]dr
= 1 1) Although the conclusion 4) of Theorem 4.1 holds, u ≤
k a ( ni=1 αi 0 r[ui (r) + vi (r) + ui (r) + vi (r)]dr)a
v ≤ w does not imply ∆α ,a (u, w) ≥ ∆α ,a (u, v) (see
=k 1−a ∆α ,a (u, v) Example 4.2). Comparing
∆α ,a (u, w)
n 1
so conclusion 6) holds. Therefore, the proof of the theorem is i=1 αi 0 r[|ui (r) − wi (r)| + |ui (r) − wi (r)|]dr
=
complete. (Mα (u) + Mα (w))a
D̃α ,p (u, v)
1/p
1
1
= [r(α1 (|r − 10 − r| + |2 − r − 12 + r|)+α2 (|r − 10 − r| + |2 − r − 12 + r|))] dr
p
2 0

1/p
1 20p
=
2 (1 + p)
D̃α ,p (u , v )
1/p
1
1
= [r(α1 (|100 + r − 110 − r| + |102 − r − 112 + r|)+α2 (|100+r − 110 − r|+|102−r−112 + r|))] dr
p
2 0

1/p
1 20p
=
2 (1 + p)
∆α ,a (u, v)
1 1
α1 0 r(|r − 10 − r| + |2 − r − 12 + r|)dr + α2 0 r(|r − 10 − r| + |2 − r − 12 + r|)dr
= 1 a
1
α1 0 r(r + 10 + r + 2 − r + 12 − r)dr + α2 0 r(r + 10 + r + 2 − r + 12 − r)dr
10
=
12a
∆α ,a (u , v )
1 1
α1 0 r(|100 + r − 110 − r| + |102 − r − 112 + r|)dr + α2 0 r(|100 + r − 110 − r| + |102 − r − 112 + r|)dr
= 1 a
1
α1 0 r(100 + r + 110 + r + 102 − r + 112 − r)dr + α2 0 r(100 + r + 110 + r + 102 − r + 112 − r)dr
10
=
212a
n 1
i=1 αi0 r[|(u + w)i (r) − (v + w)i (r)| + |(u + w)i (r) − (v + w)i (r)|]dr
∆α ,a (u + w, v + w) = 1 a
n
i=1 αi 0 r[(u + w)i (r) + (v + w)i (r) + (u + w)i (r) + (v + w)i (r)]dr
n 1
i=1 αi 0 r[|ui (r) + wi (r) − vi (r) − wi (r)| + |ui (r) + wi (r) − vi (r) − wi (r)|]dr
= 1 a
n
i=1 αi 0 r[ui (r) + wi (r) + vi (r) + wi (r) + ui (r) + wi (r) + vi (r) + wi (r)]dr
n 1
i=1r[|ui (r) − vi (r)| + |ui (r) − vi (r)|]dr
αi 0
=
(Mα (u) + Mα (v) + 2Mα (w))a
n 1
αi 0 r[|ui (r) − vi (r)| + |ui (r) − vi (r)|]dr
≤ i=1
(Mα (u) + Mα (v))a
= ∆α ,a (u, v)

to  19 19 9

 + x , if x ∈ − , −

 10
i i
10 10



 1 9 1
∆α ,a (u, v) ui (xi ) = − xi , if xi ∈ − ,
1 
 10 10 10
n 

i=1 αi 0 r[|ui (r) − vi (r)| + |ui (r) − vi (r)|]dr


= 
 0, if xi ∈ / − ,
19 1
(Mα (u) + Mα (v))a 10 10
and i = 1, 2. Then, we know that ui (r) = −(19/10) +
we can see that the untruth of ∆α ,a (u, w) ≥ r, ui (r) = (1/10) − r, vi (r) = 1, vi (r) = 1, wi (r) = 2, and
∆α ,a (u, v) is caused only by (Mα (u) + Mα (w))a ≥ wi (r) = 2 (r ∈ [0, 1], i = 1, 2), so we have ∆α ,1 (u, v) =
(Mα (u) + Mα (v))a , and when (Mα (u) + Mα (w))a and 19, ∆α ,1 (u, w) = 29/11, and ∆α ,1 (v, w) = 1/3, which
(Mα (u) + Mα (v))a are properly smaller, u ≤ v ≤ w can implies that ∆α ,1 (u, v) > ∆α ,1 (u, w) + ∆α ,1 (v, w).
imply ∆α ,a (u, w) ≥ ∆α ,a (u, v). So, in general, we may Example 4.4: Let u, v, and w be the two-cell numbers (see
choose a in (0, 1] such that ∆α ,a can reasonably char- Fig. 1) defined by: u = (u1 , u2 ), v = (2̂, 2̂), and w = (w1 , w2 ),
acterize the degree of the difference of two fuzzy n-cell where

numbers.  xi , if xi ∈ [0, 1]
2) Generally speaking, the difference value ∆α ,a does not ui (xi ) = 2 − xi , if xi ∈ (1, 2]

satisfy the property of the triangular inequality, i.e., the 0, if xi ∈/ [0, 2]
inequality ∆α ,a (u, v) ≤ ∆α ,a (u, w) + ∆α ,a (w, v) does
not necessarily hold for u, v, w ∈ L(E n ). Example 4.3 
can show it.  −1 + xi , if xi ∈ [1, 2]
Example 4.2: Let u, v, and w be the two-cell numbers defined wi (xi ) = 3 − xi , if xi ∈ (2, 3]

by: u = (0̂, 0̂), v = (v1 , v2 ), and w = (w1 , w2 ), where 0, if xi ∈
/ [1, 3]
and i = 1, 2. Then, we know that ui (r) = r, ui (r) = 2 −

 r, vi (r) = 2, vi (r) = 2, wi (r) = 1 + r, and wi (r) = 3 − r (r ∈
 −1 + xi ,
 if xi ∈ [1, 2]
[0, 1], i = 1, 2), so we can obtain that D̃α ,p (u, v) = 1/2 =
vi (xi ) = 3 − xi , if xi ∈ (2, 3] D̃α ,p (u, w) and ∆α ,a (u, v) = 2/3a = ∆α ,a (u, w). However,


0, if xi ∈
/ [1, 3] it is obvious (see Fig. 1) that the degree of the difference of
 u and v is different from the degree of the difference of u
 −2 + xi ,
 if xi ∈ [2, 3]
and w.
wi (xi ) = 4 − xi , if xi ∈ (3, 4] In fact, sometimes, the degree of the difference of two fuzzy


0, if xi ∈
/ [2, 4] numbers is not only related with the metric and the sizes of
them, but also related with the degree of fuzzy (we call it fuzzy
degree) of them. Example 4.4 shows that for the u, v, and w
and i = 1, 2. Then, we know that ui (r) = 0, ui (r) = 0, vi (r) = given, the metric D̃α ,p and the difference value ∆α ,a cannot
1 + r, vi (r) = 3 − r, wi (r) = 2 + r, and wi (r) = 4 − r (r ∈ tell us the difference of the degree of the difference u and v
[0, 1], i = 1, 2), so we have u ≤ v ≤ w, but we can see from with the degree of the difference of u and w since D̃α ,p (u, v)
∆α ,a (u, w) ≤ ∆α ,a (u, v) from ∆α ,2 (u, w), and ∆α ,2 (u, v), as = D̃α ,p (u, w) and ∆α ,a (u, v) = ∆α ,a (u, w), but we see that the
shown at the bottom of this page. two degrees of the differences indeed have some differences.
Example 4.3: Let u, v, and w be the two-cell numbers defined In order to overcome the defect, we introduce the following
by: u = (u1 , u2 ), v = (1̂, 1̂), and w = (2̂, 2̂), where concept.
1 1
α1 r[|0 − 2 − r| + |0 − 4 + r|]dr + α2 0 r[|0 − 2 − r| + |0 − 4 + r|]dr
∆α ,2 (u, w) = 0 1 2
1
α1 0 r[0 + 2 + r + 0 + 4 − r]dr + α2 0 r[0 + 2 + r + 0 + 4 − r]dr
1
=
3
1 1
α1 0 r(|0 − 1 − r| + |0 − 3 + r|)dr + α2 0 r(|0 − 1 − r| + |0 − 3 + r|)dr
∆α ,2 (u, v) = 1 2
1
α1 0 r(0 + 1 + r + 0 + 3 − r)dr + α2 0 r(0 + 1 + r + 0 + 3 − r)dr
1
= .
2
2
2

+ (2 − t)dt − 1 dt
1 2
1
= 2r dr exp(1)
0
=e
and
2
1
Λα ,1 (u, w) = αi r(|r−1−r|+|2−r−3+r|)dr
i=1 0
2

2 1

× exp αi t dt − (t − 1)dt
0 1
i=1
3
2

+ (2 − t)dt − (3 − t)dt
1 2
1
= 2r dr exp(0)
0
=1
so we see that Λα ,1 (u, v) > Λα ,1 (u, w). Therefore, in this case,

the difference value Λα ,a is more suitable than metrics and
Fig. 1. Fuzzy two-cell numbers u, v, and w in Example 4.4. difference value ∆α ,a to characterize the degree of difference
of two fuzzy n-cell numbers.
n 4.2: Let u, v ∈ L(E ), α = (α1 , α2 , . . . , αn ) ∈ R
n n
Property
Definition 4.2: Let u, v ∈ L(E n ). We denote with i=1 αi = 1 and αi > 0 (i = 1, 2, . . . , n), and a ∈
(0, +∞). Then
Λα ,a (u, v) 1) Λα ,a (u, v) ≥ 0;
n 2) Λα ,a (u, v) = 0 if and only if u = v;
1
3) Λα ,a (u, v) = Λα ,a (v, u);
= αi r[|(ui (r) − vi (r)| + |ui (r) − vi (r)|]dr
i=1 0 4) Λα ,a (u + b̂, v + b̂) = Λα ,a (u, v) for any b ∈ R;
∗ v ∗ (1) 5) Λα ,a (ku, kv) = |k|Λα ,|k |a (u, v) for any k ∈ R.

n u i (1)
i
Proof: It is obvious that the conclusions 1) and 3) hold. The
× exp a αi ui (t)dt − vi (t)dt
u i (0) v i (0) proof of conclusion 2) can also be completed by imitating the
i=1
v i (0) proof of 3) of Theorem 3.2 by using [18, Lemma 2.1].
u i (0) For any b ∈ R and i = 1, 2, . . . , n, we have

+ ui (t)dt − vi (t)dt
u ∗i (1) v i∗ (1)
(ui + b̂)(t) = sup{r ∈ [0, 1] : t ∈ [(ui + b̂)(r), (ui + b̂)(r)]}
and call Λα ,a (u, v) a difference value of u and v = sup{r ∈ [0, 1] : t − b ∈ [ui (r), ui (r)]}
(with respect
to the weight
α and parameter a), where
u∗i (1) = ui (1) + ui (1) /2, vi∗ (1) = vi (1) + vi (1) /2, and = ui (t − b).
n
α = (α1 , α2 , . . . , αn ) ∈ Rn with i=1 αi = 1 and αi ≥ 0
(i = 1, 2, . . . , n) and a ∈ (0, +∞). We denote Λa (u, v) = Hence
Λα ,a (u, v) as α = ((1/n), (1/n), . . . , (1/n)) and Λa (u, v) =
Λα ,a (u + b̂, v + b̂)
Λ1,a (u, v) as u, v ∈ E. n
Example 4.5: Let u, v, and w be the two-cell numbers defined 1
in Example 4.4. Then = αi r|(ui (r) + b − vi (r) − b|
i=1 0
2
1
Λα ,1 (u, v) = αi r(|r − 2| + |2 − r − 2|)dr + |ui (r) + b − vi (r) − b|]dr
i=1 0
2 ∗

2 1
n u i (1)+b

× exp αi t dt − 1 dt × exp a αi (ui + b̂)(t)dt
0 2 u i (0)+b
i=1 i=1
k v i (0)

v i∗ (1)+b k u i (0)
t
t
− (vi + b̂)(t)dt + ui dt − vi dt
v i (0)+b k u ∗i (1) k ∗
k v i (1) k
v i (0)+b n
u i (0)+b 1

+ (ui +b̂)(t)dt− (vi +b̂)(t)dt = k αi r[|ui (r) − vi (r)| + |ui (r) − vi (r)|]dr
u ∗i (1)+b v i∗ (1)+b 0
i=1
n ∗ v ∗ (1)
1
n u i (1)
i

= αi r|(ui (r) − vi (r)| + |ui (r) − vi (r)|]dr × exp ka αi ui (s)ds− vi (s)ds
0 u i (0) v i (0)
i=1 i=1
∗ v i (0)

n u i (1) u i (0)

× exp a αi (u + b̂)(s + b)ds + ui (s)ds − vi (s)ds
u i (0) i u ∗i (1) ∗
v i (1)
i=1
v ∗ (1)
u i (0) = |k|[Λα ,|k |a (u, v)].
i

− (vi +b̂)(s+b)ds+ (ui +b̂)(s+b)ds
v i (0) u ∗i (1)
For any k < 0, we have
v i (0)

− (vi +b̂)(s+b)ds
v i∗ (1) Λα ,a (ku, kv)
n
n
1
1
= αi r|(ui (r) − vi (r)| + |ui (r) − vi (r)|]dr = αi r[|(ku)i (r) − (kv)i (r)|
i=1 0
i=1 0
∗

n u i (1) v ∗ (1)
i
+ |(ku)i (r) − (kv)i (r)|]dr
× exp a αi (ui )(s)ds− (vi )(s)ds
u i (0) v i (0)
i=1

u i (0) v i (0) (k u ) ∗i (1)
n

+ (ui )(s)ds − (vi )(s)ds × exp a αi (ku)i (t)dt
u ∗i (1) ∗ (k u ) i (0)
v i (1) i=1
(k v ) ∗ (1)
(k u ) i (0)
= Λα ,a (u, v) i

− (kv)i (t)dt + (ku)i (t)dt
(k v ) i (0) (k u ) ∗i (1)
i.e., conclusion 4) holds.
For any k > 0, we have (k v ) i (0)

− (kv)i (t)dt
Λα ,a (ku, kv) ∗
(k v ) i (1)
n n
1 1
= αi r[|(ku)i (r)−(kv)i (r)|+|(ku)i (r)−(kv)i (r)|]dr = αi r[|kui (r)−kvi (r)|+|kui (r)−kvi (r)|]dr
i=1 0 0
i=1
∗
n (k u ) ∗i (1)

n k u i (1)
t
αi
exp a
(k u ) i (0)
(ku)i (t)dt × exp a αi u dt
i=1
k u i (0) i k
i=1
(k v ) ∗ (1) k v ∗ (1)
k u i (0)

(k u ) i (0)
i
i t t
− (kv)i (t)dt + (ku)i (t)dt − vi dt + ui dt
(k v ) i (0) (k u ) ∗i (1) k v i (0) k ∗
k u i (1) k
(k v ) i (0) k v i (0)

t
− (kv)i (t)dt − vi dt
∗
(k v ) i (1) ∗
k v i (1) k
n
1
n 1
= αi r[|kui (r)−kvi (r)| + |kui (r) − kvi (r)|]dr = |k| αi r[|ui (r) − vi (r)| + |ui (r) − vi (r)|]dr
i=1 0 i=1 0
∗ ∗
n k u i (1)
t
n u i (1) v ∗ (1)
i

× exp a αi u dt × exp a αi kui (s)ds− kvi (s)ds
k u i (0) i k u i (0) v i (0)
i=1 i=1
k v ∗ (1)

u i (0) v i (0)

i t
− vi dt + kui (s)ds − kvi (s)ds
k v i (0) k u ∗i (1) ∗
v i (1)

n 1 V. PATTERN RECOGNITION BASED ON METRICS AND
= |k| αi r[|ui (r) − vi (r)| + |ui (r) − vi (r)|]dr DIFFERENCE VALUES
i=1 0
∗ v ∗ (1) In Sections III and IV, we discussed metrics and difference

n u i (1) values on L(E n ). In this section, we put forward an algo-
i

× exp |k|a αi ui (s)ds− vi (s)ds
u i (0) v i (0) rithmic version of pattern recognition in an imprecise or un-
i=1
v i (0) certain environment based on the metrics and difference val-
u i (0)
ues defined by us and give examples to show the application
+ ui (s)ds − vi (s)ds
u ∗i (1) ∗
v i (1) (see Example 5.1) and rationality (see Example 5.2) of the
method.
= |k|[Λα ,|k |a (u, v)]. Consider a problem to identify an object (denoted by O) be-
longing to some one of l classes (denoted by C1 , C2 , . . . , Cl ) in
an imprecise or uncertain environment. Let the objects have n
If k = 0, it is obvious that Λα ,a (ku, kv) = |k|Λα ,|k |a (u, v) characteristics. Since the problem discussed by us takes on some
holds, so conclusion 5) holds. Therefore, the proof of the theo- imprecise or uncertain attributes, it is unsuitable (see Remark
rem is complete. 2.1) that we use a crisp n-dimensional vector (i.e., a standard
At the end of the section, combining the definitions of ∆α ,a n-dimensional real number vector) to express the n character
and Λα ,a , we give the following definition of difference value values of Ci (i = 1, 2, . . . , l) or O. Therefore, using the method
Γα ,a . of statistics, we construct l + 1 fuzzy n-cell numbers to express
Definition 4.3: Let u, v ∈ L(E n ) with Mα (u), Mα (v) ≥ 0 the n character values of C1 , C2 , . . . , Cl and O, respectively,
and Mα (u) + Mα (v) = 0. We denote Γα ,a (u, v), as shown at and then put forward an algorithmic version of pattern recog-
the bottom of this page, and call Γα ,a (u, v) a difference value nition based on the metrics or the difference values defined
of u and v [with respect to αand a = (a1 , a2 )], where α = by us.
n
(α1 , α2 , . . . , αn ) ∈ Rn with i=1 αi = 1 and αi ≥ 0 (i =
1, 2, . . . , n), and a = (a1 , a2 ) ∈ (0, +∞) × (0, +∞). We de-
note Γa (u, v) = Γα ,a (u, v) as α = ((1/n), (1/n), . . . , (1/n)) A. Algorithmic Version of Pattern Recognition Based on
and Γa (u, v) = Γ1,a (u, v) as u, v ∈ E. Metrics
Likewise, we have the following properties about the differ-
1) First Step: Depending on the practicality, we first find
ence value Γα ,a .
out one domain of the jth character value of Ci for each i (i =
Property 4.3: Let u, v ∈ L(E n ) with Mα (u), Mα (v) ≥
1, 2, . . . , l) and j (j = 1, 2, . . . , n) and denote the said domain
0 and Mα (u) + Mα (v) = 0, α = (α1 , α2 , . . . , αn ) ∈ Rn with
n by Dji .
i=1 αi = 1 and αi > 0 (i = 1, 2, . . . , n), and a = (a1 , a2 ) We arbitrarily take mi samples in Ci (i = 1, 2, . . . , l) and gain
∈ (0, +∞) × (0, +∞). Then
mi measure values (denoted by ci1 , ci2 . . . , cim i ) of the n charac-
1) Γα ,a (u, v) ≥ 0;
ters of the mi samples and denote cim = (cim 1 , cim 2 , . . . , cim n )
2) Γα ,a (u, v) = 0 if and only if u = v;
(m = 1, 2, . . . , mi ), i.e., we gain the following tables as shown
3) Γα ,a (u, v) = Γα ,a (v, u);
by C1 , C2 , and Cl at the bottom of this page.
4) Γα ,a (u + b̂, v + b̂) ≤ Γα ,a (u, v) for any b ∈ [0, +∞); For Ci (i = 1, 2, . . . , l),we directly work out the follow-
5) Γα ,a (ku, kv) = k 2−a 1 Γα ,b (u, v) for any k ∈ [0, +∞), ing means µij = (1/mi ) m i i
k =1 ck j and standard deviations
where b = (a1 , ka2 ). m
Proof: The proofs of the properties can be completed similarly σji = (1/(mi − 1)) k =1 i
(cik j − µij )2 (j = 1, 2, . . . , n) of
with the proofs of Properties 4.1 and 4.2, respectively, so we the n character values of Ci (i = 1, 2, . . . , l), respectively.
omit it. We construct triangular model 1-D fuzzy numbers uij
u ∗ (1) v ∗ (1) u (0) v (0)

exp a2 ni=1 αi u ii(0) ui (t)dt − v ii(0) vi (t)dt| + | u ∗i(1) ui (t)dt − v ∗i(1) vi (t)dt
Γα ,a (u, v) = 1
i
a 1 i
n
i=1 αi 0 r[ui (r) + vi (r) + ui (r) + vi (r)]dr
n
1
× αi r[|(ui (r) − vi (r)| + |ui (r) − vi (r)|]dr
i=1 0
c111 c112 ··· c11n c211 c212 ··· c21n cl11 cl12 ··· cl1n
c121 c122 ··· c12n c221 c222 ··· c22n cl21 cl22 ··· cl2n
C1 : . .. .. C2 : . .. .. ··· Cl : . .. ..
.. . . .. . . .. . .
c1m 1 1 c1m 1 2 ··· c1m 1 n c2m 2 1 c2m 2 2 ··· c2m 2 n clm l 1 clm l 2 ··· clm l n .
(i = 1, 2, . . . , l, j = 1, 2, . . . , n) as follows: We construct fuzzy n-cell numbers w = (w1 , w2 , . . . , wn )

 (w(x1 , x2 , . . . , xn ) = min {w1 (x1 ), w2 (x2 ), . . . , wn (xn )})
 x − (µij − 2σji )

 , if x ∈ [µij − 2σji , µij ] ∩ Dji and w = (w1 , w2 , . . . , wn ) (w (x1 , x2 , . . . , xn ) =

 2σ i

 j min {w1 (x1 ), w2 (x2 ), . . . , wn (xn )}), and use w or w to
uij (x) = (µij + 2σji ) − x express the object O.

 , if x ∈ (µij , µij + 2σji ] ∩ Dji

 2σ i
3) Third Step: Taking proper α = (α1 , α2 , . . . , αn ) with

 j n
 i=1 αi = 1 and αi ≥ 0, i = 1, 2, . . . , n, and p ≥ 1, we com-
0, if x∈[µ
/ ij −2σji , µij +2σji ]∩Dji pute the metrics
i = 1, 2, . . . , l, j = 1, 2, . . . , n
D̃α ,p (w, uj )
or construct Gaussian model 1-D fuzzy numbers vji (i = n
1, 2, . . . , l, j = 1, 2, . . . , n) as follows: 1 1
 = r αi (|wi (r) − uji (r)|
 (x − µij )2 2 0
 i=1
exp − , if x ∈ Dji
vji (x) = 2σji
2
p 1/p


0, if x ∈
/ Dji + |wi (r) − uji (r)|) dr , j = 1, 2, . . . , l
i = 1, 2, . . . , l j = 1, 2, . . . , n.
or
We construct fuzzy n-cellnumbers u = (ui1 , ui2 , . . . , uin
i
)
u (x1 , x2 , . . . , xn ) = min ui1 (x1 ), ui2 (x2 ), . . . , uin (xn )
i
D̃α ,p (w , v j )
and v i = (v1i , v2i , . . . , vni ) (v i (x1 , x2 , . . . , xn ) = min{v1i (x1 ), n
v2i (x2 ), . . . , vni (xn )}), i = 1, 2, . . . , l, and use ui or v i to 1 1
express the ith class Ci (i = 1, 2, . . . , l). = r αi (|wi (r) − vij (r)|
2 0 i=1
2) Second Step: For the object O to be recognized, taking t
samples in O, we can gain t classes of data about the n characters p 1/p
of O as follows: + |wi (r) − vij (r)|) dr , j = 1, 2, . . . , l.
o11 o12 · · · o1n
o21 o22 · · · o2n
O: .. .. .. 4) Fourth Step: We choose uj 0 in u1 , u2 , . . . , ul , or v j 0 in

. . .
ot1 ot2 ··· otn . v 1 , v 2 , . . . , v l such that
We work out the following means (denoted by ō1 , ō2 , . . . , ōn ) D̃α ,p (w, uj 0 )
and standard deviations (denoted by s1 , s2 , . . . , sn ) of the n
character values of O: = min{D̃α ,p (w, u1 ), D̃α ,p (w, u2 ), . . . , D̃α ,p (w, ul )}
1
t
ōi = ok i , i = 1, 2, . . . , n or
t
k =1

and D̃α ,p (w , v j 0 )

= min{D̃α ,p (w , v 1 ), D̃α ,p (w , v 2 ), . . . , D̃α ,p (w , v l )}.
1 t
si = (ok i − ōi )2 , i = 1, 2, . . . , n.
t−1
k =1 Then, we can consider that object O belongs to the j0 th class
We construct triangular model 1-D fuzzy numbers wi (i = Cj 0 , or belongs to the j0 th class Cj 0 .
1, 2, . . . , n) as follows: Remark 5.1: In the third and fourth steps of the aforemen-
 tioned method, we can use the metric D̂α ,p to replace the metric
 x−(ōi −2si )

 , if x ∈ [ōi −2si , ōi ]∩(∪lj =1 Dij ) D̃α ,p , as a result of which, we can also set up a method based

 2s
 i
on the metric D̂α ,p .
wi (x)= (ōi +2si )−x

 , if x ∈ (ōi , ōi +2si ]∩(∪lj =1 Dij )

 2si

 B. Algorithmic Version of Pattern Recognition Based on
0, / [ōi −2si , ōi +2si ]∩(∪lj =1 Dij )
if x ∈ Difference Values
i = 1, 2, . . . , n 1) First Step and the Second Step: They are same, respec-
or construct Gaussian model 1-D fuzzy numbers wi (i = tively, with the first step and the second step of the method of
1, 2, . . . , n) as follows: pattern recognition based on the metric, as mentioned earlier.

Third Step: Taking proper α = (α1 , α2 , . . . , αn ) with
 v (x) = exp − (x − ōi ) , if x ∈ ∪l Dj
2 2)n
i j =1 i i=1 αi = 1 and αi ≥ 0, i = 1, 2, . . . , n, and a = (a1 , a2 ) ∈
2s2i
 (0, +∞) × (0, +∞), we compute the difference values
0, / ∪lj =1 Dij
if x ∈ Γα ,a (w, uj ) or Γα ,a (w , v j ), as shown at the bottom of the next
i = 1, 2, . . . , n. page.

3) Fourth Step: We choose uj 0 in u1 , u2 , . . . , ul or v j 0 in C3 : Korean pine accounts for the main part; C4 : boreal and
v , v 2 , . . . , v l such that
1 broad-leaf mixture forest; and C5 : birch forest. For the five
Γα ,a (w, uj 0 ) types of land cover (C1 , C2 , C3 , C4 , C5 ) and by using the four
wave bands: MSS-4, MSS-5, MSS-6, and MSS-7, we take ten
= min{Γα ,a (w, u1 ), Γα ,a (w, u2 ), . . . , Γα ,a (w, ul )} samples and acquire the following data:
or
MSS-4 MSS-5 MSS-6 MSS-7
Γα ,a (w , v j 0 ) sample 1 18.62 20.71 58.20 26.72
= min{Γα ,a (w , v 1 ), Γα ,a (w , v 2 ), . . . , Γα ,a (w , v l )}. sample 2 19.76 17.01 51.02 24.32
Then, we can consider that object O belongs to the j0 th class sample 3 18.24 19.46 48.12 26.33
Cj 0 , or belongs to the j0 th class Cj 0 . sample 4 18.76 15.95 56.35 22.89
Remark 5.2: In the third and fourth steps of the aforemen- C1 : sample 5 18.96 18.78 45.32 28.32
tioned method, we can have the difference value ∆α ,a or Λα ,a sample 6 19.90 20.13 50.82 25.05
to replace the difference value Γα ,a , as a result of which, we sample 7 19.16 18.58 52.30 22.21
can also set up a method based on the difference value ∆α ,a or sample 8 19.36 17.32 55.02 25.28
Λα ,a . sample 9 19.36 17.98 48.26 23.86
In order to be more obvious, we may use the following dia- sample 10 18.48 16.46 46.63 27.46
gram to illustrate the methods set up by us. MSS-4 MSS-5 MSS-6 MSS-7
sample 1 24.51 28.13 52.39 22.43
sample 2 19.37 23.41 45.56 23.89
sample 3 23.12 24.82 47.80 18.05
sample 4 18.93 18.87 43.88 20.43
C2 : sample 5 21.36 32.28 53.61 24.95
sample 6 26.06 24.15 42.75 19.38
sample 7 17.64 17.08 46.73 21.16
sample 8 22.23 26.02 49.17 19.04
sample 9 20.26 21.56 41.37 24.39
sample 10 25.42 30.52 50.46 22.59
MSS-4 MSS-5 MSS-6 MSS-7
sample 1 15.01 13.30 40.50 19.37
sample 2 15.60 12.56 38.81 16.35
sample 3 15.82 12.79 37.70 18.16
sample 4 14.90 11.70 35.50 14.75
C3 : sample 5 16.10 13.80 42.10 20.75
sample 6 13.80 11.94 32.10 15.54
sample 7 15.90 10.98 30.87 14.29
sample 8 16.82 13.67 37.64 18.62
Example 5.1: Suppose that some terrain consists of five dif- sample 9 15.50 12.58 36.10 18.02
ferent types of land-based cover: C1 : road; C2 : farm or crop; sample 10 15.38 12.48 34.08 17.45

n w ∗ (1) u ji ∗ (1) j w i (0) u ji (0) j

exp a2 i=1 αi w i (0) wi (t)dt − u j (0) ui (t)dt + w ∗ (1) wi (t)dt − u j ∗ (1) ui (t)dt
i
i
a 1
i i
Γα ,a (w, uj ) = 1
n j j
i=1 αi 0 r[wi (r) + ui (r) + wi (r) + ui (r)]dr
n
1 j j
× αi r[|(wi (r) − ui (r)| + |wi (r) − ui (r)|]dr j = 1, 2, . . . , l
i=1 0

w ∗ (1) v j ∗ (1) w (0) v j (0)
exp a2 ni=1 αi w i(0) wi (t)dt − v ji (0) vij (t)dt + w ∗i (1) wi (t)dt − v ∗i(1) vij (t)dt
i i i
Γα ,a (w , v j ) = a 1
i
n 1 j j
i=1 αi 0 r[wi (r) + vi (r) + wi (r) + vi (r)]dr
n
1
j j
× αi r[|(wi (r) − vi (r)| + |wi (r) − vi (r)|]dr j = 1, 2, . . . , l.
i=1 0
MSS-4 MSS-5 MSS-6 MSS-7 we have

sample 1 17.05 13.53 43.32 22.76  x − 17.96

 , if x ∈ [17.96, 19.06]
sample 2 16.09 12.03 38.65 19.47 
 1.10


sample 3 15.44 12.87 42.21 21.20 
20.16 − x
sample 4 16.41 12.58 40.87 18.63 u11 (x) = , if x ∈ (19.06, 20.16]

 1.10
C4 : sample 5 16.32 13.01 46.56 23.58 



sample 6 17.21 13.58 45.98 21.33  0, if x ∈
/ [17.96, 20.16]
sample 7 15.98 12.14 41.09 20.23
 x − 15.08
sample 8 16.21 12.68 42.51 21.65 

 , if x ∈ [15.08, 18.24]
sample 9 15.09 12.10 38.03 21.01 
 3.16

sample 10 16.38 13.31 44.87 22.34
u12 (x) = 21.40 − x

 , if x ∈ (18.24, 21.40]
MSS-4 MSS-5 MSS-6 MSS-7  3.16

sample 1 18.12 13.45 45.76 23.64 

0, if x ∈
/ [15.08, 21.40]
sample 2 16.41 12.62 44.71 23.19  x − 42.64
sample 3 17.04 12.88 44.95 22.54 
 , if x ∈ [42.64, 51.20]

 8.56
sample 4 16.58 13.13 43.43 22.98 

C5 : sample 5 18.21 13.90 46.38 23.41 u13 (x) = 59.76 − x
sample 6 17.23 13.73 45.03 23.76 
 , if x ∈ (51.20, 59.76]

 8.56
sample 7 15.89 12.98 44.32 23.15 

sample 8 17.02 13.25 45.38 22.61 0, if x ∈
/ [42.64, 59.76]
 x − 21.28
sample 9 15.97 12.70 43.96 23.11 

 , if x ∈ [21.28, 25.24]
sample 10 17.67 13.34 46.09 23.62. 
 3.96

By u14 (x) = 29.20 − x

 , if x ∈ (25.24, 29.20]

 3.96

10 

1 0, if x ∈
/ [21.28, 29.20]
µij = cik j , i = 1, 2, 3, 4, 5, j = 1, 2, 3, 4
10  x − 16.13
k =1 
 , if x ∈ [16.13, 21.89]

 5.76
and 


u21 (x) = 27.65 − x

 , if x ∈ (21.89, 27.65]
1 i
10

 5.76
σji = (ck j − µij )2 , 

10 − 1 0, if x ∈
/ [16.13, 27.65]
k =1
 x − 15.04
i = 1, 2, 3, 4, 5, j = 1, 2, 3, 4 
 , if x ∈ [15.04, 24.68]

 9.64


we can work out the following means and standard deviations: u22 (x) = 34.32 − x

 , if x ∈ (24.68, 34.32]
µ11 = 19.06, σ11 = 0.55, µ12 = 18.24, σ21 = 1.58 
 9.64


µ13 = 51.20, σ31 = 4.28, µ14 = 25.24, σ41 = 1.98 0, if x ∈
/ [15.04, 34.32]
µ21 = 21.89, σ12 = 2.88, µ22 = 24.68, σ22 = 4.82  x − 39.19
µ23 σ32 µ24 σ42 
 , if x ∈ [39.19, 47.31]
= 47.37, = 4.09, = 21.63, = 2.39 
 8.18
µ31 = 15.46, σ13 = 1.22, µ32 = 12.58, σ23 = 0.88 

µ33 = 36.54, σ33 = 3.55, µ34 = 17.33, σ43 = 2.08 u23 (x) = 55.55 − x

 , if x ∈ (47.31, 55.55]
µ41 = 16.22, σ14 = 0.64, µ42 = 12.78, σ24 = 0.58 
 8.18


µ43 = 42.41, σ34 = 2.87, µ44 = 21.22, σ44 = 1.50 0, if x ∈
/ [39.19, 55.55]
µ51 = 17.01, σ15 = 0.82, µ52 = 13.20, σ25 = 0.42  x − 16.85
µ53 = 45.00, σ35 = 0.94, µ54 = 23.20, σ45 = 0.42. 
 , if x ∈ [16.85, 21.63]

 4.78


Taking Dij = (0, +∞)(i = 1, 2, 3, 4 and j = 1, 2, 3, 4, 5), u24 (x) = 26.41 − x
then according to 
 , if x ∈ (21.63, 26.41]

 4.78
 


 x − (µij − 2σji ) 0, if x ∈
/ [16.85, 26.41]

 , if x ∈ [µij − 2σji , µij ] ∩ Dji  x − 13.02

 2σji 
 , if x ∈ [13.02, 15.46]
uij (x) = (µij + 2σji ) − x 
 2.44
 


 , if x ∈ (µij , µij + 2σji ] ∩ Dji

 2σji u31 (x) = 17.90 − x
  , if x ∈ (15.46, 17.90]
0, if x ∈
/ [µij −2σji , µij +2σji ]∩Dji 
 2.44



i = 1, 2, 3, 4, 5, j = 1, 2, 3, 4 0, if x ∈
/ [13.02, 17.90]
 
 x − 10.82  x − 22.36

 , if x ∈ [10.82, 12.58] 
 , if x ∈ [22.36, 23.20]

 1.76 
 0.84
u2 (x) = 14.34 − x
3
u4 (x) = 24.04 − x
5

 , if x ∈ (12.58, 14.34] 
 , if x ∈ (23.20, 24.04]
 1.76
  0.84

 
0, if x ∈
/ [10.82, 14.34] 0, if x ∈
/ [22.36, 24.04]

 x − 29.44

 , if x ∈ [29.44, 36.54] and for r ∈ [0, 1]
 7.10

u33 (x) = 43.64 − x

 , if x ∈ (36.54, 43.64]

 7.10 u11 (r) = 1.10r + 17.96 u11 (r) = 20.16 − 1.10r

0, if x ∈
/ [29.44, 43.64]
 u12 (r) = 3.16r + 15.08 u12 (r) = 21.40 − 3.16r
 x − 13.17

 , if x ∈ [13.17, 17.33]

 4.16 u13 (r) = 8.56r + 42.64 u13 (r) = 59.76 − 8.56r
u34 (x) = 21.49 − x

 , if x ∈ (17.33, 21.49] u14 (r) = 3.96r + 21.28 u14 (r) = 29.20 − 3.96r

 4.16

0, if x ∈
/ [13.17, 21.49]
 x − 14.94 u21 (r) = 5.76r + 16.13 u21 (r) = 27.65 − 5.76r

 , if x ∈ [14.94, 16.22]

 u22 (r) = 9.64r + 15.04 u22 (r) = 34.32 − 9.64r
 1.28
u41 (x) = 17.50 − x

 , if x ∈ (16.22, 17.50] u23 (r) = 8.18r + 39.19 u23 (r) = 55.55 − 8.18r

 1.28

0, if x ∈
/ [14.94, 17.50] u24 (r) = 4.78r + 16.85 u24 (r) = 26.41 − 4.78r
 x − 11.62

 , if x ∈ [11.62, 12.78] u31 (r) = 2.44r + 13.02 u31 (r) = 17.90 − 2.44r


 1.16
u42 (x) = 13.94 − x u32 (r) = 1.76r + 10.82, u32 (r) = 14.34 − 1.76r

 , if x ∈ (12.78, 13.94]
 1.16


0, if x ∈
/ [11.62, 13.94] u33 (r) = 7.10r + 29.44 u33 (r) = 43.64 − 7.10r
 x − 36.67

 , if x ∈ [36.67, 42.41] u34 (r) = 4.16r + 13.17 u34 (r) = 21.49 − 4.16r


 5.74
u43 (x) = 48.15 − x u41 (r) = 1.28r + 14.94 u41 (r) = 17.50 − 1.28r

 , if x ∈ (42.41, 48.15]
 5.74

 u42 (r) = 1.16r + 11.62 u42 (r) = 13.94 − 1.16r
0, if x ∈
/ [36.67, 48.15]
 x − 18.22
 u43 (r) = 5.74r + 36.67 u43 (r) = 48.15 − 5.74r

 , if x ∈ [18.22, 21.22]

 3.00
u44 (r) = 3.00r + 18.22 u44 (r) = 24.22 − 3.00r
u44 (x) = 24.22 − x

 , if x ∈ (21.22, 24.22]
 3.00
 u51 (r) = 1.64r + 15.37 u51 (r) = 18.65 − 1.64r

0, if x ∈
/ [18.22, 24.22]
 u52 (r) = 0.84r + 12.36 u52 (r) = 14.04 − 0.84r
 x − 15.37

 , if x ∈ [15.37, 17.01]

 1.64 u53 (r) = 1.98r + 43.02 u53 (r) = 46.98 − 1.98r
u51 (x) = 18.65 − x

 , if x ∈ (17.01, 18.65]
 1.64
 u54 (r) = 0.84r + 22.36 u54 (r) = 24.04 − 0.84r.

0, if x∈
/ [15.37, 18.65]

 x − 12.36 Thus, the fuzzy four-cell numbers ui = (ui1 , ui2 , ui3 , ui4 ), i =

 , if x ∈ [12.36, 13.20]

 0.84 1, 2, 3, 4, 5, i.e.
u52 (x) = 14.04 − x

 , if x ∈ (13.20, 14.04]

 0.84 ui (x1 , x2 , x3 , x4 ) = min{ui1 (x1 ), ui2 (x2 ), ui3 (x3 ), ui4 (x4 )}

0, if x ∈
/ [12.36, 14.04]
 (x1 , x2 , x3 , x4 ) ∈ R4 , i = 1, 2, 3, 4, 5
 x − 43.02

 , if x ∈ [43.02, 45.00]

 1.98 can be used to represent Ci , i = 1, 2, 3, 4, 5, respectively.
u53 (x) = 46.98 − x

 , if x ∈ (45.00, 46.98] Using the same four wave bands: MSS-4, MSS-5, MSS-6,

 1.98 and MSS-7, we now proceed to examine some zone (i.e., ob-

0, if x ∈
/ [43.02, 46.98] ject, denoted by O) 12 times, stochastically, at various times
or positions, or using various viewers, and obtain the following
data: and for r ∈ [0, 1]

MSS-4 MSS-5 MSS-6 MSS-7 o1 (r) = 1.54r + 15.55 o1 (r) = 18.63 − 1.54r
sample 1 18.31 13.13 45.41 23.52
sample 2 17.21 12.71 45.67 23.49 o2 (r) = 1.00r + 12.17 o2 (r) = 14.17 − 1.00r
sample 3 17.00 12.57 44.82 22.56 o3 (r) = 2.02r + 43.08 o3 (r) = 47.12 − 2.02r
sample 4 16.32 13.06 43.53 22.77
sample 5 18.01 13.91 45.31 23.31 o4 (r) = 0.66r + 22.53 o4 (r) = 23.79 − 0.66r.
O sample 6 16.11 13.59 46.97 23.61 Taking α = ((1/4), (1/4), (1/4), (1/4)) and a=
sample 7 16.21 12.56 43.96 23.14 ((1/5), (1/5)), by Γα ,a (u, v), as shown at the bot-
sample 8 17.13 13.32 46.01 22.84 tom of this page, we can obtain Γα ,a (o, u1 ) =
sample 9 16.20 12.34 43.81 23.13 3.36, Γα ,a (o, u2 ) = 10.34, Γα ,a (o, u3 ) = 4.27, Γα ,a (o, u4 ) =
sample 10 17.51 13.34 45.89 23.52 1.11, and Γα ,a (o, u5 ) = 0.03. From Γα ,a (o, u5 ) =
sample 11 17.02 13.78 45.21 23.18 min{Γα ,a (o, u1 ), Γα ,a (o, u2 ), Γα ,a (o, u3 ), Γα ,a (o, u4 ), Γα ,a
sample 12 18.02 13.25 44.56 23.23. (o, u5 )}, we know that O belongs to C5 , i.e., the zone measured
We can work out the following means and standard by us is covered by birch forest.
deviations: Remark 5.3: Although, only for example, using mean vec-
ō1 = 17.09, s1 = 0.77 ō2 = 13.17, s2 = 0.50 tors (µi1 , µi2 , µi3 , µi4 ), i = 1, 2, 3, 4, 5 and (ō1 , ō2 , ō4 , ō4 ) to rep-
ō3 = 45.10, s3 = 1.01 ō4 = 23.19, s4 = 0.33. resent, respectively, C1 , C2 , C3 , C4 and O, we perhaps also
judge that O belongs to C5 by the usual Euclidean metrics,
So, we can obtain the fuzzy four-cell number o = and we still emphasize that using fuzzy n-cell numbers to deal
(o1 , o2 , o3 , o4 ), i.e. with imprecise or uncertain quantities is better than using crisp
o(x1 , x2 , x3 , x4 ) = min{o1 (x1 ), o2 (x2 ), o3 (x3 ), o4 (x4 )} n-dimensional vectors. The following example (to simplify and
(x1 , x2 , x3 , x4 ) ∈ R4 shorten the problem, we consider only a 1-D case) will show
this.
which can be used to represent O, where Example 5.2: The following two classes of ferrous quantities

 x − 15.55 (in kilograms per hundred kilogram) of minerals come from two

 , if x ∈ [15.55, 17.09] different mine areas (denoted by A and B)

 1.54
o1 (x) = 18.63 − x A : 10.20, 11.76, 8.31, 9.02, 9.63, 8.33, 11.36, 12.30,

 , if x ∈ (17.09, 18.63]

 1.54
 12.03, 7.98
0, if x ∈
/ [15.55, 18.63]
 B : 52.33, 79.34, 34.51, 62.34, 82.26, 28.36, 17.37, 25.32,
 x − 12.17

 , if x ∈ [12.17, 13.17]

 1.00 10.11, 8.34.
o2 (x) = 14.17 − x

 , if x ∈ (13.17, 14.17] Suppose that one group (denoted by C) of minerals comes
 1.00

 from the one of A and B. The problem to be solved is to identify
0, if x ∈
/ [12.17, 14.17]
 if C comes from A or B. We take samples and acquire the
 x − 43.08 following data for C:

 , if x ∈ [43.08, 45.10]
 2.02

C : 18.20, 20.31, 76.02, 9.36, 28.56, 23.32, 15.36, 20.51,
o3 (x) = 47.12 − x

 , if x ∈ (45.10, 47.12]

 2.02 13.27, 20.32.

0, if x ∈
/ [43.08, 47.12]
 We can work out: µA = 10.09, σA = 1.67, µB =
 x − 22.53 40.03, σB = 27.43, µC = 24.52, and σC = 18.86.

 , if x ∈ [22.53, 23.19]

 0.66 If we use crisp means to represent A, B, and C, then we
o4 (x) = 23.79 − x have A = 10.09, B = 40.03, C = 24.52, and d(C, A) = 14.43

 , if x ∈ (23.19, 23.79]

 0.66 < 15.51 = d(C, B). If we regard d(C, A) < d(C, B) as ev-

0, if x ∈
/ [22.53, 23.79] idence, we can draw a conclusion that C comes from A.

4 u ∗ (1) v i∗ (1) u i (0) v i (0)

exp a2 i=1 αi u i (0) ui (t)dt − v i (0) vi (t)dt + u ∗ (1) ui (t)dt − v ∗ (1) vi (t)dt
i
i i
Γα ,a (u, v) = 1 a 1
4
i=1 αi 0 r[u i (r) + vi (r) + u i (r) + v i (r)]dr

4 1
× αi r[|(ui (r) − vi (r)| + |ui (r) − vi (r)|]dr
i=1 0
However, the conclusion does not accord with fact. We should metrics are unsuitable for use in finding the difference of two
note that although d(C, A) < d(C, B), the difference of d(C, A) fuzzy n-cell numbers, so we introduced the concepts of dif-
and d(C, B) is small. Furthermore, from the statistical data, we ference values ∆α ,a , Λα ,a , and Γα ,a , studied their properties,
can see that the ferrous quantities of minerals coming from A and showed the rationality for their use in characterizing the
are more coincident, but B and C are not. It is almost impos- degree of the difference of two fuzzy n-cell numbers by re-
sible that some minerals in C come from A, such as minerals marks and examples (see Section IV). Finally, in Section V,
with ferrous quantities 76.02 and 28.56. So we may judge that we put forward an algorithmic version of pattern recognition in
C comes from B. an imprecise or uncertain environment based on the metrics and
If we use fuzzy one-cell numbers to represent A, B, and C, difference values defined by us and gave examples to show the
then we have application (see Example 5.1) and rationality (see Example 5.2)

 x − 6.75 of the methods.

 , if x ∈ [6.75, 10.09]

 3.34
A(x) = 13.43 − x ACKNOWLEDGMENT

 , if x ∈ (10.09, 13.43]

 3.34
 The authors would like to thank the Associate Editor and
0, if x ∈/ [6.75, 13.43]
 reviewers for their valuable suggestions and comments that have
 x + 14.83 greatly improved the presentation of the paper.

 , if x ∈ [0, 40.03]

 54.86
B(x) = 94.89 − x

 , if x ∈ (40.03, 94.89] REFERENCES
 54.86


0, if x ∈/ [0, 94.89] [1] D. Butnariu, “Methods of solving optimization problems and linear equa-
 tions in the space of fuzzy vectors,” Libertas Math., vol. 17, pp. 1–7,
 x + 13.20

 , if x ∈ [0, 24.52] 1997.
 37.72
 [2] S. S. L. Chang and L. A. Zadeh, “On fuzzy mappings and control,” IEEE
Trans. Syst. Man, Cybern., vol. SMC-2, no. 1, pp. 30–34, Feb. 1972.
C(x) = 62.24 − x

 , if x ∈ (24.52, 62.24] [3] P. Diamond and P. Kloeden, Metric Space of Fuzzy Sets. Singapore:

 37.72 World Scientific, 1994.

0, if x ∈ / [0, 62.24] [4] D. Dubois and H. Prade, “Operations on fuzzy numbers,” Int. J. Syst. Sci.,
vol. 9, pp. 613–626, 1978.
A(r) = 3.34r + 6.75 [5] D. Dubois and H. Prade, “Towards fuzzy differential calculus—Part I, II,
III,” Fuzzy Sets Syst., vol. 8, pp. 1–17, 105–116, 225–234, 1982.
!
54.86r − 14.83, if r ∈ (0.27, 1] [6] R. Goetschel and W. Voxman, “Elementary calculus,” Fuzzy Sets Syst.,
B(r) = vol. 18, pp. 31–43, 1986.
0, if r ∈ [0, 0.27] [7] O. Kaleva, “Fuzzy differential equations,” Fuzzy Sets Syst., vol. 24,
! pp. 301–317, 1987.
37.72r − 13.20, if r ∈ (0.35, 1] [8] A. Kaufmann and M. M. Gupta, Introduction to Fuzzy Arithmetic—Theory
C(r) =
0, if r ∈ [0, 0.35] and Applications. Reinhold, NY: Van Nostrand, 1985.
[9] M. Miyakawa, K. Nakamura, J. Ramik, and I. G. Rosenberg, “Joint canon-
A(r) = 13.43 − 3.34r ical fuzzy numbers,” Fuzzy Sets Syst., vol. 53, pp. 39–49, 1993.
[10] K. Nakamura, “Canonical fuzzy number of dimension two and fuzzy
B(r) = 94.89 − 54.86r utility difference for understanding preferential judgements,” Inf. Sci.,
vol. 50, pp. 1–22, 1990.
C(r) = 62.24 − 37.72r [11] M. L. Puri and D. A. Ralescu, “Differentials for fuzzy functions,” J. Math.
Anal. Appl., vol. 91, pp. 552–558, 1983.
so [12] J. Ramik and K. Nakamura, “Canonical fuzzy numbers of dimension two,”
Fuzzy Sets Syst., vol. 54, pp. 167–181, 1993.
Γ(0.1,0.1) (C, A) = e3.27 (14.01/3.487) = 106.39 > 21.77 [13] S. Theodoridis and K. Koutroumbas, Pattern Recognition, 3rd ed. Am-
sterdam, The Netherlands: Elsevier, 2006.
[14] G. Wang, Y. Li, and C. Wen, “On fuzzy cell numbers and dimension fuzzy
= e1.178 (15.38/4.019) = Γ(0.1,0.1) (C, B). vectors,” Fuzzy Sets Syst., vol. 158, pp. 71–84, 2007.
[15] G. Wang and C. Wu, “Fuzzy cell numbers and the differential of fuzzy cell
Thus, we can affirm that C comes from B. number value mappings,” Fuzzy Sets Syst., vol. 130, pp. 367–381, 2002.
[16] G. Wang and C. Wu, “The integral over a directed line segment of fuzzy
VI. CONCLUSION mapping and its applications,” Int. J. Uncertainty Fuzzyness Knowl.-
Based Syst., vol. 12, pp. 543–556, 2004.
In this paper, we have suggested using fuzzy n-cell numbers [17] G. Wang and C. Zhao, “The measurability and (K)-integrability about
to represent imprecise or uncertain multichannel digital signals fuzzy n-cell number value mappings,” J. Fuzzy Math., vol. 13, pp. 117–
and have put forward a method (see the first or second step 128, 2005.
[18] C. Wu and G. Wang, “Convergence of sequences of fuzzy numbers and
of the algorithmic version in Section V, or see Example 2.1) fixed point theorems for inceasing fuzzy mappings and application,” Fuzzy
of constructing such fuzzy n-cell numbers. Although the met- Sets Syst., vol. 130, pp. 383–390, 2002.
rics D and DL have been studied formerly in [14] and [15], [19] T. Wang, Y. Chen, and S. Tong, “Fuzzy reasoning models and algorithms
on type-2 fuzzy sets,” Int. J. Innovative Comput. Inf. Control, vol. 4,
in view of the roughness of D and DL , we have defined two no. 10, pp. 2451–2460, 2008.
new metrics on fuzzy n-cell number space in order that they [20] M. Shimakawa, “A proposal of extension fuzzy reasoning method,” Int.
can better characterize the degree of the difference of two J. Innovative Comput. Inf. Control., vol. 4, no. 10, pp. 2603–2615, 2008.
[21] K. Shinkai, “Decision analysis of fuzzy partition tree applying fuzzy
objects in some imprecise or uncertain environment, and we theory,” Int. J. Innovative Comput. Inf. Control., vol. 4, no. 10, pp. 2581–
have studied their properties (Section III). In some applications, 2594, 2008.
Guixiang Wang received the B.Sc. degree in mathe- Paul Messenger received the MMath Master’s degree in mathematics and the
matics from Hebei Normal University, Shijiazhuang, Ph.D. degree in relativity theory from the University of Glamorgan, Wales, U.K.,
China, in 1982, the M.Sc. degree in mathematics from in 2000 and 2005, respectively.
Hebei University, Baoding, China, in 1989, and the He is currently a Full-Time Lecturer in applied mathematics at the University
Ph.D. degree in mathematics from Harbin Institute of of Glamorgan. His research interests are relativity theory, signal processing and
Technology, Harbin, China, in 2002. related topics, and severe weather mathematics. He also does some outreach
From 1992 to 1999, he was an Associate Profes- work to local schools, explaining topics like relativity, etc., to local school
sor at the Agricultural University of Hebei, Baoding, children.
China. In 2002, he joined the Hangzhou Dianzi Uni-
versity, Hangzhou, China, as a Professor. From 2003
to 2005, he was a Postdoctoral Fellow at Harbin Engi-
neering University, Harbin, China. From 2007 to 2008, he was a Visiting Senior
Fellow at the University of Glamorgan, Pontypridd, U.K. His current research
interests include fuzzy set theory and application, nonlinear systems, signal and
information processing, data fusion, and pattern recognition.
Peng Shi (SM’97) received the B.Sc. degree in math-

ematics from Harbin Institute of Technology, Harbin,
China, in 1982, the M.E. degree in control theory from
Harbin Engineering University, Harbin, in 1985, the
Ph.D. degree in electrical engineering from the Uni-
versity of Newcastle, Newcastle, Australia, in 1994,
the Ph.D. degree in mathematics from the Univer-
sity of South Australia, Adelaide, Australia, in 1998,
and the Doctor of Science (Higher Doctorate) degree
from the University of Glamorgan, Pontypridd, U.K.,
in 2006.
He was a Lecturer at Heilongjiang University (1985–1989), the University
of South Australia (1997–1999), and a Senior Scientist at the Defence Science
and Technology Organisation, Department of Defence, Australia (1999–2005).
In 2004, he joined the University of Glamorgan as a Professor. His current re-
search interests include robust control and filtering, fault detection techniques,
Markov decision processes, and optimization techniques. He has authored or
coauthored a number of papers in these areas. In addition, he is a coauthor of the
two research monographs: Fuzzy Control and Filtering Design for Uncertain
Fuzzy Systems (Berlin, Jermany, Springer-Verlag, 2006) and Methodologies for
Control of Jump Time-Delay Systems (Boston, MA, Kluwer, 2003). He is cur-
rently the Editor-in-Chief of the International Journal of Innovative Computing
and Information and Control.
Prof. Shi is an Associate Editor for a number of other journals, such as the
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, the IEEE TRANSACTIONS ON
SYSTEMS, MAN AND CYBERNETICS, PART B, and the IEEE TRANSACTIONS ON
FUZZY SYSTEMS. He is a Fellow of the Institute of Mathematics and its Appli-
cations (U.K.).
H∞ Fuzzy Control for Systems With Repeated Scalar

Nonlinearities and Random Packet Losses
Hongli Dong, Zidong Wang, Senior Member, IEEE, and Huijun Gao, Member, IEEE
Abstract—This paper is concerned with the H∞ fuzzy control tice based on this technique is as follows. Fuzzy models have
problem for a class of systems with repeated scalar nonlinearities provided an approach to represent complex nonlinear systems
and random packet losses. A modified Takagi–Sugeno (T–S) fuzzy to a set of linear local models by using fuzzy sets and fuzzy
model is proposed in which the consequent parts are composed of
a set of discrete-time state equations containing a repeated scalar reasoning. The overall fuzzy model of the system is achieved by
nonlinearity. Such a model can describe some well-known nonlin- smoothly blending these local models together through member-
ear systems such as recurrent neural networks. The measurement ship functions. It has a convenient and simple dynamic structure
transmission between the plant and controller is assumed to be im- such that the existing results for linear systems theory can be
perfect and a stochastic variable satisfying the Bernoulli random readily extended for this class of nonlinear systems and, as a
binary distribution is utilized to represent the phenomenon of ran-
dom packet losses. Attention is focused on the analysis and design of result, a great number of important results has been reported
H∞ fuzzy controllers with the same repeated scalar nonlinearities in the literature. For example, the problem of stability analysis
such that the closed-loop T–S fuzzy control system is stochastically has been investigated in [2], [3], [9], [10], and [25], and the
stable and preserves a guaranteed H∞ performance. Sufficient stabilizing as well as H∞ control designs have been reported
conditions are obtained for the existence of admissible controllers, in [1], [4], [8], [12], [19], [21], [22], [24] and [29].
and the cone complementarity linearization procedure is employed
to cast the controller design problem into a sequential minimiza- The T–S fuzzy model has been widely employed to rep-
tion one subject to linear matrix inequalities, which can be readily resent or approximate a nonlinear system, which is described
solved by using standard numerical software. Two examples are by a family of fuzzy IF–THEN rules that represent local lin-
given to illustrate the effectiveness of the proposed design method. ear input–output relations of the system. Nevertheless, the local
Index Terms—Diagonally dominant matrix, fuzzy systems, H∞ model is not necessarily a linear one but sometimes a “simple” or
control, linear matrix inequality (LMI), random packet losses, re- “slightly” nonlinear system whose dynamics can be thoroughly
peated scalar nonlinearity. investigated. A good example of such a simple nonlinear sys-
tem is the recurrent neural network that involves a nonlinear but
I. INTRODUCTION known activation function. Therefore, there has appeared initial
INCE the concept of fuzzy sets was introduced by Zadeh research interest focusing on the extended T–S model whose sys-
S in 1965, fuzzy logic control has developed into one of
the most important and successful branch of automation and
tem dynamics is captured by a set of fuzzy implications which
characterize local relations in the state space [15], [16], [30]. In
control theory. In the past few decades, the fuzzy logic theory this case, the local dynamics of each fuzzy rule is expressed by a
has been demonstrated to be effective in dealing with a vari- well-studied nonlinear system, and the overall fuzzy model can
ety of complex nonlinear systems, which has therefore received be achieved by fuzzy “blending” of these simple nonlinear local
a rapidly growing interest in the literature. In particular, the systems. For example, a modified T–S fuzzy model has been
control technique based on the so-called Takagi–Sugeno (T–S) proposed in [16] in which the consequent parts are composed of
fuzzy model has attracted much attention. The common prac- a set of stochastic Hopfield neural networks with time-varying
delays, and a stability criterion has been derived in terms of
linear matrix inequalities (LMIs). The results of [16] have then
Manuscript received September 18, 2008; revised November 23, 2008; been extended in [15] to deal with the stability analysis prob-
accepted January 13, 2009. First published February 2, 2009; current ver- lem for T–S fuzzy cellular neural networks with time-varying
sion published April 1, 2009. This study was supported in part by the Engi-
neering and Physical Sciences Research Council (EPSRC), U.K. under Grant delays. Motivated by the works in [15], [16] and [30], in this
GR/S27658/01, in part by the Royal Society, U.K., in part by the National Out- paper, we will consider a more general yet well-known nonlin-
standing Youth Science Fund under Grant 60825303, in part by the National earity, namely repeated scalar nonlinearity [5], [7], [11] which
973 Program of China under Grant 2009CB320600, in part by the Research
Found for the Doctoral Programme of Higher Education of China under Grant covers some typical classes of nonlinearities such as the semi-
20070213084, in part by the Heilongjiang Outstanding Youth Science Fund un- linear function, the hyperbolic tangent function that has been
der Grant JC200809, and in part by the Alexander von Humboldt Foundation, extensively used for activation function in neural networks, the
Germany.
H. Dong is with the Space Control and Inertial Technology Research Center, sine function, etc.
Harbin Institute of Technology, Harbin 150001, China, and also with the College On the other hand, we notice that, in almost all aforemen-
of Electrical and Information Engineering, Daqing Petroleum Institute, Daqing tioned literature concerning T–S fuzzy control systems, it has
163318, China (e-mail: shiningdhl@gmail.com).
Z. Wang is with the Department of Information Systems and Comput- been implicitly assumed that the communication between the
ing, Brunel University, Uxbridge, Middlesex UB8 3PH, U.K. (e-mail: zidong. physical plant and controller is perfect, i.e., the signals trans-
wang@brunel.ac.uk). mitted from the plant always arrive at the controller without
H. Gao is with the Space Control and Inertial Technology Research Cen-
ter, Harbin Institute of Technology, Harbin 150001, China (e-mail: hjgao@hit. any information losses, and vice versa. Such an assumption,
edu.cn). however, is not always true in practice. For example, due to
Digital Object Identifier 10.1109/TFUZZ.2009.2014223 the unreliability of the network links, a networked control
1063-6706/$25.00 © 2009 IEEE
DONG et al.: H ∞ FUZZY CONTROL FOR SYSTEMS WITH REPEATED SCALAR NONLINEARITIES AND RANDOM PACKET LOSSES 441
system (NCS) typically exhibits significant communication de- vex functions φ : R+ → R+ such that φ(0) = 0 and φ(x) > 0
lays and data loss across the network, which gives rise to con- for x > 0. Matrices, if their dimensions are not explicitly stated,
siderable research attention on how to design control systems are assumed to be compatible for algebraic operations.
with the simultaneous consideration of the data loss issue (also
called package dropout or missing measurements). For exam- II. PROBLEM FORMULATION
ple, in [27] and [28], the control and filtering problems for
uncertain discrete-time stochastic systems with missing mea- A. Physical Plant
surements have been investigated when data travel along un- In this paper, we consider the following discrete-time fuzzy
reliable communication channels in a large, wireless, multihop systems with repeated scalar nonlinearities:
sensor network. The problem of optimally controlling a linear Plant rule i: IF θ1 (k) is Mi1 and θ2 (k) is Mi2 and · · · and
discrete-time plant has been studied in [17] when some of the θp (k) is Mip THEN
measurement and control packets were missing. However, to the 
xk +1 = Ai f (xk ) + B2i uk + B1i wk  
best of the authors’ knowledge, the T–S fuzzy control problem
for nonlinear systems under unreliable communication links zk = Ci f (xk ) + D2i uk + D1i wk (1)


has not been fully investigated, which motivates our present i = 1, . . . , r
research.
In this paper, the H∞ fuzzy control problem is addressed where Mij is the fuzzy set; xk ∈ Rn represents the state vector;
for a class of nonlinear systems under unreliable communi- uk ∈ Rm is the input vector; wk ∈ Rp is the exogenous distur-
cation links. The nonlinear system is described by a discrete- bance input which belongs to l2 [0, ∞); zk ∈ Rq is the controlled
time state equation involving a repeated scalar nonlinearity that output; Ai , B2i , B1i , Ci , D2i , and D1i are all constant matrices
typically appears in recurrent neural networks. The commu- with compatible dimensions; r is the number of IF-THEN rules;
nication links, existing between the plant and controller, are θk = [θ1 (k), θ2 (k), . . . , θp (k)] is the premise variable vector. It
assumed to be imperfect, and the packet loss phenomena are is assumed that the premise variables do not depend on the input
modeled by a Bernoulli random binary distributed white se- variable uk , which is needed to avoid a complicated defuzzifi-
quence with a known conditional probability. The objective is cation process of fuzzy controllers. f is a nonlinear function
to analyze and design a fuzzy controller such that the closed-loop satisfying the following assumption as in [6].
fuzzy control system is stochastically stable while preserving a Assumption 1: The nonlinear function f : R → R in system
guaranteed H∞ performance. Sufficient conditions that involve (1) satisfies
matrix equalities are obtained for the existence of admissible ∀a, b∈R |f (a) + f (b)| ≤ |a + b| . (2)
controllers, and the CCL procedure is employed to cast the non-
convex feasibility problem into a sequential minimization prob- In the sequel, for the vector x = [x1 x2 · · · xn ]T , we
lem subject to LMIs, which can then be readily solved by using denote
standard numerical software. Two numerical examples are given
f (x) = [f (x1 ) f (x2 ) ··· f (xn )]T .
to illustrate the effectiveness of the proposed design method.
The rest of this paper is organized as follows. Section II for- Remark 1: The model (1) is called a system with repeated
mulates the problem under consideration. The stability condi- scalar nonlinearity [5], [7], [11]. Note that f is odd (by putting
tion and H∞ performance of the closed-loop T–S fuzzy control b = −a) and 1-Lipschitz (by putting b = −b). Therefore, f en-
system are given in Section III. The H∞ fuzzy controller design capsulates some typical classes of nonlinearities, such as
problem is solved in Section IV. The validity of this approach 1) the semilinear function (i.e., the standard saturation
is demonstrated by illustrative examples in Section V. Finally, sat(s) := s if |s| ≤ 1 and sat(s) := sgn(s) if |s| > 1);
in Section VI, the conclusion is given. 2) the hyperbolic tangent function that has been extensively
Notation. The notation used in the paper is fairly standard. used for activation function in neural networks;
The superscript “T ” stands for matrix transposition, Rn de- 3) the sine function, etc.
notes the n-dimensional Euclidean space, Rm ×n is the set of all Given a pair of (xk , uk ), the final outputs of the fuzzy system
real matrices of dimension m × n; I and 0 represent the identity are inferred as follows:

matrix and zero matrix, respectively. The notation P > 0 means r

that P is real symmetric and positive definite; tr(M ) refers to xk +1 = hi (θk )[Ai f (xk ) + B2i uk + B1i wk ]  


A refers to the norm
the trace of the matrix M ; the notation i=1
(3)
of a matrix A defined by A = tr(AT A) and · 2 stands r 

hi (θk )[Ci f (xk ) + D2i uk + D1i wk ]  
for the usual l2 norm. In symmetric block matrices or com- zk = 
i=1
plex matrix expressions, we use an asterisk (∗) to represent a
term that is induced by symmetry, and diag {. . .} stands for where the fuzzy basis functions are given by
a block-diagonal matrix. In addition, E{x} and E{ x| y} will, ϑi (θk )
respectively, mean expectation of x and expectation of x con- hi (θk ) = r
ditional on y. The set of all nonnegative integers is denoted by i=1 ϑi (θk )
I+ and the set of all nonnegative real numbers is represented by with ϑi (θk ) = Πpj=1 Mij (θj (k)). Mij (θj (k)) represents the
R+ . Ξ denotes the class of all continuous nondecreasing con- grade of membership of θj (k) in Mij . Here, ϑi (θk ) has the
following basic property: and βk = 1 are true, and ρk = 0 otherwise. Then, we have

r
ϑi (θk ) ≥ 0, i = 1, 2, . . . , r, ϑi (θk ) > 0, ∀k Prob{ρk = 1} = E {ρk } := ρ̄ = ᾱβ̄
i=1
Prob{ρk = 0} = 1 − ρ̄ = 1 − ᾱβ̄.
and therefore

r

r uk = hi (θk )ρk Ki f (xk ). (6)
hi (θk ) ≥ 0, i = 1, 2, . . . , r, hi (θk ) = 1 ∀k. i=1
i=1
The closed-loop T–S fuzzy control system can now be obtained
B. Controller from (3) and (6) that
In this paper, we consider the following fuzzy control law for

r
r 
the fuzzy system (3): xk +1 = hi (θk )hj (θk )[Aij f (xk ) + B1i wk ] 



Controller rule i: IF θ1 (k) is Mi1 and θ2 (k) is Mi2 and · · · i=1 j =1

and θp (k) is Mip THEN (7)

r
r 

uck = Ki fˆ(xk ), i = 1, 2, . . . , r. zk = hi (θk )hj (θk )[Cij f (xk ) + D1i wk ] 


i=1 j =1
Here, fˆ(xk ) ∈ Rn is the input to the controller; uck ∈ Rm is the
output of the controller; Ki are the gain matrices of the controller where Aij = Ai + ρ̄B2i Kj + ρ̃k B2i Kj , Cij = Ci + ρ̄D2i Kj +
to be designed. Hence, the controller can be represented by the ρ̃k D2i Kj , and ρ̃k = ρk − ρ̄. It is clear that E{ρ̃k } = 0 and
following input–output form: E{ρ̃2k } = ρ̄(1 − ρ̄).
Before formulating the problem to be investigated, we first

r
introduce the following definitions and lemmas.
uck = hi (θk )Ki fˆ(xk ). (4)
Definition 1 [23]: The solution xk = 0 of the closed-loop T–S
i=1
fuzzy control system in (7) with wk ≡ 0 is said to be stochasti-
C. Communication Links cally stable if, for any ε > 0, there exists a δ > 0 such that
Due to the existence of the packet losses between the physical E {xk } < ε (8)
plant and controller, the measurement of the plant is probably not
equivalent to the input to the controller [i.e., f (xk )
= fˆ(xk )], whenever k ∈ I+ and x0 < δ.
and the output of the controller is probably not equivalent to the
input of the plant (i.e., uk
= uck ). As a result, we model the Definition 2: A square matrix P = [pij ] ∈ Rn ×n is called
packet losses phenomena via a stochastic approach as follows: diagonally dominant if for all i = 1, . . . , n

fˆ(xk ) = αk f (xk ) uk = βk uck pii ≥ |pij | . (9)
where {αk } and {βk } are two independent Bernoulli processes. j
= i
{αk } models the unreliable nature of the link from the sensor
to the controller, and {βk } models that from the controller to Lemma 1 [6]: If P > 0 is diagonally dominant, then for all
the actuator. Obviously, αk = 0 holds when the communication nonlinear functions f satisfying (2), the following inequality
link fails (i.e., data are lost), and αk = 1 means successful trans- holds for all xk ∈ Rn :
mission. The same happens for βk . A natural assumption on the
sequence {αk } and {βk } can be made as follows: f T (xk )P f (xk ) ≤ xTk P xk . (10)
Prob {αk = 1} = E {αk } = ᾱ Prob {αk = 0} = 1 − ᾱ Remark 2: It will be seen later that the purpose of requiring
Prob {βk = 1} = E {βk } = β̄ Prob {βk = 0} = 1 − β̄. the matrix P to satisfy (10) is to admit the quadratic Lyapunov
function V (xk ) = xTk P xk .
Based on this, we have Lemma 2 [23]: If there exist a Lyapunov function V (xk ) and

r a function φ(x) ∈ Ξ satisfying the following conditions
uk = hi (θk )βk αk Ki f (xk ). (5)
i=1 V (0) = 0 (11)
Note that such a stochastic Bernoulli approach has been ex- φ(xk ) ≤ V (xk ) (12)
tensively used for dealing with data missing problems; see,
e.g., [13] and [26] and the references therein. E {V (xk +1 )} − E {V (xk )} < 0, k ∈ I+ (13)
D. Closed-Loop System then the solution xk = 0 of the closed-loop T–S fuzzy control
system in (7) with wk ≡ 0 is stochastically stable.
In this paper, we introduce another Bernoulli process {ρk } Consider the fuzzy control problem in the presence of packet
∆
with ρk = αk βk . It is easy to know that ρk = 1 when both αk = 1 losses phenomena and suppose the parameter ρ̄ describing
intermittent transmission is known. We are now in a position to When wk ≡ 0, the difference of the Lyapunov function is cal-
state the problem of H∞ fuzzy control for systems with repeated culated as
scalar nonlinearities and random packet losses as follows.
Problem H∞ fuzzy control with data loss (HFCDL): Given a
scalar γ > 0, design a controller in the form of (4) such that
E{∆V (xk )} = E { V (xk +1 )| xk } − V (xk )
1) (stochastic stability) the closed-loop T–S fuzzy control  T
system in (7) is stochastically stable in the sense of 
 r r
Definition 1; =E  hi (θk )hj (θk )Aij f (xk ) P
2) (H∞ performance) under zero initial condition, the con- 
 i=1 j =1
trolled output zk satisfies   

r r

zE ≤ γw2 (14) × hi (θk )hj (θk )Aij f (xk ) xk
i=1 j =1 
where
− xTk P xk
  r r r r
 ∞


zE = E zkT zk =E hi (θk )hj (θk )hs (θk )ht (θk )
  i=1 j =1 s=1 t=1
k =0

and · 2 stands for the usual l2 norm. ×f T
(xk )ATij P Ast f (xk ) xk − xTk P xk
If the earlier two conditions are satisfied, the closed-loop
T–S fuzzy control system is said to be stochastically stable with

r
r
r
r
a guaranteed H∞ performance γ, and the problem HFCDL is =E hi (θk )hj (θk )hs (θk )ht (θk )
solved. i=1 j =1 s=1 t=1
T
Aij + Aj i Ast + Ats
III. H∞ FUZZY CONTROL PERFORMANCE ANALYSIS × f (xk )
T
P
2 2
In this section, the problem HFCDL formulated in the previ-
ous section will be tackled via a quadratic approach described × f (xk )| xk − xTk P xk
in the following theorem.
Theorem 1: Consider the fuzzy system in (3) and suppose 
r
r
the gain matrices Ki (i = 1, . . . , r) of the controllers in (4)
≤E hi (θk )hj (θk )f T (xk )
are given. The closed-loop fuzzy system in (7) is stochastically 
i=1 j =1
stable with a guaranteed H∞ performance γ if there exists a
T
positive diagonally dominant matrix P satisfying Aij + Aj i Aij + Aj i
× P
2 2
ΨTii P̄ Ψii + ΛTii Λii − L̄ < 0, i = 1, 2, . . . , r (15)
(Ψij + Ψj i )T P̄ (Ψij + Ψj i ) + (Λij + Λj i )T (Λij + Λj i ) × f (xk )| xk − xTk P xk . (18)
− 4L̄ < 0, 1≤i<j≤r (16)
where Note that in the earlier inequality, the elementary inequality of

2aT b ≤ aT a + bT b for a, b ∈ Rn has been used.
According to Lemma 1, we have
Ai + ρ̄B2i Kj B1i P 0
Ψij = L̄ =
qB2i Kj 0 0 γ2 I

r
r
Ci + ρ̄D2i Kj D1i P 0 E{∆V (xk )} ≤ E hi (θk )hj (θk )f T (xk )
Λij = P̄ =
qD2i Kj 0 0 P i=1 j =1
T
q = ρ̄(1 − ρ̄). Aij + Aj i Aij + Aj i
× P −P
2 2
Proof: In order to show that the fuzzy system in (7) is stochas-
tically stable with a guaranteed H∞ performance γ under condi-
tions (15) and (16), we define the following Lyapunov function × f (xk )| xk
candidate

r
= f T (xk ) h2i (θk )(ΠTii P̄ Πii − P )f (xk )
V (xk ) = xTk P xk . (17)
i=1
r
1 Ḡij + Ḡj i
+ f T (xk ) hi (θk )hj (θk ) × ηk ξk
2 i,j =1,i< j
2
r r T
× (Πij + Πj i )T P̄ (Πij + Πj i ) − 4P f (xk ) H̄ij + H̄j i
+E hi (θk )hj (θk )ηk T
i=1 j =1
2
where

Ai + ρ̄B2i Kj H̄ij + H̄j i
Πij = . × ηk ξk − ηkT L̄ηk
qB2i Kj 2
By the Schur complement lemma, we know from inequalities
r
r T
Ψij + Ψj i
(15) and (16) that = hi (θk )hj (θk )ηkT P̄
i=1 j =1
2
ΠTii P̄ Πii − P < 0, i = 1, 2, . . . , r

r
r
Ψij + Ψj i
(Πij + Πj i ) P̄ (Πij + Πj i ) − 4P < 0,
T
1 ≤ i < j ≤ r. × ηk + hi (θk )hj (θk )ηkT
2 i=1 j =1
Thus, we have
T
E {V (xk +1 )} − E {V (xk )} < 0 Λij + Λj i Λij + Λj i
× ηk − ηkT L̄ηk
2 2
which satisfies (13). Taking φ(xk ) = λm in (P )x2k such that
φ(·) ∈ Ξ, we obtain
r
1
= ηkT h2i (θk )(ΨTii P̄ Ψii + ΛTii Λii − L̄)ηk + ηkT
2
φ(xk ) = λm in (P )xk 2 = λm in (P )xTk xk i=1
≤ xTk P xk = V (xk )
r
× hi (θk )hj (θk )
which satisfies (12). Considering V (0) = 0, it follows readily i,j =1,i< j
from Lemma 2 that the closed-loop system in (7) with wk ≡ 0
× (Ψij + Ψj i )T P̄ (Ψij + Ψj i )
is stochastically stable.
Next, the H∞ performance criteria for the closed-loop system + (Λij + Λj i )T (Λij + Λj i ) − 4L̄ ηk
in (7) will be established. Assuming zero initial conditions, an
index is introduced as follows: From inequalities (15) and (16), we know that J¯ ≤ 0 and, ac-
cording to Lemma 1, we have
J¯ = E { V (xk +1 )| ξk } + E zkT zk ξk
E xTk +1 P xk +1 ξk + E zkT zk ξk − γ 2 wkT wk − xTk P xk ≤ 0.
− γ 2 wkT wk − f T (xk )P f (xk ).
Defining Taking mathematical expectation on both sides, we obtain

ξk = [ xTk wkT ]T ηk = [ f T (xk ) wkT ]T E xTk +1 P xk +1 ξk − E xTk P xk ξk + E{zkT zk }
Ḡij = [ Aij B1i ] H̄ij = [ Cij D1i ] − γ 2 wkT wk ≤ 0.
similar to the derivation of (18), we have For k = 0, 1, 2, . . . , summing up both sides under zero initial
 T 
 condition and considering E{xT∞ P x∞ } ≥ 0, we arrive at
 r r r r
J¯ = E  hi (θk )hj (θk )Ḡij ηk  P  hi (θk ) ∞ ∞

 i=1 j =1 i=1 j =1 E z k zk −
T
γ 2 wkT wk ≤ 0
  k =0 k =0


× hj (θk )Ḡij ηk  ξk which is equivalent to (14). The proof is now completed.
 Remark 3: In Theorem 1, with given controller gain and dis-
 T turbance attenuation level γ, we obtain the stochastic stability

 r r conditions of the nominal fuzzy system (7), which are repre-
+E  hi (θk )hj (θk )H̄ij ηk  sented via a set of matrix inequalities in (15) and (16). We will

 i=1 j =1 show later in the next section that such inequalities can be con-
   verted into LMIs when designing the actual controllers. Note
r 
r
that the feasibility of LMIs can be easily checked by using the
× hi (θk )hj (θk )H̄ij ηk  ξk − ηkT L̄ηk MATLAB LMI toolbox.
i=1 j =1  Remark 4: Let us now consider the standard H∞ performance
T criterion for a discrete-time fuzzy closed-loop system with per-

r
r
Ḡij + Ḡj i
≤E hi (θk )hj (θk )ηkT P fect communication links between the plant and controller. In
i=1 j =1
2 this case, we have ρ̄ = 1 in (7) and then the inequalities (15)
and (16) reduce to Υ22 = diag {−L, −L, −I, −I}

XiiT P̄ Xii + YiiT Yii − L̄ < 0, i = 1, 2, . . . , r (19) B̄2 = B2i Mj + B2j Mi D̄2 = D2i Mj + D2j Mi
(Xij + Xj i )T P̄ (Xij + Xj i ) + (Yij + Yj i )T (Yij + Yj i ) 1 ≤ i < j ≤ r.
− 4L̄ < 0, 1≤i<j≤r (20) Furthermore, if the earlier conditions have feasible solutions,
the gain Ki of the subsystem controller in (4) is given by
where
Ki = Mi L−1 . (27)
Xij = [ Ai + B2i Kj B1i ] Yij = [ Ci + D2i Kj D1i ] .
Proof: From Theorem 1, we know that the closed-loop system
Later, we will show via simulation that, in the case there indeed in (7) is stochastically stable with a guaranteed H∞ performance
exist random packet losses, the main results given by Theorem 1 γ if there exists a diagonally dominant matrix P > 0 satisfying
will provide much improved performance over the standard (15) and (16). By the Schur complement, the following inequal-
H∞ approach that does not take into account the data missing ities are obtained:
problem.  
−P ∗ ∗ ∗ ∗ ∗
 
IV. H∞ FUZZY CONTROLLER DESIGN  0 −γ 2 I ∗ ∗ ∗ ∗ 
 
 Ai + ρ̄B2i Ki B1i −P −1 ∗ ∗ ∗ 
In this section, we aim at designing a controller in the form of  
 <0
(4) based on Theorem 1, i.e., we are interested in determining the  qB2i Ki 0 0 −P −1
∗ ∗ 
 
controller parameters such that the closed-loop fuzzy system in  
 Ci + ρ̄D2i Ki D1i 0 0 −I ∗ 
(7) is stochastically stable with a guaranteed H∞ performance.
The following theorem provides sufficient conditions for the qD2i Ki 0 0 0 0 −I
existence of such H∞ fuzzy controller for system (7). (28)
Theorem 2: Consider the fuzzy system in (3). There exists a
Υ̂11 ∗
state-feedback controller in the form of (4) such that the closed- <0 (29)
loop system in (7) is stochastically stable with a guaranteed H∞ Υ̂21 Υ̂22

performance γ, if there exist matrices 0 < P = [pij ], L > 0, where
T

Mi , i = 1, . . . , r, R = R = [rij ] satisfying −4P ∗
Υ̂11 =
  0 −4γ 2 I
−L ∗ ∗ ∗ ∗ ∗  
  (Ai + Aj ) + ρ̄B̂2 B1i + B1j
 0 −γ I
2
∗ ∗ ∗ ∗ 
   
 Ai L + ρ̄B2i Mi −L ∗ ∗ ∗   q B̂2 0
 B1i  Υ̂21 = 
 (C + C ) + ρ̄D̂


 qB2i Mi 0 0 −L ∗
<0
∗   i j 2 D1i + D1j 

 
 
 Ci L + ρ̄D2i Mi D1i 0 0 −I ∗  q D̂2 0
−1 −1

qD2i Mi 0 0 0 0 −I Υ̂22 = diag −P , −P , −I, −I
i = 1, 2, . . . , r (21) B̂2 = B2i Kj + B2j Ki D̂2 = D2i Kj + D2j Ki .

Υ11 ∗
<0 (22) Performing congruence
transformations
to inequalities (28) and
Υ21 Υ22 (29) by diag P −1 , I, I, I, I, I , we have
 
pii − (pij + 2rij ) ≥ 0 (23) −P −1 ∗ ∗ ∗ ∗ ∗
j
= i  
 0 −γ 2 I ∗ ∗ ∗ ∗ 
rij ≥ 0 ∀i
= j  
(24)  (Ai + ρ̄B2i Ki )P −1 B1i −P −1 ∗ ∗ ∗ 
 
pij + rij ≥ 0 ∀i
= j  
(25)  qB K P −1
0 0 −P −1
∗ ∗ 
 2i i 
 
PL = I (26)  (Ci + ρ̄D2i Ki )P −1 D1i 0 0 −I ∗ 
where qD2i Ki P −1 0 0 0 0 −I
<0
−4L ∗
Υ11 =
0 −4γ 2 I Υ̌11 ∗
  <0
(Ai + Aj )L + ρ̄B̄2 B1i + B1j ˇ 21 Υ̌22
Υ̂
 q B̄2 0 
  where
Υ21 =  
 (Ci + Cj )L + ρ̄D̄2 D1i + D1j  −4P −1 ∗
Υ̌11 =
q D̄2 0 0 −4γ 2 I
 
(Ai + Aj )P −1 + ρ̄B̌2 B1i + B1j Remark 5: As is well known, the packet dropout problem may
  occurs in the area of networked fault detection and isolation
 qB̌2 0 
Υ̌21 =



(FDI); see, e.g., [18] and [20] and the references therein. One
 (Ci + Cj )P −1 + ρ̄Ď2 D1i + D1j  of the future research topics would be the extension of our main
qĎ2 0 results to network-based FDI and filtering problems.
Remark 6: Our main results are based on the LMI conditions.
Υ̌22 = diag −P −1 , −P −1 , −I, −I The LMI control toolbox implements state-of-the-art interior-
point LMI solvers. While these solvers are significantly faster
B̌2 = (B2i Kj + B2j Ki )P −1 than classical convex optimization algorithms, it should be kept
Ď2 = (D2i Kj + D2j Ki )P −1 . in mind that the complexity of LMI computations remains higher
than that of solving, say, a Riccati equation. For instance, prob-
Defining L = P −1 and Mi = Ki P −1 , we can obtain (21) and lems with a thousand design variables typically take over an hour
(22) readily. Furthermore, from (23) to (25), we have on today’s workstations. However, research on LMI optimiza-
tion is a very active area in the applied math, optimization and
pii ≥ (pij + 2rij ) = (|pij + rij | + |−rij |) ≥ |pij | the operations research community, and substantial speedups
j
= i j
= i j
= i can be expected in the future.
which guarantees the positive definite matrix P to be diagonally
dominant, and the proof is then complete. V. ILLUSTRATIVE EXAMPLES
It is worth noting that, by far, we are unable to apply the
LMI approach in the design of controller because of the matrix In this section, two simulation examples are presented to
equality in Theorem 2. Fortunately, this problem can be ad- illustrate the fuzzy controller design method developed in this
dressed with help from the cone complementarity linearization paper.
(CCL) algorithm proposed in [14]. The basic idea behind the Example 1: Consider a T–S fuzzy model (1) with repeated
CCL algorithm is that if the LMI scalar nonlinearities and random packet losses. The rules are
given as follows:
P I Plant rule 1: IF f1 (xk ) is h1 (f1 (xk )) THEN
≥0
I L
xk +1 = A1 f (xk ) + B21 uk + B11 wk
is feasible in the n × n matrix variables L > 0 and P > 0, then (31)
zk = C1 f (xk ) + D21 uk + D11 wk .
tr(P L) ≥ n; and tr(P L) = n if and only if P L = I. Based
on this, it is likely to solve the equalities in (26) by using the Plant rule 2: IF f1 (xk ) is h2 (f1 (xk )) THEN
CCL algorithm. In view of this observation, we put forward
the following nonlinear minimization problem involving LMI xk +1 = A2 f (xk ) + B22 uk + B12 wk
conditions instead of the original nonconvex feasibility problem (32)
formulated in Theorem 2. zk = C2 f (xk ) + D22 uk + D12 wk .
The nonlinear minimization problem: min tr(P L) subject to
Controller rule 1: IF f1 (xk ) is h1 (f1 (xk )) THEN uck =
(21)–(25) and
K1 fˆ(xk ).
P I Controller rule 2: IF f1 (xk ) is h2 (f1 (xk )) THEN uck =
≥ 0. (30)
I L K2 fˆ(xk ).
The final outputs of the fuzzy system are inferred as follows:
If the solution of min tr(P L) subject to (21)–(25) exists and
min tr(P L) = n, then the conditions in Theorem 2 are solvable.
2
Finally, the following algorithm is suggested to solve the xk +1 = hi (f1 (xk ))[Ai f (xk ) + B2i uk + B1i wk ]
earlier problem. i=1
(33)
Algorithm HinfFC (HinfFC: H∞ Fuzzy Control)
2
Step 1: Find a feasible set (P(0) , L(0) , Mi (0) , R(0) ) satisfy- zk = hi (f1 (xk ))[Ci f (xk ) + D2i uk + D1i wk ].
ing (21)–(25) and (30). Set q = 0. i=1
Step 2: According to (21)–(25) and (30), solve the LMI
The model parameters are given as follows:
problem: min tr(P L(q ) + P(q ) L).
Step 3: Substitute the obtained matrix variables (P, L, Mi ,    
1.0 0.31 0 0.2 0 0
R) into (15) and (16). If conditions (15) and (16)    
are satisfied with |tr(P L) − n| < τ for some suffi- A1 =  0 0.33 0.21  C1 =  0 0 0 
ciently small scalar τ > 0, then output the feasible 0 0 −0.52 0 0 0.1
solutions. Exit.    
0.8 −0.38 0 1 1
Step 4: If K > N , where N is the maximum number of    
iterations allowed. Exit. Else, set q = q + 1 and go A2 =  −0.2 0 0.21  B21 = 0 1
to Step 2. 0.1 0 −0.55 0 1
   
0.1 −0.12 0 0.1
   
B11 =  0  C2 =  0 0 0 
0 0 0 0.1
     
0 1 1 1 0
     
B12 =  0.12  D21 =  0 1 B22 =  0 1 
0 1 0 0 1
     
0 0.15 1 1
     
D12 =  0  D11 = 0  D22 =  0 1 
0.22 0 0 1
and γ = 0.8. The membership function is assumed to be
#
1, f1 (x0 ) = 0
h1 (f1 (xk )) = (34)
| sin(f1 (x0 ))|/f1 (x0 ), else
h2 (f1 (xk )) = 1 − h1 (f1 (xk )). (35)
Earlier, the nonlinear function f (xk ) = sin(xk ) satisfies (2). Fig. 1. State evolution xk of uncontrolled systems.
Our aim is to design a state-feedback paralleled controller in
the form of (4) such that the system (33) is stochastically stable
with a guaranteed H∞ norm bound γ.
Let ρ̄ = 0.8. By applying Theorem 2 with help from Algo-
rithm HinfFC, we can obtain admissible solutions as follows:
 
3.0596 −0.4887 0.0629
 
L =  −0.4887 0.7434 0.3136 
0.0629 0.3136 0.8091
 
0.3874 0.2799 −0.1065
 
P =  0.2799 2.0222 −0.9528 
−0.1065 −0.9528 1.8081

−0.4004 −0.0364 −0.1353
K1 =
−0.0157 −0.1077 0.0209

−0.1291 0.1054 −0.2146
K2 = .
0.0513 −0.0160 0.0166
Fig. 2. State evolution xk of controlled systems.
First, we assume the initial condition to be
x0 = [ 0 0.01 0 ]T (36) z2 = 0.1669 and w2 = 0.9381, and therefore, γ = 0.4218,
which stays below the prescribed upper bound γ ∗ = 0.8.
and the external disturbance wk ≡ 0. Fig. 1 gives the state Example 2: In this example, we aim to show the advantage
responses for the uncontrolled fuzzy systems, which are ap- of considering the probabilistic packet losses. We demonstrate
parently unstable. Fig. 2 gives the state simulation results of through numerical simulation that, when the measurement trans-
the closed-loop fuzzy system, from which we can see that the mission between the plant and controller is indeed imperfect, the
closed-loop system is stochastically stable. design method proposed in this paper gives better performance
Next, to illustrate the disturbance attenuation performance, than the standard H∞ approach without taking into account the
we choose the initial condition x0 ≡ 0 and the external distur- packet loss problem. For this purpose, we let
bance wk as follows:  
 1.0 0.31 0
 0.2, 20 ≤ k ≤ 30 A1 =  0 0.33 0.21 
wk = −0.2, 40 ≤ k ≤ 50 (37) 0 0 −1.5

0, else
and the other system data of (7) be the same as those in
Fig. 3 shows the controller output, and Fig. 4 shows the evolu- Example 1 with ρ̄ = 0.6. First, we assume the initial condi-
tion of the state variables. The disturbance input wk and con- tion to be x0 = [ 0 0.01 0.2 ]T , and the external disturbance
trolled output zk are depicted in Fig. 5. It can be calculated that wk ≡ 0.
Fig. 3. Controllers u k .
Fig. 6. Closed-loop state variables by taking the standard H ∞ performance
criterion.
Fig. 4. State evolution xk of controlled systems when w(k)

= 0.
Fig. 7. Closed-loop state variables by taking the Theorem 2.
In such a case, if we use the standard H∞ performance crite-

rion described in (19) and (20), the controller gain matrices can
be obtained as follows:

−0.3231 0.0460 −0.7181
K1 =
−0.0076 −0.1092 0.6953

−0.3231 0.0597 −0.2481
K2 =
0.0129 −0.0018 0.2066
and the evolution of the corresponding state variables is given
in Fig. 6. If we use Theorem 2, which accounts for the packet
losses, the controller gain matrices are obtained as

−1.7265 0.0129 −1.7443
K1 =
0.3543 −0.1478 1.5592

−0.8813 0.4437 −0.1694
K2 =
0.1682 −0.1474 0.9820
Fig. 5. Controlled output zk and disturbance input w k . with the evolution of the state variables depicted in Fig. 7.
VI. CONCLUSION
In this paper, we have investigated the H∞ fuzzy control prob-
lem for systems with repeated scalar nonlinearities and random
packet losses. The nonlinear system is described by a discrete-
time state equation containing a repeated scalar nonlinearity
and the control strategy takes the form of parallel distributed
compensation. The missing measurements are modeled by a
stochastic variable satisfying the Bernoulli random binary dis-
tribution. The quadratic Lyapunov function has been used to
design H∞ fuzzy controllers such that, for the admissible ran-
dom measurement missing and repeated scalar nonlinearities,
the closed-loop T–S fuzzy control system is stochastically sta-
ble and preserves a guaranteed H∞ performance. By using the
CCL algorithm, sufficient conditions have been established that
ensure the stochastic stability of the closed-loop system, and the
controller gains have been obtained by the solution of a set of
LMIs. Two illustrative simulation examples have been given to
Fig. 8. Closed-loop state variables by taking the standard H ∞ performance
criterion when w(k)
= 0. illustrate the effectiveness of the proposed design method.
We would like to point out that our main results can be
extended to more general/practical systems such as Itô-type
stochastic systems, Markovian jumping systems and time-delay
systems, and the corresponding results will appear in the near
future.
REFERENCES
[1] W. Assawinchaichote, S. K. Nguang, P. Shi, and E. Boukas, “H ∞ fuzzy
state-feedback control design for nonlinear systems with D-stability con-
straints: An LMI approach,” Math. Comput. Simul., vol. 78, pp. 514–531,
2008.
[2] Y. Y. Cao and P. M. Frank, “Analysis and synthesis of nonlinear time-
delay systems via fuzzy control approach,” IEEE Trans. Fuzzy Syst.,
vol. 8, no. 2, pp. 200–211, Apr. 2000.
[3] Y. Cao and P. M. Frank, “Stability analysis and synthesis of nonlinear
time-delay systems via linear Takagi–Sugeno fuzzy models,” Fuzzy Sets
Syst., vol. 124, pp. 213–229, 2001.
[4] B. Chen, X. Liu, and S. Tong, “Robust fuzzy control of nonlinear systems
with input delay,” Chaos Solitons Fractals, vol. 37, pp. 894–901, 2006.
[5] Y. Chu, “Further results for systems with repeated scalar nonlinearities,”
IEEE Trans. Autom. Control, vol. 44, no. 12, pp. 2031–2035, Dec. 2001.
[6] Y. Chu and K. Glover, “Bounds of the induced norm and model reduc-
tion errors for systems with repeated scalar nonlinearities,” IEEE Trans.
Fig. 9. Closed-loop state variables by taking the Theorem 2 when w(k)
= 0. Autom. Control, vol. 44, no. 3, pp. 471–483, Mar. 1999.
[7] Y. Chu and K. Glover, “Stabilization and performance synthesis for sys-
tems with repeated scalar nonlinearities,” IEEE Trans. Autom. Control,
vol. 44, no. 3, pp. 484–496, Mar. 1999.
[8] G. Feng, “Controller synthesis of fuzzy dynamic systems based on piece-
wise Lyapunov function,” IEEE Trans. Fuzzy Syst., vol. 11, no. 5, pp. 605–
Furthermore, to illustrate the disturbance attenuation perfor- 612, Oct. 2003.
mance, we set the initial condition x0 ≡ 0 and the external [9] G. Feng and J. Ma, “Quadratic stabilization of uncertain discrete-time
disturbance wk be the same as that in (37). Again, we apply the fuzzy dynamic systems,” IEEE Trans. Circuits Syst.-I: Fundam. Theory
Appl., vol. 48, no. 11, pp. 1337–1344, Nov. 2001.
standard H∞ performance criterion as described in (19) and [10] G. Feng, “Stability analysis of discrete-time fuzzy dynamic systems based
(20) and Theorem 2, and display the evolutions of the state on piecewise Lyapunov functions,” IEEE Trans. Fuzzy Syst., vol. 12, no. 1,
variables under these two situations in Figs. 8 and 9, respec- pp. 22–28, Feb. 2004.
[11] H. Gao, J. Lam, and C. Wang, “Induced l2 and generalized H 2 filtering for
tively. The disturbance attenuation performance indexes can systems with repeated scalar nonlinearities,” IEEE Trans. Signal Process.,
also be calculated as γ = 0.4670 and γ = 0.4445, respectively. vol. 53, no. 11, pp. 4215–4226, Nov. 2005.
By comparing Fig. 6 with Fig. 7 and noticing the H∞ perfor- [12] H. Gao, Z. Wang, and C. Wang, “Improved H ∞ control of discrete-time
fuzzy systems: A cone complementarity linearization approach,” Inf. Sci.,
mance indexes, we can conclude that the controller design by vol. 175, no. 1–2, pp. 57–77, 2005.
Theorem 2 has given better dynamical behavior as well as a [13] H. Gao, Y. Zhao, J. Lam, and K. Chen, “H ∞ fuzzy filtering of nonlinear
better disturbance rejection attenuation level, which confirms systems with intermittent measurements,” IEEE Trans. Fuzzy Syst., to be
published.
our theoretical analysis for the problem of H∞ fuzzy control for [14] L. El, Ghaoui, F. Oustry, and M. A. Rami, “A cone complementarity
systems with repeated scalar nonlinearities and random packet linearization algorithm for static output-feedback and related problems,”
losses. IEEE Trans. Autom. Control, vol. 42, no. 8, pp. 1171–1176, Aug. 1997.
[15] Y.-Y. Hou, T.-L. Liao, and J.-J. Yan, “Stability analysis of Takagi–Sugeno Zidong Wang (M’03–SM’03) was born in Jiangsu,
fuzzy cellular neural networks with time-varying delays,” IEEE Trans. China, in 1966. He received the B.Sc. degree in math-
Syst., Man, Cybern.: B, Cybern, vol. 37, no. 3, pp. 720–726, Jun. 2007. ematics from Suzhou University, Suzhou, China, in
[16] H. Huang, D. W. C. Ho, and J. Lam, “Stochastic stability analysis of fuzzy 1986, and the M.Sc. degree in applied mathematics
hopfield neural networks with time-varying delays,” IEEE Trans. Circuits and the Ph.D. degree in electrical and computer en-
Syst. II: Exp. Briefs, vol. 52, no. 5, pp. 251–255, May 2005. gineering from Nanjing University of Science and
[17] O. C. Imer, S. Yuksel, and T. Basar, “Optimal control of LTI systems over Technology, Nanjing, China, in 1990 and 1994, re-
unreliable communication links,” Automatica, vol. 42, no. 9, pp. 1429– spectively.
1439, 2006. He is currently a Professor of dynamical systems
[18] B. Jiang, M. Staroswiecki, and V. Cocquempot, “Fault accommodation and computing, Brunel University, Uxbridge, Mid-
for nonlinear dynamic systems,” IEEE Trans. Autom. Control, vol. 51, dlesex, U.K. His current research interests include
no. 9, pp. 1578–1583, Sep. 2006. dynamical systems, signal processing, bioinformatics, and control theory and
[19] X. Liu, “Delay-dependent H ∞ control for uncertain fuzzy systems with applications. He has published more than 120 papers in refereed international
time-varying delays,” J. Comput. Appl. Math., vol. 68, no. 5, pp. 1352– journals.
1361, 2008. Dr. Wang is currently serving as an Associate Editor for 12 international jour-
[20] Z. H. Mao and B. Jiang, “Fault identification and fault-tolerant control for nals including IEEE TRANSACTIONS ON AUTOMATIC CONTROL, IEEE TRANSAC-
a class of networked control systems,” Int. J. Innovative Comput., Inf. TIONS ON NEURAL NETWORKS, IEEE TRANSACTIONS ON SIGNAL PROCESSING,
Control, vol. 3, no. 5, pp. 1121–1130, 2007. IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C, and
[21] S. K. Nguang and P. Shi, “Fuzzy H ∞ output feedback control of nonlinear IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY.
systems under sampled measurements,” Automatica, vol. 39, pp. 2169–
2174, 2003.
[22] S. K. Nguang and P. Shi, “H ∞ fuzzy output feedback control design for Huijun Gao (M’06) was born in Heilongjiang
nonlinear systems: An LMI approach,” IEEE Trans. Fuzzy Syst., vol. 11, Province, China, in 1976. He received the M.S. de-
no. 3, pp. 331–340, Jun. 2003. gree in electrical engineering from Shenyang Univer-
[23] B. Shen, Z. Wang, H. Shu, and G. Wei, “On nonlinear H ∞ filtering sity of Technology, Shengyang, China, in 2001, and
for discrete-time stochastic systems with missing measurements,” IEEE the Ph.D. degree in control science and engineering
Trans. Autom. Control, vol. 53, no. 9, pp. 2170–2180, Oct. 2008. from Harbin Institute of Technology, Harbin, China,
[24] P. Shi and S. K. Nguang, “H ∞ output feedback control of fuzzy system in 2005.
models under sampled measurements,” Comput. Math. Appl., vol. 46, He is currently a Professor at Harbin Institute
no. 5–6, pp. 705–717, 2003. of Technology. His current research interests include
[25] K. Tanaka, T. Hori, and H. O. Wang, “A multiple Lyapunov function network-based control, robust control/filter theory,
approach to stabilization of fuzzy control systems,” IEEE Trans. Fuzzy model reduction, time-delay systems, and multidi-
Syst., vol. 11, no. 4, pp. 582–589, Aug. 2003. mensional systems, and their applications. He has published more than 80
[26] Z. Wang, D. W. C. Ho, and X. Liu, “Variance-constrained filtering for papers in refereed international journals.
uncertain stochastic systems with missing measurements,” IEEE Trans. Dr. Gao is an Associate Editor or Member of the editorial board for
Autom. Control, vol. 48, no. 7, pp. 1254–1258, Jul. 2003. several journals, such as IEEE TRANSACTIONS ON SYSTEMS, MAN, AND
[27] Z. Wang, D. W. C. Ho, and X. Liu, “Variance-constrained control for CYBERNETICS—PART B, International Journal of Systems Science, Journal of
uncertain stochastic systems with missing measurements,” IEEE Trans. Intelligent and Robotics Systems, and Journal of the Franklin Institute.
Syst., Man Cybern. A, Syst., Humans, vol. 35, no. 5, pp. 746–753, Sep.
2005.
[28] Z. Wang, F. Yang, D. W. C. Ho, and X. Liu, “Robust H ∞ control for
networked systems with random packet losses,” IEEE Trans. Syst., Man
Cybern. B, Cybern., vol. 37, no. 4, pp. 916–924, Aug. 2007.
[29] H. Wu, “Delay-dependent H ∞ fuzzy observer-based control for discrete-
time nonlinear systems with state delay,” Fuzzy Sets Syst, vol. 159,
pp. 2696–2712, 2008.
[30] K. Yuan, J. Cao, and J. Deng, “Exponential stability and periodic solutions
of fuzzy cellular neural networks with time-varying delays,” Neurocom-
puting, vol. 69, no. 13–15, pp. 1619–1627, 2006.
Hongli Dong received the B.Sc. degree in computer

science and technology from Heilongjiang Institute
of Science and Technology, Harbin, China, in 2000,
and the M.Sc. degree in control theory and engineer-
ing from Daqing Petroleum Institute, Daqing, China,
in 2003. She is currently working toward the Ph.D.
degree in control science and engineering from the
Harbin Institute of Technology, Harbin.
She is currently a Lecturer at Daqing Petroleum
Institute. Her current research interests include robust
control and networked control systems. She is an ac-
tive reviewer for many international journals.
The Model of Fuzzy Variable Precision Rough Sets

Suyun Zhao, Student Member, IEEE, Eric C. C. Tsang, Member, IEEE, and Degang Chen
Abstract—The fuzzy rough set (FRS) model has been introduced perturbation represents small change of real values. In recent
to handle databases with real values. However, FRS was sensitive years, a great number of researchers have paid attention to the
to misclassification and perturbation (here misclassification means development of RS. Many extensions of RS, such as general-
error or missing values in classification, and perturbation means
small changes of numerical data). The variable precision rough sets ization of approximation space, concepts approximation, rough
(VPRSs) model was introduced to handle databases with misclas- fuzzy sets, fuzzy rough sets (FRSs) and variable precision rough
sification. However, it could not effectively handle the real-valued sets (VPRSs), covering RSs, have been proposed and studied.
datasets. Now, it is valuable from theoretical and practical aspects Interested readers may consult [3] for summary of some of these
to combine FRS and VPRS so that a powerful tool, which not only extensions. The focus of this paper is on two types of general-
can handle numerical data but also is less sensitive to misclassifi-
cation and perturbation, can be developed. In this paper, we set izations: FRSs and VPRSs.
up a model named fuzzy VPRSs (FVPRSs) by combining FRS and Since theories of fuzzy sets and RSs were distinct and com-
VPRS with the goal of making FRS a special case. First, we study plementary on dealing with uncertainty, many researchers were
the knowledge representation ways of FRS and VPRS, and then, inspired to construct some new models, in which both indiscerni-
propose the set approximation operators of FVPRS. Second, we bility and fuzziness were considered. The results of these stud-
employ the discernibility matrix approach to investigate the struc-
ture of attribute reductions in FVPRS and develop an algorithm to ies led to the construction of FRSs. Recently, most researchers
find all reductions. Third, in order to overcome the NP-complete focused on the study of knowledge representation of FRS [5]–
problem of finding all reductions, we develop some fast heuristic [15]. These studies mainly have been done from two branches:
algorithms to obtain one near-optimal attribute reduction. Finally, the constructive and axiomatic. The axiomatic branch took the
we compare FVPRS with RS, FRS, and several flexible RS-based lower and upper approximation operators as the primary notion
approaches with respect to misclassification and perturbation. The
experimental comparisons show the feasibility and effectiveness of and focused on studying the mathematical structure of FRS,
FVPRS. such as algebraic and topologic structures [10]–[15]. On the
contrary, the constructive branch focused on studying the con-
Index Terms—Attribute reduction, fuzzy sets, knowledge repre-
sentation, rough sets (RSs). struction of the lower and upper approximation operators and
was often used to develop the applications of FRS [5]–[9].
Attribute reduction, as one important application of FRS, has
I. INTRODUCTION been studied in [16]–[23]. Shen and Jensen first proposed the
OUGH set (RS) theory, as one kind of generalization of attribute reduction method with FRS by keeping dependency
R set theory, was proposed to deal with uncertainty and in-
discernibility [1], [2]. It assumed that objects characterized by
function invariant [18]–[21]. Their work performed well on
some practical applications, but their algorithms lacked math-
the same information were indiscernible (similar) in the view of ematical foundation and theoretical analysis. Unlike [18]–[21],
the available information about them [3]. The indiscernibility an attribute reduction method by adopting information entropy
relation generated in this way was the mathematical basis of to measure attribute significance was proposed in [22]. How-
RS. It made RS work well on some problems, but it also limited ever, the notions of attribute reduction in [22] were not based
the further applications of RS. For example, RS could not work on knowledge representation of FRS. That is to say, the knowl-
effectively on the datasets with real values. Furthermore, RS edge representation part and the attribute reduction part were
was sensitive to misclassification and perturbation. Here, mis- independent. By considering the limitations of the aforemen-
classification represents erroneous or missing classification and tioned attribute reduction methods, the authors in [16] proposed
a unified framework of attribute reduction with FRS that not
Manuscript received October 4, 2008; accepted December 30, 2008. Cur- only proposed a formal notion of attribute reduction based on
rent version published April 1, 2009. This work was supported by the Hong the lower approximation operators, but also analyzed the math-
Kong Research Grants Council (RGC) Competitive Earmarked Research Grant ematical structure of attribute reduction. Although this method
(CERG) under Grant PolyU 5273/05E (B-Q943) and Grant PolyU 5281/07E
(B-Q06C). The work of D. Chen was supported by the National Science Finance proposed in [16] worked well on some datasets with fuzzified
of China (NSFC) under Grant 70871036. values, by experiments we found that their attribute reduction
S. Zhao is with the Department of Computing, Hong Kong Polytechnic results on the real-valued datasets were not compact because
University, Kowloon, Hong Kong, and also with the School of Mathematics
and Computer Sciences, Hebei University, Baoding 071002, China (e-mail: the lower approximation operators in FRS could not effectively
zsyrose2007@yahoo.com.cn). deal with the problem of misclassification and were sensitive to
E. C. C. Tsang is with the Department of Computing, Hong Kong Polytechnic perturbation of the original data [17], [50]. Now, it is necessary
University, Kowloon, Hong Kong (e-mail: etsang@mlcri.hk).
D. Chen is with the Department of Mathematics and Physics, and important to improve FRS to a robust model with respect to
North China Electric Power University, Beijing 102206, China (e-mail: misclassification and perturbation.
chengdegang@263.net). As a robust model, VPRS was proposed to handle the prob-
at http://ieeexplore.ieee.org. lems with misclassification by incorporating a controlled de-
Digital Object Identifier 10.1109/TFUZZ.2009.2013204 gree of misclassification into knowledge representation of RS
1063-6706/$25.00 © 2009 IEEE
formalism [24]. Recently, many researchers have generalized II. PRELIMINARIES

VPRS to some more general application areas [25]–[27]. Some In this section, we review RS, FRS, and several flexible RS-
of them have been focused on the parameter, i.e., the vari- based methods with respect to misclassification and/or pertur-
able precision. For example, one generalization of VPRS with
bation, which are necessary preliminaries for constructing the
asymmetric bounds was proposed that allowed the lower and model of FVPRSs.
upper approximation operators with different variable preci-
sions [25]. More researches have been put forward on the
investigation of attribute reduction with VPRS. For example, A. Rough Sets
by considering the unreasonable definition of attribute reduc- RS theory proposed by Pawlak [1] was an extension of set
tion, some new concepts named beta lower and upper dis- theory for the study of intelligent systems characterized by
tribution attribute reductions were then proposed in [26]. Al- insufficient and incomplete information. It is formally based
though VPRS and its existing generalizations worked well on on an information system that is a pair IS = (U, A), where
the problems with misclassification, they could not effectively U is a nonempty finite set of objects called the universe and
deal with the numerical problems with misclassification and/or A is a nonempty finite set of attributes. With every subset
perturbation. of attributes B ⊆ A, we associate a binary relation defined
By considering the characteristics of FRS and VPRS, some as IND(B) = {(x, y) ∈ U : a(x) = a(y), ∀a ∈ B}. The sets
researchers were inspired to construct a model, which can handle {x ∈ U : [x]B ⊆ X} and {x ∈ U : [x]B ∩ X = φ} are called
the numerical problems with misclassification and/or perturba- B-lower and B-upper approximations of X ⊆ U , and are de-
tion, by combining FRS and VPRS. Thus, several flexible RS- noted by BX and BX, respectively. By using these approxima-
based generalizations were proposed. For example, VPRS was tions, knowledge hidden in data can be represented as possible
introduced into a specific FRS framework (VPRS-FRS) in [28]. and certain rules.
It replaced the normal aggregation by β-precision aggregation If the set of attributes in the information system consists of
to weaken the significant influence of some misclassification two parts: a set of condition attributes R and a set of decision
samples on the values of lower and upper approximation oper- attributes D, the information system is called a decision system,
ators. Furthermore, VPRS was generalized into a general FRS or decision table, denoted by DS = (U, R ∪ D). One impor-
framework (VPFRS) [37]–[39]. In this generalization, misclas- tant application of RS was attribute reduction that distinguished
sification was controlled by a crisp cut set that was introduced RS itself from other approaches since no additional information
to the lower and upper approximation operators of a general such as expert knowledge and/or data distribution was needed
FRS framework. VPRS-FRS and VPFRS were effective to han- in the process of attribute reduction. However, RS could not
dle misclassification. However, they could not effectively deal effectively handle nonsymbolic valued databases with misclas-
with the problems with perturbation. Recently, a robust model sification and perturbation. Thus, several generalizations of tra-
with respect to both misclassification and perturbation, vaguely ditional RSs were considered. Among these generalizations, we
quantified rough sets (VQRSs), has been proposed [40]. It in- emphasize FRS and VPRS in this paper.
troduced a fuzzy quantification measure, i.e., quantifier, into the
set approximation operators. This quantifier made VQRS work
B. Fuzzy Rough Sets
well on deducing the influence of misclassification and pertur-
bation, but it also limited the further applications of VQRS. 1) Fuzzy Logical Operator: To provide a clear introduction
For example, it was difficult for VQRS to design an attribute of FRS in the constructive branch, we present and exemplify
reduction method since some important properties such as the some fuzzy logical operators used in the construction of the set
approximation quality being a monotonic measure did not hold approximation operators in FRS [6], [10], [11], [15]: triangular
in VQRS. norm, triangular conorm, negator, dual, T -residuated implica-
Now, it is valuable from both theoretical and practical aspects tion, and its dual operation.
to set up a framework by combining FRS and VPRS. From the A triangular norm, or shortly T -norm, is a function T :
theoretical aspect, we can improve the knowledge representation [0, 1] × [0, 1] → [0, 1] that satisfies the following conditions:
power of FRS and set up a robust framework so that a further monotonicity [if x < α, y < β, then T (x, y) ≤ T (α, β)], com-
generalization of RSs is gained. From the practical aspect, we mutativity [T (x, y) = T (y, x)], associativity [T (T (x, y), z) =
can propose some new attribute reduction methods that are less T (x, T (y, z))], and boundary condition [T (x, 1) = x]. The
sensitive to both misclassification and perturbation in a fuzzy most popular continuous T -norms include the standard min
information system. operator TM (x, y) = min{x, y} and the bounded intersection
The rest of this paper is organized as follows. Section II re- (also called the Lukasiewicz T -norm) TL (x, y) = max{0, x +
views RS, FRS, and several RS-based flexible methods with re- y − 1}.
spect to misclassification and/or perturbation. Section III builds A triangular conorm, or shortly T -conorm, is an in-
a framework named fuzzy VPRSs (FVPRSs) and develops some creasing, commutative, and associative function S : [0, 1] ×
algorithms to find the reductions by employing the discernibility [0, 1] → [0, 1] that satisfies the boundary condition: ∀x ∈
matrix approach. In Section IV, we by experiments demonstrate [0, 1], S(x, 0) = x. The most well-known continuous. T -
the feasibility and effectiveness of FVPRS. The last section conorms include the standard max operator SM (x, y) =
concludes this paper. max{x, y} and the bounded sum SL (x, y) = min{1, x + y}.
ZHAO et al.: MODEL OF FUZZY VARIABLE PRECISION ROUGH SETS 453
A negator N is a decreasing function N : [0, 1] → [0, 1]

that satisfies N (0) = 1 and N (1) = 0. A negator N is called
involutive iff N (N (x)) = x for all x ∈ [0, 1]. The standard
negator is defined as NS (x) = 1 − x. Given a negator N ,
T -norm T and T -conorm S are called dual with respect to
(shortly denoted by w.r.t.) N iff De Morgan laws are satis-
fied, i.e., S(N (x), N (y)) = N (T (x, y)) and T (N (x), N (y)) =
N (S(x, y)).
Let X : U → [0, 1] be a fuzzy set and F (U ) be the fuzzy
power set on U , i.e., the collection of all fuzzy sets on U . For
every X ∈ F (U ), the symbol coN X is used to denote the fuzzy
complement of X determined by a negator N , i.e., for every
x ∈ U , (coN X)(x) = N (X(x)).
Given a lower semicontinuous triangular norm T , the resid-
uation implication, or called the T -residuated implication is
a function ϑ : [0, 1] × [0, 1] → [0, 1] that satisfies ϑ(x, y) =
sup{z |z ∈ [0, 1], T (x, z) ≤ y } for every x, y ∈ [0, 1]. T -
residuated implications include the Godel implication ϑM based
on

1, x ≤ y
TM : ϑM =
y, x > y
and the Lukasiewicz implication ϑL based on TL : ϑL =

min{1 − x + y, 1}.
Given an upper semicontinuous triangular conorm S,
the dual of T -residuated implication w.r.t. N is a
function σ : [0, 1] × [0, 1] → [0, 1] that satisfies σ(x, y) =
inf{z |z ∈ [0, 1], S(x, z) ≥ y } for every x, y ∈ [0, 1]. Note
that σ(N (x), N (y)) = N (ϑ(x, y)) and ϑ(N (x), N (y)) =
N (σ(x, y)) hold for an involutive negator N [15]. Fig. 1. Fuzzy sets. (a) f1 . (b) f2 . (c) f3 .
2) Fuzzy Rough Set: FRS was first proposed by Dubois and
Prade [5], [6] and then studied in [8], [10]–[12], and [15]. Inter-
ested readers can consult [15] for more detailed summary of the From the viewpoint of duality, RT A and RS A are one pair of
development of FRS. Here, we present the general definition of dual approximation operators, and Rσ A and Rϑ A are another
the fuzzy approximation operators. pair of dual approximation operators [15].
Suppose U is a nonempty universe (may not be finite) and R From the viewpoint of granular computing, these four approx-
is a binary fuzzy relation on U , then the fuzzy approximation imation operators can be represented by fuzzy granules RT xλ
operators can be summarized as follows. For every fuzzy set and Rσ xλ [14]. They are
A ∈ F (U )
Rϑ A = ∪{RT xλ : RT xλ ⊆ A}; RT A = ∪{RT xλ : xλ ⊆ A}
RT A(x) = sup T (R(x, u), A(u))
u ∈U RS A = ∪{Rσ xλ : Rσ xλ ⊆ A}; Rσ A = ∪ {Rσ xλ : xλ ⊆ A}.
RS A(x) = inf S(N (R(x, u)), A(u)) There are two conditions RT xλ ⊆ A and Rσ xλ ⊆ A in the
u ∈U
aforementioned granular representation of fuzzy approximation
Rσ A(x) = sup σ(N (R(x, u)), A(u)) operators. These conditions are too harsh for the construction of
u ∈U
the fuzzy lower approximation operators Rϑ A and RS A. The
Rϑ A(x) = inf ϑ(R(x, u), A(u)). graphs to illustrate this idea are shown in Figs. 1 and 2.
u ∈U
In Figs. 1 and 2, we use some fuzzy sets to illustrate the
Note that in the applications of FRS [5], [10], [16]–[22], fuzzy granules RT xλ and Rσ xλ , a box to illustrate a crisp set
[28], a binary fuzzy relation R is often specified as a fuzzy A. Fig. 1 is the demonstration graph of three fuzzy sets f1 ,
T -similarity relation, which is a fuzzy binary relation R on U f2 , and f3 that satisfy f1 ≥ f2 ≥ f3 . Fig. 2 shows the inclusion
satisfying the following conditions: reflexivity (R(x, x) = 1), relations between f1 , f2 , f3 , and A, respectively. We find that f3
symmetry [R(x, y) = R(y, x)], and T -transitivity [R(x, y) ≥ is included in A, whereas only part of f1 and f2 are included in
T (R(x, z), R(z, y))], for every x, y, z ∈ U [10]. In this paper, A. From Fig. 2(b), it is easy to see that if we ignore these small
we focus on studying the fuzzy approximation operators based membership degrees, which may be caused by misclassification
on a fuzzy T -similarity relation. and perturbation, f2 should be included in A. All these show
C. Several Flexible RS-Based Methods With Respect

to Misclassification and Perturbation
In this section, we review two types of RS generalizations
with threshold. One type focused on generalizing the knowl-
edge representation power of RS. They are VPRS, VPRS-FRS,
VPFRS, and VQRS [24], [28], [37], [40]. The other type fo-
cused on studying the criteria of attribute reduction, such as the
methods of fuzzy reductions (Fuzzy-RED) [49] and approxi-
mate reductions (Approximate-RED) [46]–[48].
1) VPRS: VPRS has been proposed to handle the problems
with misclassification by incorporating a controlled de-
gree of misclassification into knowledge representation of
RS formalism. It was a robust model and has been widely
used in practical applications. However, VPRS could not
effectively deal with real-valued databases with misclas-
sification and perturbation.
2) VPRS-FRS: VPRS-FRS introduced VPRS into Dubois
and Prade’s FRS framework. By considering that one mis-
classification sample may significantly influence the value
of fuzzy approximation operators, the authors in [28] in-
troduced a concept named β-precision aggregation to the
lower and upper approximation operators. Their approxi-
mation operators were proposed as follows:
R∗ Aβ (x) = max(IR ∗ A β , A(x))
R∗ Aβ (x) = min(IR ∗ A β , A(x))
with
IR ∗ A β (x) = maxβ min(R(x, y), A(y))
y (y = x)
and
IR ∗ A β (x) = minβ max(1 − R(x, y), A(y)).
y (y = x)
By employing β-precision aggregation, we find that some

maximum or minimum values of min(R(x, y), A(y)) or
max(1 − R(x, y), A(y)), which may be caused by mis-
classification, were omitted by controlling the threshold β.
As a result, VPRS-FRS was robust to handle real-valued
datasets with misclassification. However, it was sensitive
to perturbation since perturbation may exist in each value
of min(R(x, y), A(y)) or max(1 − R(x, y), A(y)).
3) VPFRS: The authors in [37] introduced VPRS in a general
FRS framework. Here, we focus on reviewing the lower
Fig. 2. Inclusion relations between f1 , f2 , f3 , and A. (a) f1 and A. (b) f2 approximation of VPFRS since it was important for at-
and A. (c) f3 and A. tribute reduction. The lower approximation was proposed
as follows:
that the conditions of RT xλ ⊆ A and Rσ xλ ⊆ A are too harsh µR u F (Xi )
for the construction of the lower approximation operators and 
limit the boundary of the positive region. Thus, the knowledge  x∈Sinf ϑ(µX i (x), µF (x)),
iu
representation power of FRS is relatively weak. = if ∃αu = sup{α ∈ (0, 1] : eα (Xi , F ) ≤ 1−u}

By the aforementioned analysis, we find that the approxima- 0, otherwise
tion operator of FRS is sensitive to misclassification and per-
turbation. Some small misclassification and perturbation may where Si u = sup p(Xi ∩ XiFα u ).
affect the values of the positive region and further affect the This definition can be equivalently represented as
performance of attribute reduction with FRS. shown as the bottom of the page.
The proof of the equivalence of two statements is presented reduction. Here, we focus on reviewing the key idea of attribute
as follows. reduction.
Proof: 1) Approximate-RED: The concept of approximate reduction
was proposed by keeping the quality measure of certain
Si u = sup p(Xi ∩ XiFα u ) = sup p(Xi ) ∩ sup p(XiFα u ) subset almost unchanged after attribute reduction. The
= sup p(Xi ) ∩ Xi α u F approximate reduction proposed in this way was more
flexible than VPRS and FRS because it was not only free
(since Aα = {x ∈ X, µA (x) ≥ α}) from the unreasonable behavior of reduction criteria in
= sup p(Xi ) ∩ {x ∈ X : ϑ(µX i (x), µF (x))≥α} VPRS, but also free from the harsh criteria of attribute
reduction in FRS.
= {x ∈ X : ϑ(µX i (x), µF (x)) ≥ α and µX i (x) > 0}. 2) Fuzzy-RED: In Fuzzy-RED, after describing a monotonic
measure, the authors in [49] proposed the fuzzy decision
We have
reduction by keeping this measure almost unchanged.
inf ϑ(µX i (x), µF (x)) Fuzzy-RED was more general than Approximate-RED
x∈S i u since the quality measure was a special case of the mono-
= inf ϑ(µX i (x), µF (x)). tonic measure.
ϑ(µ X i (x),µ F (x))≥α u Up until now, we have reviewed two types of flexible
RS-based generalizations. The former focused on studying the
From the equivalent definition of VPFRS, we find that the
lower approximation, while the latter focused on relaxing the
threshold in VPFRS was put in the position of aggrega-
criteria of attribute reduction. The main differences between
tion operator. As a result, some misclassification could be
them are listed as follows: 1) in the former type, the threshold
omitted by controlling this threshold. However, VPFRS
was put in set approximation operators, whereas in the latter,
was still sensitive to perturbation since the introduction of
it was put in attribute reduction and 2) in the former type, mis-
the threshold did not reduce the effect of perturbation that
classification and/or perturbation were controlled in knowledge
may exist in each value of ϑ(µX i i (x), µF (x)).
representation, whereas in the latter, they were controlled in
VPRS-FRS and VPFRS shared some strengths: 1) they
attribute reduction.
took FRS as a special case and 2) they were effective to
Note that in this paper, we employ the idea of keeping the
handle misclassification. Also, they shared some limita-
quality measure invariant in attribute reduction since misclas-
tions: 1) they were sensitive to perturbation and 2) attribute
sification and perturbation have been omitted in knowledge
reduction was not mentioned in these models.
representation.
4) VQRS: Another flexible generalization is VQRS. In this
model, a fuzzy quantification measure, named quantifier,
was introduced to measure whether one fuzzy set Ry was
III. FUZZY VARIABLE PRECISION ROUGH SETS MODEL
most included in another fuzzy set A: Qu (|Ry ∩ A|/|A|).
Then, the lower approximation was proposed as In this section, we improve the knowledge representation
follows: power of FRS by incorporating one controlled degree into the
fuzzy approximation operators so that the undesirable effect
|Ry ∩ A|
R ↓Q u A(y) = Qu . of misclassification and perturbation can be weakened. First,
|A| we propose four fuzzy approximation operators with variable
VQRS had the following strengths: 1) it was a flexible precision, and then discuss the properties of them. Second,
model since it was less sensitive to misclassification and we propose the concepts of attribute reduction based on the
perturbation; 2) it was suitable for both symbolic problems proposed approximation operators. Also, we investigate the
and real-valued problems; and 3) it took RS and VPRS as structure of these reductions by employing the discernibility
the special cases. Also, VQRS had some limitations. One matrix approach. Finally, we design some algorithms to find
main limitation was that some important properties about the reductions. Thus, we build a unified framework named
attribute reduction did not hold. For example, the approx- FVPRSs.
imation quality measure did not monotonically increase
with the number of selected attributes. As a result, the al-
gorithm of attribute reduction by keeping approximation A. Set Approximation in FVPRS and Their Properties
quality invariant could not be designed in VQRS. In this section, we first describe some new concepts: fuzzy
In the following, we review another type of RS generaliza- lower and upper approximation operators with variable preci-
tions that introduced the threshold in the concept of attribute sion and fuzzy positive region with variable precision. Then,


 inf ϑ(µX i (x), µF (x)), if ∃αu = sup{α ∈ (0, 1] : eα (Xi , F ) ≤ 1 − u}
ϑ(µ X i (x),µ F (x))≥α u
µR u F (Xi ) =


µ X i (x)> 0
0, otherwise.
we present and prove some properties and theorems that are ∨ sup N (S(N (R(x, u)), N (A(u))))
valuable for the following construction of discernibility matrix. N (A (u ))> α
Definition 3.1: Given F D = (U, R ∪ D) and A ∈ F (U ), = sup T (R(x, u), N (α))

the fuzzy lower and upper approximation operators with vari- A (u )≥N (α )
able precision α ∈ [0, 1), also called the lower and upper
∨ sup T (R(x, u), A(u)) = RT α A.
approximation operators of FVPRS, are, respectively, defined A (u )< N (α )
as follows. ∀x ∈ U :
D3.1.1) Rϑ α A(x) = inf ϑ(R(x, u), α) The proof of P3.1.2) is similar to P3.1.1) since ϑ and σ are
A (u )≤α
∧ inf ϑ(R(x, u), A(u)); dual w.r.t. the involutive negator N .
A (u )> α The previous proposition shows that Rϑ α A and Rσ α A,
D3.1.2) RT α A(x) = sup T (R(x, u), N (α)) RS α A and RT α A are dual w.r.t. the involutive negator N ,
A (u )≥N (α )
respectively. In the following, we focus on the pair of fuzzy
∨ sup T (R(x, u), A(u));
A (u )< N (α ) approximation operators Rϑ α and RT α . Some properties of an-
D3.1.3) RS α A(x) = inf S(N (R(x, u)), α) other pair of fuzzy approximation operators of RS α A and RT α A
A (u )≤α
can be obtained in a similar way by using the duality.
∧ inf S(N (R(x, u)), A(u));
A (u )> α If A is an arbitrary crisp subset of U , the lower and upper ap-
D3.1.4) Rσ α A(x) = sup σ(N (R(x, u)), N (α)) proximation operators of FVPRS degenerate into the following
A (u )≥N (α ) formulae. ∀x ∈ U :
∨ sup σ(N (R(x, u)), A(u)). D3.1.5) Rϑ α A(x) = inf ϑ(R(x, u), α);
A (u )< N (α ) A (u )=0
Since T , S, ϑ, and σ are dual, respectively, we have the D3.1.6) RT α A(x) = sup T (R(x, u), N (α));
following properties. A (u )=1
Proposition 3.1: For A ∈ F (U ), the following statements D3.1.7) RS α A(x) = inf S(N (R(x, u)), α);
A (u )=0
hold.
D3.1.8) Rσ α A(x) = sup σ(N (R(x, u)), N (α)).
P3.1.1) RS α A = coN (RT α (coN A)), A (u )=1
RT α A = coN (RS α (coN A)); A fuzzy positive region with variable precision α ∈ [0, 1),
P3.1.2) Rϑ α A = coN (Rσ α (coN A)), also called α-positive region of D relative to R, can be defined
Rϑ α A = coN (Rσ α (coN A)). as a union of fuzzy lower approximations, i.e., P OSR α (D) =
Proof: ∪ Rϑ α [x]D . If no confusion arises, we still call it a posi-
x∈U
P3.1.1) ∀x ∈ U tive region. If k = γR α (D) = |P OSR α (D)|/|U |, we say that
D is α-dependent on R in a degree k, where k is called α-
dependency degree of D on R. Attribute a ∈ P ⊆ R is α-
coN (RT α (coN A))(x)
dispensable in P w.r.t. D if P osP α D = P os(P −{a}) α D, oth-
= N (RT α (coN A))(x) erwise it is α-indispensable in P w.r.t. D. If every attribute in
P is α-indispensable, then we say that P is α-independent in R
=N sup T (R(x, u), N (α)) w.r.t. D.
N (A (u ))≥N (α ) When α = 0, FRS becomes the special case of FVPRS. We

show this fact in the following properties.
∨ sup T (R(x, u), coN A(u)) Proposition 3.2: For α ≤ β and α, β ∈ [0, 1) :
N (A (u ))< N (α )
P3.2.1) Rϑ A = Rϑ 0 A;
= inf N (T (R(x, u), N (α))) P3.2.2) Rϑ A ⊆ Rϑ α A;
A (u )≤α
P3.2.3) Rϑ α A ⊆ Rϑ β A;
∧ inf N (T (R(x, u), N (A(u)))) P3.2.4) RT A = RT 0 A;
A (u )> α
P3.2.5) RT α A ⊆ RT A;
= inf S(N (R(x, u)), α) P3.2.6) RT β A ⊆ RT α A;
A (u )≤α
P3.2.7) P OSR D = P OSR 0 D;
∧ inf S(N (R(x, u)), A(u)) = RS α A P3.2.8) P OSR D ⊆ P OSR α D.
A (u )> α
Proof:
coN (RS α (coN A))(x) = N (RS α (coN A))(x) P3.2.1) ∀x ∈ U :

Rϑ 0 A(x) = inf ϑ(R(x, u), 0) ∧ inf ϑ(R(x, u), A(u))
=N inf S(N (R(x, u)), α) A (u )≤0 A (u )> 0
co N A (u )≤α
= Rϑ A(x).
∧ inf S(N (R(x, u)), coN A(u))
co N A (u )> α P3.2.2) ∀x ∈ U :
= sup N (S(N (R(x, u)), α)) Rϑ A(x) = inf ϑ(R(x, u), A(u))∧ inf ϑ(R(x, u), A(u))
N (A (u ))≤α A (u )≤α A (u )> α
≤ inf ϑ(R(x, u), α) ∧ inf ϑ(R(x, u), A(u))

A (u )≤α A (u )> α
(since ϑ(·, ·) is monotonically increasing

in the right argument)
= Rϑ A(x) ≤ Rϑ α A(x).
Hence, Rϑ A ⊆ Rϑ α A.
P3.2.3) ∀x ∈ U :
Rϑ α A(x) = inf ϑ(R(x, u), α)∧ inf ϑ(R(x, u), A(u)) Fig. 3. Practical meaning of fuzzy granule R T xλ .
A (u )≤α α < A (u )≤β
∧ inf ϑ(R(x, u), A(u)) reduction [16]. From the practical viewpoint, the fuzzy granule
A (u )> β
RT xλ functions as the crisp granule, i.e., the equivalence class.
≤ inf ϑ(R(x, u), β) ∧ inf ϑ(R(x, u), β) In the following, we explain the practical meaning of RT xλ
A (u )≤α α < A (u )≤β
by using a demonstration graph (see Fig. 3). The gray square
∧ inf ϑ(R(x, u), A(u)) = Rϑ β A. in Fig. 3 represents the available information R by employing
A (u )> β
which we obtain the values of λ, R(x, y), and RT xλ . First, λ =
Hence, Rϑ α A ⊆ Rϑ β A. Rϑ A(x) is the degree of the object x ∈ U certainly belonging
P3.2.4) It is straightforward to get RT A = RT 0 A by defini- to the decision class A. Second, R(x, y) is the similarity degree
tion of upper approximation of FVPRS. between x and y. Finally, RT xλ (y) = T (R(x, y), λ) represents
P3.2.5) ∀x ∈ U : the degree of the object y belonging to the decision class A.
The value RT xλ (y) is obtained by aggregating the similarity
RT A(x) = sup T (R(x, u), A(u))
A (u )≥N (0) relation R(x, y) and the degree of the object x belonging to the
decision class A, i.e., y. The aforementioned analysis shows that
∨ sup T (R(x, u), A(u)) RT xλ measures the supporting degree of condition attributes R
A (u )< N (0)
to the decision class A. For example, if R(x, y) = 1 and λ = 1,
≥ sup T (R(x, u), N (α)) then the set of condition attributes R supports the decision class
A (u )≥N (α )
A 100% on the objects x and y.
∨ sup T (R(x, u), A(u)) = RT α A(x). In the following, we focus on the granular representation of
A (u )< N (α ) the fuzzy lower and upper approximation operators by using the
P3.2.6) ∀x ∈ U : basic granules in FRS. As preliminaries, we need to review a
concept called fuzzy cut set. In fuzzy sets, let f be a fuzzy set
RT α A(x) defined on U , then the fuzzy α-cut set of fuzzy set f is defined as

= sup T (R(x, u), N (α)) f (x), f (x) > α
A (u )≥N (α )
α
f (x) = , for α ∈ [0, 1).
0, f (x) ≤ α
∨ sup T (R(x, u), N (α))
N (β )≤A (u )< N (α ) By using the fuzzy granule RT xλ , the granular representation
of the fuzzy lower and upper approximation operators in FVPRS
∨ sup T (R(x, u), A(u))
A (u )< N (β ) is described as follows.
Definition 3.2: Given F D = (U, R ∪ D) and A ∈ F (U ), the
≥ sup T (R(x, u), N (β)) fuzzy lower and upper approximations of FVPRS are, respec-
A (u )≥N (α )
tively, defined as follows.
∨ sup T (R(x, u), N (β)) D3.2.1) Rϑ α A = ∪{RT xλ : (RT xλ )α ⊆ A};
N (β )≤A (u )< N (α )
D3.2.2) RT α A = ∪{RT xλ : xλ ⊆ AN (α ) }.
∨ sup T (R(x, u), A(u)) = RT β A(x). where
A (u )< N (β )
N (α), A(x) ≥ N (α)
Hence, RT β A ⊆ RT α A. AN (α ) = .
A(x), otherwise
By Properties P3.2.1) and P3.2.2), it is straightforward to get
P OSR D = P OSR 0 D and P OSR D ⊆ P OSR α D . D3.1.1) and D3.1.2) can be seen as the membership function

It has been proposed and proven in [14] that the elements in representation of Rϑ α A and RT α A. D3.2.1) and D3.2.2) can
MT = {RT xλ : x ∈ U, λ ∈ (0, 1]} can be applied as the basic seen as the granular representation of Rϑ α A and RT α A. Now,
granules for the FRS framework FTR (U ) = {A : Rϑ A = A = the task is to prove these two types of representation ways are
RT A}. The fuzzy granule RT xλ is important and meaningful equivalent. As preliminaries, we present a theorem about the
in a fuzzy rough framework. From the theoretical viewpoint, the granular representation of Rϑ α A and RT α A as follows.
granule RT xλ is not only used to construct the fuzzy approxi- Theorem 3.1:
mation operators, but also used to study the structure of attribute T3.1.1) Let γ = (∪{RT xλ : (RT xλ )α ⊆ A})(x) for x ∈
U , then the following statements hold: (RT xγ )α ⊆
A always holds; (RT xβ )α ⊆ A does not hold for (∪{RT xλ : (RT xλ )α ⊆ A})(x) = λ. Thus,
any β > γ.
T3.1.2) ∪{RT xλ : xλ ⊆ AN (α ) } = ∪{RT xλ : λ = inf ϑ(R(x, u), α) ∧ inf ϑ(R(x, u), A(u))
A (u )≤α A (u )> α
AN (α ) (x)}.
Proof: = (∪{RT xλ : (RT xλ )α ⊆ A})(x)
T3.1.1) By the condition of γ = (∪{RT xλ : (RT xλ )α ⊆ holds.
A})(x), we get for any x ∈ U there ex- Theorem 3.3: Given F D = (U, R ∪ D) and A ∈ F (U ), the
ist t ∈ (0, 1] and y ∈ U satisfying γ = RT yt (x) following two statements are equivalent.
and (RT yt )α ⊆ A. Then, the two statements, T3.3.1) RT α A(x) = sup T (R(x, u), N (α))
A (u )≥N (α )
T (R(y, x), t) = γ and if T (R(y, z), t) > α, then
T (R(y, z), t) ≤ A(z) for any z ∈ U , hold. Thus, the ∨ sup T (R(x, u), A(u)), ∀x ∈ U ;
A (u )< N (α )
statement ∀z ∈ U , RT xγ (z) = T (R(x, z), γ) = T3.3.2) RT α A = ∪{RT xλ : xλ ⊆ AN (α ) }.
T (R(x, z), T (R(y, x), t)) ≤ T (R(y, z), t) holds. If
Proof: By Theorem 3.1, RT α A = ∪{RT xλ : λ =
T (R(x, z), t) > α, then T (R(y, z), t) > α. Thus, we
AN (α ) (x)}. Then,
have T (R(x, z), t) ≤ A(z). Hence, (RT xγ )α ⊆ A.
Assume (RT xβ )α ⊆ A for β > γ, then RT α A(x)
γ ≥ RT xβ (x) = β by γ = (∪{RT xλ : (RT xλ )α ⊆
A})(x). That is to say, γ ≥ β. This contradicts the = (∪{RT uλ : λ = AN (α ) (u)})(x)
condition β > γ. Thus, for any β > λ, the formula
= sup{RT uλ (x) : λ = AN (α ) (u)}
(RT xβ )α ⊆ A does not hold.
T3.1.2) If xλ ⊆ AN (α ) , then λ ≤ AN (α ) (x); ∀x, z ∈ U , = sup {RT uN (α ) (x)} ∨ sup {RT uA (u ) (x)}
RT xλ (z) = T (R(x, z), λ). By the monotonic- A (u )≥N (α ) A (u )< N (α )
ity of the T -norm, we have RT xλ ⊆ RT xλ for = sup T (R(x, u), N (α)) ∨ sup T (R(x, u), A(u))
∀xλ ⊆ AN (α ) and λ = AN (α ) (x). Thus, (∪{RT xλ : A (u )≥N (α ) A (u )< N (α )
xλ ⊆ AN (α ) }) ⊆ (∪{RT xλ : λ = AN (α ) (x)}). i.e.,

Hence, ∪{RT xλ : xλ ⊆ AN (α ) } = ∪{RT xλ : λ =
sup T (R(x, u), N (α)) ∨ sup T (R(x, u), A(u))
AN (α ) (x)}. A (u )≥N (α ) A (u )< N (α )
By employing Theorem 3.1, we describe and prove the fol-
lowing two theorems. = (∪{RT xλ : xλ ⊆ AN (α ) })(x).
Theorem 3.2: Given F D = (U, R ∪ D) and A ∈ F (U ), the Theorems 3.2 and 3.3 show that the membership function
following two statements are equivalent. representation and granular representation of Rϑ α A and RT α A
T3.2.1) Rϑ α A(x) = inf ϑ(R(x, u), α) ∧ inf ϑ(R(x, u), are equivalent. They are the key theorems for the construction
A (u )≤α A (u )> α
A(u)), ∀x ∈ U ; of discernibility matrix.
T3.2.2) Rϑ α A = ∪{RT xλ : (RT xλ )α ⊆ A}. From Theorem 3.2, we find that a fuzzy cut set is introduced
Proof: ∀x ∈ U into the lower approximation operators in FVPRS, and the con-
dition RT xλ ⊆ A in the granular representation of FRS is re-
Rϑ α A(x) = inf ϑ(R(x, u), α) ∧ inf ϑ(R(x, u), A(u))
A (u )≤α A (u )> α laxed by the condition (RT xλ )α ⊆ A. We can illustrate this fact
by some demonstration graphs (see Fig. 4). Assuming α = 0.2,
= inf (sup{c ∈ [0, 1], T (R(x, u), c) ≤ α}) we need to check whether the fuzzy cut sets (f1 )0.2 , (f2 )0.2 , and
A (u )≤α
(f3 )0.2 are included in A0.2 (since (RT xλ )α ⊆ A is equivalent
∧ inf (sup{c∈[0, 1], T (R(x, u), c)≤A(u)}) to (RT xλ )α ⊆ Aα ).
A (u )> α
Fig. 4 shows that (f1 )0.2 is not included in A0.2 , whereas
= inf (sup{c ∈ [0, 1], (RT xc )α (u) ≤ A(u)}) (f2 )0.2 and (f3 )0.2 are included in A0.2 . That is to say, when
A (u )≤α
we ignore some small membership degrees, some more fuzzy
∧ inf (sup{c ∈ [0, 1], (RT xc )α (u) ≤ A(u)}) sets (e.g., f2 ) are then selected to construct the fuzzy lower
A (u )> α
approximation. As a result, the positive region becomes larger
= inf (sup{c ∈ [0, 1], (RT xc )α (u) ≤ A(u)}). and the knowledge representation power becomes stronger.
u ∈U
In FVPRS, we omit some misclassification or small perturba-
Let λ = inf u ∈U λu = inf u ∈U (sup{c ∈ [0, 1], (RT xc )α (u) ≤ tion by controlling the threshold α in the fuzzy cut set (RT xλ )α .
A(u)}), here λu = sup{c ∈ [0, 1], (RT xc )α (u) ≤ A(u)}. As a result, the values of set approximations are flexible and less
Then, for any u ∈ U , (RT xλ )α (u) ≤ (RT xλu )α (u) ≤ A(u)} sensitive to misclassification and perturbation, and further, the
holds. For any γ > λ, γ ∈ [0, 1], ∃u ∈ U , (RT xγ )α (u) > attribute reduction results are robust. In a word, FVPRS is less
A(u)} exists. That is to say, λ is the maximum value to sensitive to misclassification and perturbation.
make (RT xλ )α (u) ≤ A(u) hold for any u ∈ U . Then, Since P OSR α (D)(x) = ∪z ∈U Rϑ α [z]D (x), we know that
(RT xλ )α ⊆ A holds for any x ∈ U , and for any β > λ, P OSR α (D)(x) can get its maximum value at cer-
(RT xβ )α ⊆ A does not hold. By Theorem 3.1, we get tain Rϑ α [z]D (x). The following theorem implies that
Theorem 3.4: Given F D = (U, R ∪ D) with the decision

symbolic attributes and ∀x ∈ U , the following statement holds:
P OSR α (D)(x) = Rϑ α [x]D (x).
Proof: Since P OSR α (D)(x) = ∪z ∈U Rϑ α [z]D (x), we get
that P OSR α (D)(x) can get its maximum value at certain
Rϑ α [z]D (x). Let λ = P OSR α (D)(x), then Rϑ α [z]D (x) = λ
and Rϑ α [y]D (x) ≤ λ for every y ∈ U . By Theorem 3.1,
(RT xλ )α ⊆ [z]D holds. By Lemma 3.2, (RT xλ )α ⊆ [x]D
holds. Then, λ ≤ Rϑ α [x]D (x) holds. Thus, λ = Rϑ α [x]D (x),
i.e., P OSR α (D)(x) = Rϑ α [x]D (x).
By Theorem 3.4, we know that in a fuzzy decision table
with the decision symbolic attributes, P OSR α (D)(x) always
gets its value at its lower approximation value of [x]D , i.e.,
Rϑ α [x]D (x). Thus, we conclude that keeping the positive re-
gion value P OSR α (D)(x) invariant is equivalent to keeping
Rϑ α [x]D (x) invariant.
Theorem 3.5: For two T -similarity relations R and P , if
R ⊆ P , then the following hold.
T3.5.1) Rϑ α A ⊇ Pϑ α A;
T3.5.2) RT α A ⊆ PT α A;
T3.5.3) P OSR α D ⊇ P OSP α D.
Proof:
T3.5.1) R ⊆ P ⇒ ∀x, u ∈ U, R(x, u) ≤ P (x, u) ⇒ ∀x, u ∈ U,
inf ϑ(R(x, u), α) ∧ inf ϑ(R(x, u), A(u)) ≥
A (u )≤α A (u )> α
inf ϑ(P (x, u), α) ∧ inf ϑ(P (x, u), A(u)) ⇒
A (u )≤α A (u )> α
Rϑ α A ⊇ Pϑ α A.
T3.5.2) R ⊆ P ⇒ ∀x, u ∈ U, R(x, u) ≤ P (x, u) ⇒ ∀x, u ∈ U,
sup T (R(x, u), 1 − α) ∧ sup T (R(x, u), A(u))
A (u )≥1−α A (u )< 1−α
≤ sup T (P (x, u), 1 − α) ∧ sup T (P (x, u),
A (u )≥1−α A (u )< 1−α
A(u)) ⇒ RT α A ⊆ PT α A.
T3.5.3) R ⊆ P ⇒ Rϑ α A ⊇ Pϑ α A [by T3.5.1)] ⇒
P OSR α D ⊇ P OSP α D (by Theorem 3.4).
It is easy to see from Theorem 3.5 that the smaller the fuzzy
T -similarity relation is, the larger the fuzzy lower approximation
and fuzzy positive region are.
Fig. 4. Inclusion relation. (a) (f1 )0 . 2 and A 0 . 2 . (b) (f2 )0 . 2 and A 0 . 2 . (c)
(f3 )0 . 2 and A 0 . 2 .
B. Attribute Reduction With FVPRS

In this section, we propose a general concept of attribute
P OSR α (D)(x) always takes Rϑ α [x]D (x) as its maximum
reduction by employing the fuzzy approximation operators in
value in a fuzzy decision table with the decision symbolic
FVPRS. A fuzzy decision table with condition fuzzy attributes
attributes.
and decision symbolic attributes is adopted as the platform of
Lemma 3.1: Rϑ α and RT α are monotone.
attribute reduction since it occurs in most practical datasets.
Proof: Suppose A ⊆ B ⊆ U ; if (RT xλ )α ⊆ A, then Definition 3.3: Given F D = (U, R ∪ D), P ⊆ R, and α ∈
(RT xλ )α ⊆ B. By Rϑ α A = ∪{RT xλ : (RT xλ )α ⊆ A}, we [0, 1), if the following two statements hold: P OSR α (D) =
get Rϑ α A ⊆ Rϑ α B. Hence, Rϑ α is monotone. By RT α A = P OSP α (D) and P is α-independent in R, then P is called
∪{RT xλ : xλ ⊆ A1−α } and A1−α ⊆ B1−α , we get RT α A ⊆ the α-reduction of R.
RT α B. Hence, RT α is monotone. It is easy to see that P is a minimal subset of R to keep the
Lemma 3.2: If (RT xλ )α ⊆ [z]D , then (RT xλ )α ⊆ [x]D . positive region invariant.
Proof: If [z]D = [x]D , it is straightforward to get (RT xλ )α ⊆ The α-core of R is the collection of the attributes that are
[x]D . If [z]D = [x]D , RT xλ (x) ≤ α, i.e., λ ≤ α holds (by the α-indispensable in R.
definition of fuzzy cut set). Thus, ((RT xλ )α = φ) ⊆ [x]D al- By Theorem 3.4, we have the following theorem to charac-
ways holds for λ ≤ α. terize the α-reduction.
Theorem 3.6: Given F D = (U, R ∪ D), P ⊂ R, and α ∈ where ∨(cij ) is the disjunction of all variables a such
[0, 1). The following two statements are equivalent: T3.6.1) that a ∈ cij . Let gα (F D) be the reduced disjunctive form
P ⊂ R contains one α-reduction of R and T3.6.2) ∀x ∈ U, of fα (F D) obtained from fα (F D) by applying the mul-
T (P (x, z), λ) ≤ α for every z ∈ / [x]D and λ = Rϑ α [x]D (x). tiplication and absorption laws as many times as possi-
Proof: T3.6.1) ⇒ T3.6.2) If P contains one α-reduction of R, ble. Then, there exist l and Rk ⊆ R for k = 1, . . . , l such
then P OSR α (D) = P OSP α (D). Let λ = P OSR α (D)(x) = that gα (F D) = (∧R1 ) ∨ · · · ∨ (∧Rl ) where every element
P OSP α (D)(x) ∀x ∈ U . By Theorems 3.1 and 3.4, we have in Rk only appears one time. We have the following
λ = Rϑ α [x](x) = Pϑ α [x](x) and (PT xλ )α ⊆ [x]D . By the theorem.
definition of fuzzy cut set, we have T (P (x, z), λ) ≤ α for every Theorem 3.9: RedD α (R) = {R1 , . . . , Rl }; here, RedD α (R)
z∈ / [x]D . is the collection of all the α-reductions of R.
T3.6.2) ⇒ T3.6.1) By Theorem 3.5, λ = Rϑ α [x]D (x) ≥ The proof is omitted since this theorem is similar to the one
Pϑ α [x]D (x) holds. If T (P (x, z), λ) ≤ α for every in [16].
z∈ / [x]D , then (PT xλ )α ⊆ [x]D . By Rϑ α A = ∪{RT xλ : Theorem 3.9 shows that we can find all reductions of R by
(RT xλ )α ⊆ A}, Pϑ α [x]D (x) ≥ (PT xλ )α (x) = λ holds. using an α-discernibility matrix.
Thus, λ = Rϑ α [x](x) = Pϑ α [x](x). By Theorem 3.4, An algorithm to compute all α-reductions for a fuzzy decision
λ = P OSR α (D)(x) = P OSP α (D)(x) holds. This shows that system is then designed (see Algorithm 3.1).
P ⊂ R contains one α-reduction of R. Algorithm 3.1: To find all α- reductions of FVPRS:
Clearly, P is a α-reduction in F D = (U, R ∪ D) if and Step 1: Compute the similarity relation of the set of all condi-
only if P is a minimal subset satisfying the conditions in tion attributes: SIM (R);
Theorem 3.6. Now, we can construct a α-discernibility matrix. Step 2: Compute Rϑ α ([x]D ) for every x ∈ U ;
Suppose U = {x1 , x2 , x3 , . . . , xn }, by Mα (U, R, D) we de- Step 3: Compute cij : cij = {α : T (a(xi , xj ), λ) ≤ α}, λ =
note an n × n matrix (cij ), called the α-discernibility matrix of Rϑ α [xi ]D (xi ) for D(xi , xj ) = 0; otherwise Cij = φ;
(U, R ∪ D), such that Step 4: Compute CoreD α (R) = {a : cij = {a}}; delete
M3.1) cij = {a : T (a(xi , xj ), λ) ≤ α}, those cij with nonempty overlap with CoreD α (R);
λ = Rϑ α [xi ]D (x) for D(xi , xj ) = 0; Step 5: Define fα (F D) = ∧{∨(cij )} with nonemtpy cij
M3.2) cij = φ, for D(xi , xj ) = 1. and computing gα (F D) = (∧Rl ) ∨ · · · ∨ (∧Rl ) by
Theorem 3.7: CoreD α (R) = {a : cij = {a}} for some 1 ≤ fα (DS);
i, j ≤ n. Step 6: Output RedD α (R) = CoreD α (R) ∪ Rl , . . . , Core ∪
Proof: a ∈ CoreD α (R) ⇔ P OSR α D = P OS(R − {a}) α D ⇔ Rl .
there exists xi ∈ U such that P OSR α D(xi ) > This algorithm to find all α-reductions is an NP-complete
P OS(R − {a}) α D(xi ) ⇔ Rϑ ([xi ]D )(xi ) > (R − {a})ϑ problem [30]. In practical applications, it is enough to employ
([xi ]D )(xi ) ⇔ Let λ(xi ) = Rϑ ([xi ]D )(xi ) and a near-optimal reduction to address the real problems. Thus, we
t(xi ) = (R − {a})ϑ ([xi ]D )(xi ), then λ(xi ) > t(xi ), design a heuristic algorithm to find one near-optimal reduction
by using the following steps to replace steps 5 and 6 in Algorithm
RT (xi )λ(x i ) ⊆ [xi ]D , (R − {a})T (xi )t(x i ) ⊆ [xi ]D , and
3.1:
(R − {a})T (xi )λ(x i ) ⊂ [xi ]D (by Theorem 3.1) ⇔ there Step 5: Let Reduct = CoreD α (R);
exists xj ∈ U such that T ((R − {a})(xi , xj ), λ(xi )) > α Step 6: Add the element a with maximum of frequency of
and T (R(xi , xj ), λ(xi )) ≤ α ⇔ for any b ∈ R − {a}, occurrence into Reduct, and delete those cij with
T (b(xi , xj ), λ(xi )) > α, and T (a(xi , xj ), λ(xi )) ≤ α hold ⇔ nonempty overlap with Reduct;
cij = {a}. Step 7: If there still exist some cij = φ, go to step 6; otherwise,
The statement cij = {a} implies that a is the unique attribute go to step 8;
to maintain T (R(xi , xj ), λ(xi )) ≤ α. Step 8: If Reduct is not independent, delete the redundant ele-
This theorem shows that by using α-discernibility matrix, we ments in Reduct;
can find the α-core. Step 9: Output Reduct.
Theorem 3.8: If P ⊂ R, then P contains α-reduction of R if
and only if P ∩ cij = φ for every cij = φ.
The proof is straightforward by Theorem 3.6 and the defini- IV. EXPERIMENTAL COMPARISONS AND SCALABILITY
tion of cij . ANALYSIS
Corollary 3.2: If P ⊂ R, then P is one α-reduction of R if In this section, we compare and analyze FVPRS with sev-
and only if P is a minimal subset of R satisfying P ∩ cij = φ eral dimension reduction methods. These methods can be
for every cij = φ. roughly divided into two types. One type focuses on several
Theorem 3.8 and Corollary 3.2 show that by using α- RS-based methods. They are RS, FRS, VPRS, VPRS-FRS,
discernibility matrix, we can find the α-reduction. and VPFRS. The other type focuses on two-feature (i.e., at-
A discernibility function fα (F D) for F D = (U, R ∪ D) is tributes) dependency-based methods. They are MRMR and CA-
a Boolean function with m Boolean variables a1 , . . . , am cor- CLF [31], [42].
responding to the attributes a1 , . . . , am , respectively, and is de- Before we conduct comparison and analysis, two issues
fined as fα (F D)(a1 , . . . , am ) = ∧{∨(cij ) : 1 ≤ j < i ≤ n}, should be specified. One issue is that in this section, some
TABLE I
INFORMATION OF SOME DATASETS FROM UCI MACHINE
LEARNING REPOSITORY
Fig. 5. Procedure and scalability analysis of attribute reduction in FVPRS.
methods (i.e., RS, VPRS, and FRS) are designed to find the
reductions by using the discernibility matrix approach. Others or VPFRS,” the algorithms to find one near-optimal reduction
(i.e., VPRS-FRS and VPFRS) are designed by keeping the de- of VPRS-FRS or VPFRS are then obtained.
pendency function invariant. Another issue is that the algorithm From the aspect of computational complexity, we find that
mentioned in this section is the heuristic to find a near-optimal the main difference of FVPRS, VPRS-FRS, and VPFRS lies
reduction, since it is NP-hard to find the optimal one. in the lower approximations. The computational complexity to
compute the lower approximation of FVPRS and VPRS-FRS is
A. Scalability Analysis and Comparison similar (i.e., O(|U |2 × |D|) in crisp decision case and O(|U |3 )
in real-valued decision case); the computational complexity of
In this section, we illustrate the procedure of attribute reduc- VPFRS is O(|U |4 × |D|) in crisp decision case (O(|U |5 ) in
tion in FVPRS (see Fig. 5). The procedure of FVPRS is similar real-valued decision case). From the aforementioned analysis,
to the procedures of RS, VPRS, and FRS. The main differ- we find that the real-valued decision case spends more time
ence among them lies in the steps covered by “Preprocessing.” and space than the crisp decision case to compute the lower
In the fuzzy cases, i.e., FRS and FVPRS, the “Preprocessing” approximation.
step means computing similarity relation and lower approxi- Algorithm 4.1 (heuristic):
mation, whereas in crisp cases, i.e., RS and VPRS, it means Step 1: Compute SIM (R);
the discretization of the original datasets. As a result, the com- Step 2: Compute Rϑ α ([x]D ) for every x ∈ U ;
putational complexity in the fuzzy cases is O(|U |2 × |A|) (the Step 3: Compute P OSR α (D) = ∪x∈U Rϑ α [x]D and γR α(D)=
similarity relation) and O(|U |2 × |D|) (the lower approxima- |P OSR α (D)| / |U |;
tion), whereas the computational complexity in the crisp cases Step 4: Add the attribute with maximum γ{{α }∪R edu ct}α (D)
depends on the selection of discretization method (the compu- into Reduct and delete it from R;
tational complexity of the selected discretization method in this Step 5: If γ{R edu ct}α (D) = γR α (D), go to step 4; otherwise,
paper is O(|U |2 × |A|) [43]). By the aforementioned analysis, it go to step 6;
is easy to see that FVPRS and FRS spend similar time and space Step 6: If Reduct is not independent, delete the redundant ele-
to find one reduction, whereas RS and VPRS spend similar time ments in Reduct;
and space. Step 7: Output Reduct.
In Fig. 5, |U | is the size of the universe, |A| is the number of
attributes, |D| is the number of the decision classes, and |k| is
B. Experimental Setup
the number of selected attributes in the obtained reduction.
Since, up to now, there was no algorithm that used discerni- The experiments in this section are set up as follows.
bility matrix to find the reductions for VPRS-FRS and VPFRS, Dataset: Ten datasets from University of California, Irvine
we need design some heuristic algorithms by keeping depen- (UCI) Machine Learning Repository [35] are used (see Table I).
dency function invariant for them. In the following, we design Classifier: Nearest neighbor algorithm in [36] is used to be
a heuristic algorithm by keeping dependency function invariant the classifier. [Note that decision tree is used in the regression
for FVPRS (see Algorithm 4.1). If we replace step 2 in Algo- problems (i.e., those datasets with continuous decision attributes
rithm 4.1 by “computing the lower approximation of VPRS-FRS such as forest fires and housing).]
Fig. 6. Trend of the number of selected attributes varying with the threshold.
Dataset Split: In the process of classification, the dataset after (multiclass or two-class or continuous). We study the different
attribute reduction is split into two parts. The randomly chosen dimension reduction methods on these datasets and we believe
50% of objects are used as the training set and the remainder that these datasets suit for the dimension reduction methods
as the testing set. The classification result is the average of 20 under different conditions.
times training and testing.
Indexes: They are: 1) the number of selected attributes and
2) the classification accuracy of the reduction.
Parameter Specification: In FVPRS, VPRS-FRS, and C. Effect of Threshold in FVPRS
VPFRS, we try the threshold from 0 to 1 (not including 1) To clearly show the effect of the threshold in FVPRS, we set
with step 0.01. In VPRS, we try the threshold from 0 to 0.5 the threshold from 0.9 to 1 with the step 0.005 and summarize the
(not including 0.5) with step 0.01. In MRMR, we try the results in Figs. 6 and 7, where the horizontal axis represents the
parameter from 1 to the number of attributes in the correspond- threshold in FVPRS and the vertical axes represent the number
ing dataset. of attributes and the accuracy, respectively. The line with
Triangular norm: “Lukasiewicz” T -norm is selected to con- in Fig. 7 is the accuracy of the original dataset, which provides
struct the lower approximations of these RS-based generalized a baseline to evaluate the performance of the reductions.
methods, i.e., FRS, FVPRS, VPRS-FRS, and VPFRS. Fig. 6 shows that the number of attributes moves downward
The first six datasets with different target classes (multiclass along the threshold. That is to say, the larger the threshold is,
or two-class) are used in studying the effect of the threshold in the smaller the number of selected attributes is. Fig. 7 shows
FVPRS. What is more, these six datasets are used as the datasets that the accuracy after reduction is close to the accuracy before
with crisp target class in comparing FVPRS with other RS-based reduction at the beginning, but it suddenly becomes smaller
methods. The last two datasets are used as the datasets with real- than the accuracy before reduction when the threshold is large
valued decision in comparing FVPRS, VPFRS, and VPRS-FRS. enough. All these give us a guideline on how to specify the
The remaining two datasets are used in comparing FVPRS with threshold for the further applications of FVPRS, i.e., the larger
CACLF. the threshold is, the smaller the number of selected attribute
Table I shows that the selected ten datasets differ greatly from is, whereas if the threshold is too large, the accuracy becomes
sample size, feature number, data distribution, and target class smaller than the accuracy before reduction.
Fig. 7. Trend of classification accuracy varying with the threshold.
D. Experimental Comparison and Analysis of FVPRS- and TABLE II

COMPARISON OF FVPRS, RS, FRS, AND VPRS
RS-Based Methods
In this section, we first compare FVPRS with RS, FRS, and
VPRS [17], [24], [41], and then we compare FVPRS with VPRS-
FRS, VPFRS, and VQRS.
1) Experimental Comparison of FVPRS With RS, FRS, and
VPRS: We would like to point out the main differences of these
methods as follows: 1) RS and VPRS are the methods to deal
with the datasets with symbolic values, whereas FVPRS and
FRS are the methods to handle the datasets with real values and
2) in the preprocessing step, RS and VPRS need to discretize the
real-valued datasets, whereas FVPRS and FRS need to compute
the similarity relation and the fuzzy lower approximations.
Discretizaton of real-valued attributes is an important step
for RS and VPRS since their attribute reduction performances
depend on this step. In this paper, the discretization method These show that RS has maintained the information of the orig-
proposed by Nguyen and Nguyen is chosen since it not only inal datasets, but its attribute reduction result is not compacter
preserves the discernibility information in the original datasets than FVPRS because RS was sensitive to misclassification.
but also is fast in discretization [43]. Next, Table II shows that FVPRS often finds compacter at-
We summarize the comparison results in Table II. First, we tribute reductions than FRS does. That is to say, the dimension
find that RS has comparable classification power (0.8438) with reduction power of FRS on the datasets with real values is
FVPRS (0.8351 and 0.8471), but the number of selected at- relatively weak. The reason is that FRS was sensitive to mis-
tributes in RS (8.5) is much larger than FVPRS (4.33 and 5). classification and perturbation.
Finally, Table II shows that VPRS has a better reduction power TABLE III
COMPARISON AMONG FVPRS, VPFRS, AND VPRS-FRS ON DATASETS WITH
than FVPRS (from the viewpoint of the number of selected THE CRISP DECISION
attributes), whereas it misses some useful information. It is easy
to see that the classification accuracy of VPRS (0.7646) is far
less than the accuracy of FVPRS (0.8351 and 0.8471). This
is because VPRS was not flexible enough for the real-valued
attributes. In contrast, FVPRS is a flexible model.
2) Comparisons of FVPRS With VPRS-FRS, VPFRS, and
VQRS: Before comparison, we highlight some differences
among these methods.
One main difference of these methods is on the concepts of
lower approximations. VPRS-FRS was proposed by introducing
the β-precision aggregation into the set approximations of one
specified FRS framework; VPFRS was proposed by introducing
one crisp cut set into the set approximations of a general FRS
framework; FVPRS is proposed by incorporating a fuzzy cut
set into a general FRS framework; VQRS was proposed by
introduced a fuzzy quantifier into the set approximations.
Another difference is that VPRS-FRS and VPFRS are ef-
fective to handle misclassification but sensitive to perturbation,
whereas FVPRS and VQRS are less sensitive to both misclas-
sification and perturbation.
A third difference is about the applications of these methods,
i.e., attribute reduction. In VPRS-FRS and VPFRS, attribute
reduction was not yet mentioned; FVPRS not only proposes
the concepts of attribute reduction, but also designs some de-
tailed algorithms of attribute reduction; VQRS could not design
the algorithm of attribute reduction by keeping approximation
quality invariant since the monotonicity of quality measure did perturbation, whereas VPRS-FRS and VPFRS are sensitive to
not hold in VQRS. Note that although monotonicity of quality perturbation.
measure no longer holds, it is still possible to define attribute Two types of reductions are selected to compare the classifi-
reductions in VQRS [51]. Since the key idea of keeping approx- cation accuracy: one type has the classification accuracy com-
imation quality measure invariant is adopted in this paper, we parable to or better than the accuracy of the original dataset, as
omit the comparison of attribute reduction between FVPRS and summarized in Table III(b), while the other has the best clas-
VQRS. sification accuracy in each method on the specified dataset, as
The last difference among these methods is about the compu- summarized in Table III(a).
tational complexity. We omit this point since we have mentioned Table III shows that the numbers of selected attributes in
it in Section IV-A. VPFRS are obviously smaller than the ones in VPRS-FRS in
In the following, we conduct the detailed experimental com- most cases, and the numbers of selected attributes in FVPRS are
parison among these methods (except VQRS). The experimental significantly smaller than the ones in VPFRS and VPRS-FRS.
comparison is twofold. One is experimental comparison on the What is more, FVPRS can find the attribute subset with high-
datasets with the crisp decision while the other is on the datasets est classification and smallest number of selected features [see
with the real-valued decision. Table III(a)]. All these show that FVPRS has a better perfor-
a) Experimental comparison on the databases with the mance of dimension reduction than VPFRS and VPRS-FRS
crisp decision: The experimental comparison results are sum- because FVPRS is less sensitive to both misclassification and
marized in Table III and Fig. 8. Fig. 8 summarizes the distri- perturbation, whereas VPFRS and VPRS-FRS were only less
bution of those reductions that have the classification accuracy sensitive to misclassification.
comparable to or better than the accuracy of the original dataset. b) Experimental comparison on the datasets with the real-
In each subfigure, the horizontal axis represents the number of valued decision: The comparison results have been summarized
selected attributes, and each line with some points represents the in Table IV, which shows that under the condition with similar
reduction distribution of a certain method. For example, the line regression results, FVPRS has the smallest number of selected
with diamonds represents the reduction distribution of FVPRS. attributes; VPFRS has a smaller number of selected attributes
Fig. 8 shows that FVPRS distributes more densely than than VPRS-FRS. We then draw a conclusion that VPFRS is
VPRS-FRS and VPFRS on the left side of each subfigure. That time-consuming (we have mentioned this in Section IV-A);
is to say, FVPRS often finds some compacter reductions than VPRS-FRS has a poor reduction performance; FVPRS is an
VPRS-FRS and VPFRS do. The reason is that FVPRS can ef- acceptable choice to find the compacter attribute subset within
fectively handle the problems with both misclassification and the reasonable time and space.
Fig. 8. Reduction distribution of FVPRS, VPRS, and VPRS-FRS.

TABLE IV Table V. Here, we randomly choose n (n < 10) reductions to
COMPARISON AMONG FVPRS, VPFRS, AND VPRS-FRS ON THE DATASETS
WITH THE REAL-VALUED DECISION
construct the ensembles of reductions.
Table V shows that under the condition with the similar num-
ber of attributes, FVPRS (ensemble) often has the better clas-
sification accuracy than FVPRS (single). FVPRS (single) has
the comparable classification accuracy with MRMR. For exam-
ple, the average of classification accuracy in FVPRS (ensem-
ble) (0.861, 0.855, 0.847) is often higher than FVPRS (single)
(0.836, 0.829, 0.816) and MRMR (0.829, 0.826, 0.818). The
E. Comparison of FVPRS With Two-Feature Dependency- computational complexity of MRMR increases with the number
Based Methods: MRMR and CACLF of selected attributes. The computational complexity of FVPRS
increases with the square of the number of attributes at the worst
In this section, we would like to compare FVPRS with MRMR case. All these show that MRMR and FVPRS suit for different
and CACLF. applications: MRMR is fast to find feature selection, whereas
1) Comparison of FVPRS With MRMR: MRMR was pro- FVPRS has a better classification performance (especially the
posed as a well-designed feature selection method based on the ensemble approach).
mutual information [31]. Its key idea to find the feasible subset Note that to further improve the ensemble, a genetic algo-
of attributes was to minimize feature redundancy and maxi- rithm, as mentioned in [44], can be used to find the best ensem-
mizing the feature dependency of the target classes, which was ble of reductions. Also, the method proposed in [45] by using the
similar to FVPRS. Now, it is necessary to compare FVPRS with accuracy-guided forward searching strategy to find the best en-
MRMR. semble of reductions is another choice. Since these two methods
It is already pointed out in [44] and [45] that the ensembles of are computationally consuming, we omit them here and inter-
reductions have a better performance than a single reduction. As ested readers are referred to [44] and [45] for details.
a result, it is necessary to conduct the ensembles of reductions to 2) Comparison of FVPRS With CACLF: In this section, we
compare with the classifiers constructed on a single reduction. compare FVPRS with CACLF, which is one method of dimen-
In the following, we conduct the comparisons among FVPRS sion reduction by considering the feature dependency. Before
(ensemble), FVPRS (single), and MRMR, as summarized in
TABLE V ter than CACLF. For example, under the condition with similar
COMPARISON AMONG FVPRS (ENSEMBLE), FVPRS (SINGLE), AND MRMR
number of features, the accuracy of FVPRS is comparable to
the accuracy of CACLF with the measure Q1 . If we relax this
condition, FVPRS can find the feature subset with better accu-
racy than CACLF with the measure Qm o d in most cases. All the
aforementioned comparison shows that these two methods are
suitable for different applications: FVPRS can find the reduc-
tions with better classification performance, whereas CACLF
can find the compacter feature subset.
Remarks: FVPRS has its strengths in comparing with other
methods. However, it still has some limitations. One limitation
of FVPRS is that there is a parameter, i.e., the threshold in
fuzzy lower approximation operators, to be specified. Although
the specification of the parameter depends on the domain of
problems and the requirements of the classification accuracy, it
is necessary to point out the extent of the reasonable parameter.
Now one guide of specifying this parameter is that the larger the
threshold is, the smaller the number of selected attributes is.
V. CONCLUSION
In this paper, we have set up a framework named FVPRS by
combining FRS and VPRS. In FVPRS, a controlled threshold is
introduced into knowledge representation of FRS (in the sight of
granular computing, a fuzzy cut set is introduced into FRS), and
then, the concepts of the fuzzy lower and upper approximations
with variable precision were proposed. We briefly list several
characterizations of FVPRS as follows.
1) FVPRS takes the existing FRS as a special case when the
threshold is set to zero.
2) FVPRS can find all the reductions by employing the dis-
TABLE VI
COMPARISON BETWEEN FVPRS AND CACLF cernibility matrix approach.
3) The methodology of FVPRS is less sensitive to misclassi-
fication and perturbation.
ACKNOWLEDGMENT
The authors would like to thank the reviewers for their valu-
able suggestions. They would also like to thank M. Kyle for his
help in the presentation of this paper.
experimental comparison, we would like to point out some dif- REFERENCES

ferences between them. The main difference between them is
[1] Z. Pawlak, “Rough sets,” Int. J. Comput. Inf. Sci., vol. 11, pp. 341–356,
that FVPRS is a feature selection method, whereas CACLF 1982.
was a feature extraction method since it found new features by [2] Z. Pawlak, Rough Sets: Theoretical Aspects of Reasoning About Data [M].
combining some original features. Another difference is that in Dordrecht, The Netherlands: Kluwer, 1991.
[3] Z. Pawlak and A. Skowron, “Rough sets: Some extensions,” Inf. Sci.,
CACLF, the number of extracted features can be specified in ad- vol. 177, pp. 28–40, 2007.
vance, whereas FVPRS cannot. What is more, we find that the [4] L. A. Zadeh, “Fuzzy sets,” Inf. Control, vol. 8, pp. 338–353, 1965.
computational complexity of FVPRS is O(|U |2 × |A| × |K|), [5] D. Dubois and H. Prade, “Rough fuzzy sets and fuzzy rough sets,” Int. J.
Gen. Syst., vol. 17, pp. 191–208, 1990.
whereas the computational complexity of CACLF is decided by [6] D. Dubois and H. Prade, “Putting rough sets and fuzzy sets together,” in
the optimal problems of finding the suitable parameters for the Intelligent Decision Support, Handbook of Applications and Advances of
linear combination. Here, we compare these two methods on the Rough Sets Theory, R. Slowinski, Ed. Norwell, MA: Kluwer, 1992,
pp. 203–232.
the datasets that have been analyzed in CACLF and summarize [7] M. Quafafou, “α-RST: A generalization of rough set theory,” Inf. Sci.,
them in Table VI. vol. 124, pp. 301–316, 2000.
Table VI shows that under the condition with similar num- [8] W. Z. Wu and W. X. Zhang, “Constructive and axiomatic approaches of
fuzzy approximation operators,” Inf. Sci., vol. 159, pp. 233–254, 2004.
ber of features, CACLF has better classification accuracy than [9] R. Slowinski and D. Vanderpooten, “A generalized definition of rough
FVPRS, whereas if we relax this condition, it is easy to find approximations based on similarity,” IEEE Trans. Knowl. Data Eng.,
that the classification accuracy of FVPRS is significantly bet- vol. 12, no. 2, pp. 331–336, Mar./Apr. 2000.
[10] N. N. Morsi and M. M. Yakout, “Axiomatics for fuzzy rough sets,” Fuzzy [39] A. Mieszkowicz-Rolka and L. Rolka, “Variable precision fuzzy rough
Sets Syst., vol. 100, pp. 327–342, 1998. sets,” Trans. Rough Sets, vol. 1, pp. 144–160, 2004.
[11] A. M. Radzikowska and E. E. Kerre, “A comparative study of fuzzy rough [40] C. Cornelis, M. D. Cock, and A. M. Radzikowska, “Vaguely quantified
sets,” Fuzzy Sets Syst., vol. 126, pp. 137–155, 2002. rough sets,” in Proc. Rough Sets, Fuzzy Sets, Data Mining Granular
[12] J. S. Mi and W. X. Zhang, “An axiomatic characterization of a fuzzy Comput. (Lecture Notes in Artificial Intelligence), vol. 4482, 2007, pp. 87–
generalization of rough sets,” Inf. Sci., vol. 160, pp. 235–249, 2004. 94.
[13] W. Z. Wu, J. S. Mi, and W. X. Zhang, “Generalized fuzzy rough sets,” [41] H. S. Nguyen and A. Skowron, “Boolean reasoning for feature extraction
Inf. Sci., vol. 151, pp. 263–282, 2003. problems,” in Proc. 10th Int. Symp. Methodol. Intell. Syst. (Lecture Notes
[14] D. G. Chen, X. Z. Wang, D. S. Yeung, and E. C. C. Tsang, “Rough approx- in Artificial Intelligence), vol. 1325, Berlin, Germany, 1997, pp. 117–126.
imations on a complete completely distributive lattice with applications to [42] D. Slezak and J. Wroblewski, “Classification algorithms based on linear
generalized rough sets,” Inf. Sci., vol. 176, pp. 1829–1848, 2006. combination of features,” in Proc. 3rd Eur. Conf. Principles Data Mining
[15] D. S. Yeung, D. G. Chen, E. C. C. Tsang, J. W. T. Lee, and X. Z. Wang, Knowl. Discov. (Lecture Notes in Artificial Intelligence), vol. 1704, 1999,
“On the generalization of fuzzy rough sets,” IEEE Trans. Fuzzy Syst., pp. 548–553.
vol. 13, no. 3, pp. 343–361, 2005. [43] S. H. Nguyen and H. S. Nguyen, “Some effective algorithms for rough
[16] E. C. C. Tsang, D. G. Chen, D. S. Yeung, X. Z. Wang, and J. W. T. Lee, set methods,” in Proc. Conf. Int. Process. Manage. Uncertainty Knowl.
“Attributes reduction using fuzzy rough sets,” IEEE Trans. Fuzzy Syst., Based Syst., Granada, Spain, 1996, pp. 1451–1456.
to be published. [44] J. Wroblewski, “Ensembles of classifiers based on approximate reducts,”
[17] D. G. Chen, E. C. C. Tsang, and S. Y. Zhao, “Attributes reduction with Fundam. Inf., vol. 47, pp. 351–360, 2001.
T L –fuzzy rough sets,” in Proc. 2007 IEEE Int. Conf. Syst., Man, Cybern., [45] Q. H. Hu, D. R. Yu, Z. X. Xie, and X. D. Li, “EROS: Ensemble rough
vol. 1, pp. 486–491. subspaces,” Pattern Recognit., vol. 40, pp. 3728–3739, 2007.
[18] R. Jensen and Q. Shen, “Fuzzy-rough attributes reduction with application [46] D. Slezak, “Approximate reducts in decision tables,” in Proc. New Dir.
to web categorization,” Fuzzy Sets Syst., vol. 141, pp. 469–485, 2004. Rough Sets, Data Mining, Granular-Soft Comput. (LNCS), vol. 1711,
[19] Q. Shen and R. Jensen, “Selecting informative features with fuzzy-rough pp. 137–145.
sets and its application for complex systems monitoring,” Pattern Recog- [47] D. Slezak, “Approximate entropy reducts,” Fundam. Inf., vol. 53, pp. 365–
nit., vol. 37, pp. 1351–1363, 2004. 390, 2002.
[20] R. Jensen and Q. Shen, “Semantics-preserving dimensionality reduction: [48] H. S. Nguyen and D. Slezak, “Approximate reducts and association rules-
Rough and fuzzy-rough-based approaches,” IEEE Trans. Knowl. Data corresponding and complexity results,” Proc. RSFDGrC 1999 (Lecture
Eng., vol. 16, no. 12, pp. 1457–1471, Dec. 2004. Notes in Artificial Intelligence), vol. 1711, N. Zhong, A. Skowron, and
[21] R. Jensen and Q. Shen, “Fuzzy-rough sets assisted attribute selection,” S. Ohsuga, Eds., pp. 137–145.
IEEE Trans. Fuzzy Syst., vol. 15, no. 1, pp. 73–89, Feb. 2007. [49] C. Cornelis, G. H. Martı́n, R. Jensen, and D. Slezak, “Feature selection
[22] Q. H. Hu, D. R. Yu, and Z. X. Xie, “Information-preserving hybrid data with fuzzy decision reducts,” in Proc. 3rd Int. Conf. Rough Sets Knowl.
reduction based on fuzzy-rough techniques,” Pattern Recognit. Lett., Technol., 2008, pp. 284–291.
vol. 27, pp. 414–423, 2006. [50] E. C. C. Tsang, S. Y. Zhao, and J. W. T. Lee, “Rule induction based on
[23] B. B. Rajen and M. Gopal, “On fuzzy rough sets approach to feature fuzzy rough sets,” in Proc. 2007 Int. Conf. Mach. Learn. Cybern., vol. 5,
selection,” Pattern Recognit. Lett., vol. 26, pp. 1632–1640, 2005. pp. 3028–3033.
[24] W. Ziarko, “Variable precision rough set model,” J. Comput. Syst. Sci., [51] C. Cornelis and R. Jensen, “A noise-tolerant approach to fuzzy-rough
vol. 46, no. 1, pp. 39–59, 1993. feature selection,” in Proc. 17th Int. Conf. Fuzzy Syst., 2008, pp. 1598–
[25] J. D. Katzberg and W. Ziarko, “Variable precision rough sets with asym- 1065.
metric bounds,” in Rough Sets, Fuzzy Sets and Knowledge Discovery,
W. Ziarko, Ed. Berlin, Germany: Springer-Verlag, 1994, pp. 167–177. Suyun Zhao (S’06) received the Bachelor’s and
[26] J. S. Mi, W. Z. Wu, and W. X. Zhang, “Approaches to knowledge reduction Master’s degrees from the School of Mathematics
based on variable precision rough set model,” Inf. Sci., vol. 159, no. 3–4, and Computer Science, Hebei University, Baoding,
pp. 255–272, 2004. China, in 2001 and 2004, respectively. She is cur-
[27] M. Beynon, “Reducts within the variable precision rough sets model: A rently working toward the Ph.D. degree in the Depart-
further investigation,” Eur. J. Oper. Res., vol. 134, pp. 592–605, 2001. ment of Computing, Hong Kong Polytechnic Univer-
[28] J. M. F. Salido and S. Murakami, “Rough set analysis of a general type sity, Kowloon, Hong Kong.
of fuzzy data using transitive aggregations of fuzzy similarity relations,” Her current research interests include machine
Fuzzy Sets Syst., vol. 139, pp. 635–660, 2003. learning, pattern recognition, fuzzy sets, and rough
[29] G. J. Klir and B. Yuan, Fuzzy Logic: Theory and Applications. Engle- sets.
wood Cliffs, NJ: Prentice-Hall, 1995. Ms. Zhao is a Student Member of the IEEE Sys-
[30] A. Skowron and C. Rauszer, “The discernibility matrices and functions tem, Man, and Cybernetics Society.
in information systems,” in Intelligent Decision Support—-Handbook of
Applications and Advances of the Rough Sets Theory, R. Slowinski, Ed. Eric C. C. Tsang (A’04–M’04) received the B.Sc.
Norwell, MA: Kluwer, 1992, pp. 331–362. degree in computer studies from the City University
[31] H. C. Peng, F. H. Long, and C. Ding, “Feature selection based on mu- of Hong Kong, Kowloon, Hong Kong, in 1990, and
tual information: Criteria of max-dependency, max-relevance and min the Ph.D. degree in computing from the Hong Kong
redundancy,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 27, no. 8, Polytechnic University, Hong Kong, in 1996.
pp. 1226–1238, Aug. 2005. His current research interests include fuzzy
[32] Y. F. Yuan and M. J. Shaw, “Introduction of fuzzy decision tree,” Fuzzy rough sets, fuzzy neural networks, genetic algorithm,
Sets Syst., vol. 69, pp. 125–139, 1995. fuzzy support vector machine, and multiple classifier
[33] J. C. Bezdek and J. O. Harris, “Fuzzy partitions and relations: An axiomatic systems.
basis of clustering,” Fuzzy Sets Syst., vol. 84, pp. 143–153, 1996.
[34] T. Sudkamp, “Similarity, interpolation, and fuzzy rule construction,”
Fuzzy Sets Syst., vol. 58, pp. 73–86, 1993. Degang Chen received the M.S. degree from the
[35] UCI Machine Learning Repository. (2005). [Online]. Available: Northeast Normal University, Changchun, China, in
http://www.ics.uci.edu/∼mlearn/MLRepository.html. 1994, and the Ph.D. degree from Harbin Institute of
[36] D. G. Stork and Y.-T. Elad, Computer Manual in MATLAB to Accompany Technology, Harbin, China, in 2000.
Pattern Classification, 2nd ed. Hoboken, NJ: Wiley, 2004. He was a Postdoctoral Fellow at Xi’an Jiaotong
[37] A. Mieszkowicz-Rolka and L. Rolka, “Variable precision fuzzy rough sets University, Xi’an, China (from 2000 to 2002), and
model in the analysis of process data,” in Proc. Int. Workshop Rough Sets, at Tsinghua University, Beijing, China (from 2002
Fuzzy Sets, Data Mining, Granular Comput. (Lecture Notes in Artificial to 2004). Since 2006, he has been a Professor at
Intelligence), vol. 3641, Regina, SK, Canada, 2005, pp. 354–363. the North China Electric Power University, Beijing,
[38] A. Mieszkowicz-Rolka and L. Rolka, “Remarks on approximation quality China. His current research interests include fuzzy
in variable precision fuzzy rough sets model,” in Proc. Rough Sets Curr. group, fuzzy analysis, rough sets, and support vector
Trends Comput., 2004, pp. 402–411. machine (SVM).
Short Papers
Intermediate Variable Normalization for Gradient Descent the THEN part instead of the IF part of upper layers. However, their
Learning for Hierarchical Fuzzy System method needs many parameters while obtaining less accurate results.
Zeng and Keane [16] investigated the approximation capability and
Di Wang, Xiao-Jun Zeng, and John A. Keane have theoretically shown that hierarchical fuzzy systems can achieve
better approximation accuracy with fewer parameters and rules for a
large class of systems.
Abstract—When applying gradient descent learning methods to hierar-
Gradient descent methods are widely used in parameter learning.
chical fuzzy systems, there is great difficulty in handling the intermediate
variables introduced by the hierarchical structures, as the intermediate Despite the aforementioned research results for hierarchical fuzzy sys-
variables may go outside their definition domain that makes gradient de- tems, there is no effective scheme to handle intermediate variables—a
scent learning invalid. To overcome this difficulty, this paper proposes a feature that does not occur in gradient descent learning for flat fuzzy
learning scheme that integrates a normalization process for intermediate systems. However, in order to make the gradient decent learning ap-
variables into gradient descent learning. This ensures that gradient descent
methods are applicable to, and correctly used for, learning general hier- plicable to, and effective for, hierarchical fuzzy systems, it is nec-
archical fuzzy systems. Benchmark datasets are used to demonstrate the essary and important to handle these intermediate variables properly
validity and advantages of the proposed learning scheme over other existing for two reasons. First, in order to define the membership functions of
methods in terms of better accuracy, better transparency, and fewer fuzzy the upper-layered fuzzy subsystems, it is necessary to know the def-
rules and parameters.
inition domain of intermediate variables which are the outputs of the
Index Terms—Fuzzy systems, gradient descent method, hierarchical lower-layered fuzzy subsystems and the inputs to the upper-layered
fuzzy systems, learning. fuzzy subsystems. Unfortunately, it is impossible to know the domain
of intermediate variables before the final hierarchical fuzzy system is
determined. Second, after the initial definition of membership func-
I. INTRODUCTION
tions in the gradient descent learning process, often the intermediate
Standard fuzzy systems have been widely and successfully applied variables go outside their definition domain, which results in no activa-
in function approximation [1]–[3], system control [4]–[6], classification, and hence, no parameter updating is done for the current training
tion [7], and clustering [8]. However, when fuzzy systems are applied instance. If many training instances cause no parameter updating in this
to more complex and high-dimensional systems, the “curse of dimen- way, then the gradient descent learning method is invalid. Hence, the
sionality” becomes increasingly apparent as a bottleneck to their wider purpose of this research is to address this gap by introducing a normal-
application. To overcome this, hierarchical fuzzy systems were pro- ization process to handle intermediate variables in the gradient descent
posed in the early 1990s by Raju and Zhou [9] and have attracted much learning method to make it applicable to, and effective for, hierarchical
attention in recent years. In hierarchical fuzzy systems, instead of us- fuzzy systems.
ing a standard (flat) high-dimensional fuzzy system, a number of lower In this paper, a gradient descent learning scheme integrated with the
dimensional fuzzy subsystems are linked in a hierarchical manner. normalization of intermediate variables for general hierarchical fuzzy
The main topics of hierarchical fuzzy systems research have been systems is proposed to overcome the particular difficulty in determining
construction, learning, and analysis (see the survey by Torra [10]). the definition domain of intermediate variables. Further, it is shown that,
Wang and his colleagues [5], [11] proposed a kind of hierarchical although some extra errors may be introduced by the normalization
fuzzy system with one fuzzy subsystem in each layer that has one process for intermediate variables, such errors can be corrected in the
original input and one input from the output of the lower subsystem. following learning iterations. The proposed learning scheme leads to
They [5], [11] applied the gradient descent method for learning and advantages such as better accuracy, better transparency, and fewer rules
analyzed the relative importance of input variables as a criterion for the and parameters. These advantages are validated both theoretically in
hierarchical structure construction. Although their proposed hierarchi- the Appendix and experimentally through simulation in Section III.
cal fuzzy systems managed to decrease the exponential growth of fuzzy This paper is organized as follows. The learning algorithm with the
rules, the exponential growth of parameters remains inherent. Chung normalization of intermediate variable is proposed and validated in
and Duan [12] discussed how to design a hierarchical structure based Section II. In Section III, the proposed algorithm is applied to bench-
on the correlated or coupled relationship between input variables and mark problems to demonstrate its advantages over existing methods for
showed the applicability of their approach by simulation. Campello and hierarchical fuzzy systems. Then, conclusions are given in Section IV.
Amaral [13] developed a method to construct hierarchical fuzzy sys-
tems by using Gaussian membership functions. Joo and Lee [14], [15] II. GRADIENT DESCENT LEARNING WITH INTERMEDIATE
proposed a scheme to construct another kind of general hierarchical VARIABLE NORMALIZATION
fuzzy system, in which the outputs of lower layers are used only in
A. Overview of Gradient Descent Learning for General Hierarchical
Fuzzy Systems
Manuscript received April 27, 2007; revised April 4, 2008 and September 29,
In a general multiple-input and single-output (MISO) hierarchical
2008; accepted December 16, 2008. First published February 10, 2009; current
version published April 1, 2009. This work is supported by the U.K. Engineering fuzzy system (see Fig. 1), there may be multiple layers and multi-
and Physical Sciences Research Council (EPSRC) under Grant EP/C513355/1. ple fuzzy subsystems in each layer. Outputs of lower-layered fuzzy
D. Wang is with ThinkAnalytics Ltd., Glasgow G3 7QF, U.K. (e-mail: subsystems form as inputs to their neighboring upper-layered fuzzy
dwang@thinkanalytics.com). subsystems. These are termed as intermediate variables. The inputs to
X.-J. Zeng and J. A. Keane are with the School of Computer Sci-
ence, University of Manchester, Manchester M13 9PL, U.K. (e-mail: the lowest layer are all original input variables. The inputs to the lth (l
x.zeng@manchester.ac.uk; john.keane@manchester.ac.uk). >1) layer are the combination of its lower-layered (the (l−1)th layer)
Digital Object Identifier 10.1109/TFUZZ.2009.2014940 outputs and some (or none) of the original input variables.
1063-6706/$25.00 © 2009 IEEE
Fig. 2. Example of triangular membership function.
If
Fig. 1. General structure for hierarchical fuzzy systems.
Pl,1,p
j
Ul , p = ul ,kp , k (ŷl −1 , p , k ) (5)
k=1
In general, a fuzzy subsystem Fl , p (the pth fuzzy subsystem in the and

lth layer) is represented in the form as

Q l −1 , p
ŷl , p = fl , p (ŷl −1 , p , 1 , ŷl −1 , p , 2 , . . . , ŷl −1 , p , P l −1 , p , Vl , p = vli,kp , k (xl −1 , p , k ) (6)

k=1
xl −1 , p , 1 , . . . , xl −1 , p , Q l −1 , p ) (1)
then (4) can be represented as (7) as follows. For a simpler representa-
where ŷl , p is the output of fuzzy subsystem Fl , p , ŷl −1 , p , j is the output tion, we use Ul , p and Vl , p in the remainder of this paper
from fuzzy subsystem Fl −1 , p , j (the jth fuzzy subsystem of the (l−1)th
layer) to Fl , p , where Pl −1 , p is the total number of outputs from the  
(l−1)th layer to Fl , p , xl −1 , p , j is the jth original input variable to Fl , p ,
and Ql −1 , p is the total number of original input variables to Fl , p .
 U Vl , p 
ŷl , p =  l,p 
Assuming L is the total number of layers (there is only one fuzzy Ul , p Vl , p
j 1 j 2 ···j P i i ···i Q
l −1 , p 1 2 l −1 , p j 1 j 2 ···j P i i ···i Q
subsystem in the top layer, the Lth layer, for an MISO system), the l −1 , p 1 2 l −1 , p
final output (the output of the Lth layer) is represented by j 1 j 2 ···j P i i ···i Q
l −1 , p 1 2 l −1 , p
× yl , p . (7)
ŷ = ŷL , 1 = fL , 1 (ŷL −1 , 1 , 1 , ŷL −1 , 1 , 2 , . . . , ŷL −1 , 1 , P L −1 , 1 ,
xL −1 , 1 , 1 , . . . , xL −1 , 1 , Q L −1 , 1 ). (2)
In the aforementioned discussion, µjl ,kp , k (ŷl −1 , p , k ) and
i
vl ,kp , k (xl −1 , p , k )
in (5) and (6) are the membership functions,
Uniquely, for the first (lowest) layer, the output of the pth fuzzy
which can be defined by many types of functions, such as Gaussian,
subsystem F1 , p is
triangular, trapezoid, or bell-shape. Specially, triangular functions are
chosen as the membership functions in this paper. Furthermore, the
ŷ1 , p = f1 , p (x0 , p , 1 , . . . , x0 , p , Q 0 , p ). (3)
edge of one triangular membership function is chosen to intersect to the
middle point of its neighboring triangular membership functions (as
In the proposed algorithm, for each fuzzy subsystem Fl , p , a
shown in Fig. 2). Detailed reasons for the use of triangle membership
commonly used defuzzifier, center-average defuzzifier, is applied
functions are given in Remark 2.1 at the end of this section. Based
based on Mamdani reasoning [17] (4) as shown at the bottom j
on the analysis before, the membership functions µl ,kp , k (ŷl −1 , p , k )
of this page, where j1 j2 · · · jP l −1 , p i1 i2 · · · iQ l −1 , p is the index of
j are designed as µjl ,kp , k (ŷl −1 , p , k ) : Γl −1 , p , k ⊂ R → Γl , p ⊂ R,
rules for the Fl , p , µl ,kp , k (ŷl −1 , p , k ) is the membership function for
ik
where Γl −1 , p , k = [αl −1 , p , k , βl −1 , p , k ]. If ŷl −1 , p , k is evenly par-
ŷl −1 , p , k , vl , p , k (xl −1 , p , k ) is the membership function for xl −1 , p , k , and titioned with Nl −1 , p , k , then Γl −1 , p , k = [αl −1 , p , k , βl −1 , p , k ] =
j 1 j 2 ···j P
l −1 , p
i 1 i 2 ···i Q
l −1 , p

N l , 1 , p , k −1 i jk + 1 jk
[el −1 , p , k , eil −1 , p , k ], and el −1 , p , k − el −1 , p , k = (αl −1 , p , k
+1
yl , p is the parameters to be learned during gradi- i= 0
ent descent learning. − βl −1 , p , k )/Nl −1 , p , k .
ŷl,p =fl,p (ŷl−1,p,1 , ŷl−1,p,2 , . . . , ŷl−1,p,P l −1 , p , xl−1,p,1 , . . . , xl−1,p,Q l −1 , p )

 P l −1 , p j k Q l −1 , p i k 
µ (ŷl−1,p,k ) v (x l−1,p,k ) j j ···j i i ···i
=  k =1 l,p,k
P l −1 , p j k
k =1 l,p,k
Q l −1 , p i k y 1 2 P l −1 , p 1 2 Q l −1 , p
l,p
j 1 j 2 ···j P i 1 i 2 ···i Q
l −1 , p
j 1 j 2 ···j P
l −1 , p
i 1 i 2 ···i Q
l −1 , p
k =1 µl,p,k (ŷl−1,p,k )
l −1 , p
k =1 vl,p,k (xl−1,p,k )
(4)
j
The triangular membership function µl ,kp , k (ŷl −1 , p , k ) for ŷl −1 , p , k is functions have the decomposition property [18], i.e., by properly di-
defined as follows: viding the input space into subinput spaces, a general fuzzy system
is decomposed into several fuzzy subsystems that are the simplest
j
µl ,kp , k (ŷl −1 , p , k ) fuzzy systems in the subinput spaces. This feature is particular useful
 k −1
in advanced techniques for fuzzy control design and analysis such as
 ŷl −1 , p , k − elj−1

 ,p ,k j k −1 jk piecewise Lyapounov methods [20].
 , el −1 , p , k ≤ ŷl −1 , p , k < el −1 , p , k
 jk
 l −1 , p , k
e − e
j k −1
In the remainder of this section, a gradient descent learning scheme
 l −1 , p , k
jk + 1 integrated with normalization of intermediate variables for general hi-
= e − ŷ
l −1 , p , k (8)


l −1 , p , i
, ejl −1
k jk + 1
, p , k ≤ ŷl −1 , p , k < el −1 , p , k
erarchical fuzzy systems is proposed. For this purpose, it is assumed

 jk + 1
−
jk


e l −1 , p , k e l −1 , p , k that the hierarchical structure has been determined, and the discussion
 will focus on the parameter learning process for general hierarchical
0, otherwise.
fuzzy systems.
i
In a similar way, vl ,kp , k (xl −1 , p , k ) is defined as
i
vl ,kp , k (xl −1 , p , k ) B. Normalization for the Intermediate Variables
 i k −1 The definition domain for each variable must be known in order
 xl −1 , p , k − el −1

 ,p ,k i k −1 ik to define triangular membership functions given in (8) and (9). The
 , el −1 , p , k ≤ xl −1 , p , k < el −1 , p , k

 e
ik
− e
i k −1
definition domain of the original input variables can be obtained directly
 l −1 , p , k l −1 , p , k
from the training data. However, the possible values of the intermediate
= ei k + 1 − x l −1 , p , k (9)


l −1 , p , i
, k
el −1
i
, p , k ≤ xl −1 , p , k < el −1 , p , k
k i +1 variables are unknown. As a result, it is impossible to define the domain

 ik + 1
− ik
 l −1 , p , k

e e l −1 , p , k of [αl −1 , p , k , βl −1 , p , k ] for intermediate variable ŷl −1 , p , k in hierarchical
 fuzzy systems before training.
0, otherwise.
In Wang’s research [11], the domain for the intermediate variable
For the aforementioned membership functions, only 2n rules are (y1 in his simulation example) is defined as [0, 1] during training.
activated each time, where n is the number of input variables. As However, the value of the intermediate variable y1 is, in fact, not
shown by [3, Th. 2] always between zero and one during gradient descent learning. On the
contrary, y1 might be outside of the range [0, 1]. A value of y1 outside
Ul , p Vl , p = 1 (10) of the range [0, 1] will not activate any membership function. In such
j 1 j 2 ···j P
l −1 , p
i 1 i 2 ···i Q
l −1 , p cases, the parameters (dl i j and ȳ1p q [11]) will not be updated during the
gradient descent learning process. As a result, such gradient descent
then (7) can be simplified as learning is not valid for hierarchical fuzzy systems with triangular
j 1 j 2 ···j P i i ···i Q membership functions. This might be the reason why Wang’s system
l −1 , p 1 2 l −1 , p
ŷl , p = Ul , p Vl , p yl , p . in [11] is less accurate than ours, despite using more parameters. This
j 1 j 2 ···j P i i ···i Q is illustrated by simulation comparison in Section III.
l −1 , p 1 2 l −1 , p
(11) To bridge this gap, a normalization process is proposed to overcome
Remark 2.1: There are several important reasons which motivate this particular difficulty in determining the definition domain of in-
the use of triangular membership functions in this paper. First, fuzzy termediate variables. Our research has found that there is a valuable
systems with triangular membership functions are more transparent property related to intermediate variables: if a hierarchical fuzzy sys-
and interpretable than fuzzy systems with other types of membership tem can achieve a desired approximation accuracy within a definition
functions. Second, fuzzy systems with triangular membership func- domain [αl −1 , p , k , βl −1 , p , k ] for an intermediate variable ŷl −1 , p , k , then
tions lead to a much simpler gradient decent learning algorithm for by choosing its definition domain to be any interval [αl −1 , p , k , βl −1 , p , k ],
two reasons: one is the much simpler mathematical expression possi- the same approximation accuracy can be achieved by an associated hi-
ble, as shown in (11) where the complicated denominator terms in (7) erarchical fuzzy system. In other words, the approximation accuracy of
disappear based on (10); the other is that only 2n rules are activated hierarchical fuzzy systems is independent of the choice of the defini-
for each training sample that leads to much simpler calculation in each tion domains for intermediate variables. As a result, we can always find
iteration of the training process. This is significantly different to fuzzy the optimal hierarchical fuzzy system by restricting the intermediate
systems with other types of membership functions, especially Gaus- variables to any definition domain of [αl −1 , p , k , βl −1 , p , k ]. Particularly,
sian membership functions where all rules are active for each training in our work, optimal hierarchical fuzzy systems are constructed by
sample. Due to these two reasons, the corresponding gradient decent restricting the definition domain of all intermediate variables to [0, 1].
algorithm takes much less time to learn and is much quicker to con- A normalization process is introduced to restrict the definition domain
verge. Such simplification and improved performance are especially of intermediate variables to [0, 1], which is integrated into the gradi-
important for hierarchical fuzzy systems that have a much more com- ent descent learning algorithm. We now give the following theorem
plex system structure than (single layered) flat fuzzy systems and are (Theorem 2.1) that provides the theoretical justification of this normal-
thus much more complicated to learn. Third, fuzzy systems with trian- ization process of intermediate variables to [0, 1].
gular membership functions are one of the most widely used classes Theorem 2.1 (Approximation Accuracy Preservation Prop-
of fuzzy systems and have shown better performance than other types erty of Normalization): Define, in a hierarchical fuzzy sys-
of membership functions in many applications (for example, [19]). In tem H(X), the pth fuzzy subsystem in the lth layer as
particular, a valuable property of fuzzy systems with triangular mem- Fl , p :ŷl , p = fl , p (Yl −1 , p , Xl −1 , p ), where ŷl , p is the output, Xl −1 , p =
bership functions is that they can reproduce (i.e., represent with no (xl −1 , p , 1 , xl −1 , p , 2 , . . . , xl −1 , p , Q l −1 , p ) are original inputs to Fl , p ,
error) many important and widely used functions and systems such as and Yl −1 , p = (ŷl −1 , p , 1 , ŷl −1 , p , 2 , . . . , ŷl −1 , p , P l −1 , p ) are outputs of the
linear functions, multilinear functions, and piecewise linear functions, connected fuzzy subsystems in the (l−1)th layer and ŷl , p , j ∈
as shown in [18]. Fourth, fuzzy systems with triangular membership [αl −1 , p , i , βl −1 , p , i ] (i = 1, . . . , Ql , p ). Under this definition, for a
given function G(X), if there exists a hierarchical fuzzy system H(X)

defined by (1)–(4)
H(X) = fL , 1 (fL −1 , 1 {. . . [f2 , 1 (f1 , 1 (X0 , 1 )
, . . . , f1 , P 1 (X0 , P 1 ), X1 , 1 ), . . .] . . .}
, . . . , fL −1 , P L −1 {. . .} , XL −1 , 1 ) (12)
such that, for any given input vector X, |G(X) − H(X)| ≤ ε (where
ε is a small constant representing the approximation error), then there
exists an associated hierarchical fuzzy system H (X) defined by (1)–
Fig. 3. Illustration of two neighboring fuzzy subsystems.
(4), where ŷl , p , j ∈ [0, 1], (l < L), which has the same hierarchical
structure as H(X) and is given by
H (X) = fL , 1 (fL −1 , 1 {. . . [f2 , 1 (f1 , 1 (X0 , 1 ) where ep (k) is the error of fuzzy subsystem p, which is propagated
from its neighboring upper-layered fuzzy subsystem q, and eq (k) is
, . . . , f1 , P 1 (X0 , P 1 ), X1 , 1 ), . . .] . . .} the propagated error of fuzzy subsystem q.
, . . . , fL −1 , P L −1 {. . .} , XL −1 , 1 ) (13) Similar to (5) and (6), if we define
such that H(X) = H (X) and |G(X) − H (X)| ≤ ε.

Pq
The proof of Theorem 2.1 can be found in Appendix A. Uq (k) = ujq i, i (k) (16)
Remark 2.2: As shown in Theorem 2.1, for a given hierarchical i= 1
fuzzy system H(X) with the approximation error ε, there exists an

Qq
associated hierarchical fuzzy system H (X)with its all intermediate Vq (k) = νqj ,i i (xp , i ) (17)
variables yl , p = fl, p ∈ [0, 1](l < L) that can achieve the same accu- i= 1
racy as H(X). This is why the aforementioned theorem is called the
approximation accuracy preservation property of normalization. Fur- then,
ther, by the same proof, it can be shown that the theorem still holds if
[0, 1] is replaced by any interval. For simplicity, in our research, the ∂ ŷq (k) j 1 j 2 ···j P q , p i 1 i 2 ··· i Q q , p ∂(Uq (k)Vq (k))
= yq
definition domain of all intermediate variables is restricted to [0, 1]. ∂ ŷp (k) ∂ ŷp (k)
j 1 j 2 ···j P q , p i 1 i 2 ···i Q q , p
(18)
C. Learning Based on Gradient Descent Algorithm where
j
In this section, the formula of gradient descent learning for hier- ∂(Uq (k)Vq (k) Uq (k)Vq (k) ∂(µq p, p (k))
= j
× . (19)
archical fuzzy systems with the integrated normalization process is ∂ ŷp (k) µq p, p (k) ∂ ŷp (k)
given. From (5), we have
In gradient descent learning, the error of the final output is propa-  j j
gated back from upper layers to lower ones. The parameter updating
j

 1, epp −1 (k) ≤ ŷp (k) < epp (k)
of the lower layers is based on the errors propagated back from their ∂(µq p, p (k)) j j
= −1, epp (k) ≤ ŷp (k) < epp + 1 (k) (20)
upper layers. The objective is to minimize total error. ∂ ŷp (k) 

The error between the actual output y(k) and the model output ŷL (k) 0, otherwise.
at time k is defined as
Hence, (21) shown at the bottom of this page.
eL (k) = ŷL (k) − y(k). (14) The formula to update the parameters yp
j 1 j 2 ···j P q , p i 1 i 2 ···i Q q , p
(k) in
For simplicity, consider two connected fuzzy subsystems in two gradient descent learning is
consequent layers, fuzzy subsystem q and fuzzy subsystem p, where q j 1 j 2 ···j P q , p i 1 i 2 ···i Q q , p
is the neighboring upper-layered fuzzy subsystem of fuzzy subsystem yp (k + 1)
p (as shown in Fig. 3). j 1 j 2 ···j P q , p i 1 i 2 ···i Q q , p
Further, for simplifying the expressions, in this section, the variables = yp (k) − η × Uq (k)Vq (k) × ep (22)
in discussion and equations are omitted. For example, we represent
j j where η is the learning rate.
µq p, p (ŷq , p , j , k) by µq p, p (k), where k means “at time k.”
As discussed, we aim to assign the definition domain for the inter-
Then, the error propagated back from fuzzy subsystem q to p is
mediate variables as [0, 1]. However, the outputs of a fuzzy subsystem
defined as j 1 j 2 ···j n q
∂ ŷq (k) ŷl , p computed by (11) are dependent on the parameters yl , p and
ep (k) = eq (k) × (15)
∂ ŷp (k) cannot always be guaranteed to be between zero and one. To force the
 j ,j ···j
1 2 P q , p i 1 i 2 ···i Q q , p

 y q (k)U q (k)V q (k) j j

 eq (k) × , if epp −1 (k) ≤ ŷp (k) < epp (k)

 µ
jp
(k)
 j 1 j 2 ···j P q , p i 1 i 2 ···i Q q , p
j j ···j
q ,p

ep (k) = 1 2 P q , p i 1 i 2 ···i Q q , p (21)
 y q (k)U q (k)V q (k) j j

 eq (k) × , if epp (k) ≤ ŷp (k) < epp + 1 (k)

 jp
µq ,p (k)

 j 1 j 2 ···j P q , p i 1 i 2 ···i Q q , p
0, otherwise.
intermediate variables to be in the range [0, 1], normalization has to be TABLE I

applied by using (23), shown at the bottom of this page. COMPARISON WITH WANG’S METHOD [11]
After this normalization, all ŷl , p are always between zero and one.
This can be formally expressed in Theorem 3.1.
Theorem 3.1: After the normalization process by (23), for any fuzzy
subsystem Fl , p (i.e., the pth fuzzy subsystem of the lth layer), all the
j 1 j 2 ···j P q , p i 1 i 2 ···i Q q , p
THEN part yl , p (say, the j1 j2 · · · jP q , p i1 i2 · · · ith
Qq ,p
j 1 j 2 ···j P i 1 i 2 ···i Q
rule of Fl , p ) is between zero and one (i.e., yl , p q ,p q ,p
∈
[0, 1]). Thus, the output of Fl , p , ŷl , p , is also between zero and one (i.e.,
ŷl , p ∈ [0, 1]).
The proof of Theorem 3.1 can be found in Appendix B.
However, the normalization process, according to (23), introduces
extra error; hence, the parameters may no longer be optimal after nor-
malization. The error introduced by this normalization process can be
efficiently corrected in the following gradient descent learning itera-
tions. This is shown by the following theorem.
Theorem 3.2: In the proposed learning algorithm, the extra error
TABLE II
introduced in the normalization process [according to (23)] can be COMPARISON WITH JOO’S WORK 2002 (300 CASES)
eliminated by its following gradient descent training process.
The proof of Theorem 3.2 is given in Appendix C.
III. SIMULATION AND COMPARISON

In this section, we illustrate the advantages of our proposed method
over other methods for hierarchical fuzzy system learning using two
simulation examples.
Simulation 1: This is a function approximation problem referred by
Wang [11]
1
g(x1 , x2 , x3 ) =
1 + sin (πx1 ) + sin2 (πx2 ) + sin2 (πx3 )
2
on U = [0, 1]3 .
For comparison, 63 input–output pairs (xi0j k ; g(xi0j k )), where

xi0j k= [0.2(i − 1), 0.2(j − 1), 0.2(k − 1)] for i, j, k = 1, 2, . . . , 6,
are created for training [11], and 63 uniformly distributed input–output proposed method is based on Mamdani reasoning and the number of pa-
pairs are used for testing. Wang used six equally spaced triangular rameters is equal to the number of fuzzy rules, i.e., only 62 + 62 = 72
membership functions over [0, 1] on both the inputs x1 , x2 , x3 , and the parameters are needed in our proposed scheme for the same hierarchi-
intermediate variable. cal structure and partitions. Therefore, far fewer parameters are used in
In Wang’s method, there is one fuzzy subsystem in each layer, and our scheme than Wang’s, and, in turn, much better accuracy is obtained,
there are two inputs to each fuzzy subsystem created by using a grid as shown in Table I, where it can be seen that by using 72 parameters,
partition. Hence, there are (n−1) layers altogether, and N 2 fuzzy rules we obtain an accuracy of average percentage error (APE) = 0.44067%
for each layer, where N is the partition for each variable (both the for the training data, and an accuracy of APE = 2.9721% for the testing
original and intermediate variables) and n is the input dimension (i.e., data in comparison with 10% by Wang’s method [11]. More results in
the total number of the original input variables). In Wang’s scheme, terms of APE by using different number of parameters and fuzzy rules
the conclusion part for lowest layer is a constant, i.e., the number of in our scheme are shown in Table I. All these results show that our
parameters for the lowest layer is equal to the number of fuzzy rules proposed scheme constructs a hierarchical fuzzy system with far fewer
N 2 . For the other layers, the conclusion part of a fuzzy rule is an parameters and better accuracy than Wang’s method.
(N −1) order polynomial (hence, N parameters for each fuzzy rule) of Joo and Lee [14], [15] have proposed a hierarchical fuzzy system
its lower layer outputs. Therefore, the total number of parameters used [14], where the outputs of the previous layer are not used in the IF
in Wang’s method is (n − 2) ∗ N 3 + N 2 . parts, but only the THEN parts of the fuzzy rules of the current layer.
In this simulation, N = 6 and n = 3, so the total number of param- Using Simulation 2 (used in [14] and [15]), we illustrate the better
eters used is 63 + 62 = 252 for Wang’s method. On the contrary, our performance of our proposed algorithm over Joo and Lee’s approach.
j 1 j 2 ···j P q , p i 1 i 2 ···i Q q , p j 1 j 2 ···j P q , p i 1 i 2 ···i Q q , p

yl,p − min (yl,p )
j 1 j 2 ···j P q , p i 1 i 2 ···i Q q , p j 1 j 2 ...j n q
yl,p = j 1 j 2 ···j P q , p i 1 i 2 ···i Q q , p j 1 j 2 ···j P q , p i 1 i 2 ···i Q q , p
max (yl,p )− min (yl,p )
j 1 j 2 ···j n q j 1 j 2 ···j n q
(23)
TABLE III Nonetheless, there remain some related and open problems for learn-
COMPARISON WITH JOO’S WORK 2005 (3000 CASES) ing hierarchical fuzzy systems, such as how to identify and learn their
associated hierarchical structure. Developments in these areas should
result in wider application of hierarchical fuzzy systems and, in turn,
help to extend fuzzy systems to successfully solve more complicated
and high-dimensional problems.
APPENDIX
A. Proof of Theorem 2.1
To prove Theorem 2.1, a lemma is proved first as follows.
Lemma A.1: Let F (X) = f1 [f0 , 1 (X1 ), f0 , 2 (X2 ), . . . , f0 , p
(Xp ), X0 ] be a function defined on X = [X1 , X2 , . . . ,
X p , X 0 ] ∈ U = U1 × U2 × · · · × Up × U0 and ŷi = f0 , i (Xi )
∈ [αi , βi ] for Xi ∈ Ui with αi = minX i ∈U i f0 , i (Xi ) and
βi = maxX i ∈U i f0 , i (Xi ) (i = 1, 2, . . . , p). Then, there ex-
ists an associated function F (X) = f1 [f0 , 1 (X1 ), f0 , 2

(X2 ), . . . , f0 , p (Xp ), X0 ] defined on U = U1 × U2 × · · · × Up × U0
Simulation 2: This simulation is to stabilize a ball and beam control
such that F (X) = F (X) for all X ∈ U and ŷi = f0 , i (Xi ) ∈ [0, 1]
system. The control law obtained is as follows:
for Xi ∈ Ui with minX i ∈U i f0 , i (Xi ) = 0 and maxX i ∈U i f0 , i (Xi ) =
1 (i = 1, 2, . . . , p).
4BGx4 cos x3 +6BG sin x3 −4x2 −x1 −BGx24 sin x3
u ∗ (x)= Proof: Define ŷi = f0 , i (Xi ) = (f0 , i (Xi ) − αi )/(βi − αi ) (i =
−BG cos x3
1, 2, . . . , p) for Xi ∈ Ui , then ŷi = f0 , i (Xi ) ∈ [0, 1] for Xi ∈
Ui with minX i ∈U i f0 , i (Xi ) = 0 and maxX i ∈U i f0 , i (Xi ) = 1 (i =
with B and G being 0.7143 and 9.81, respectively. So, the task of this
1, 2, . . . , p).
simulation is to find a hierarchical fuzzy system to approximate the
Now, define
aforementioned control function.
The simulation results are compared with Joo and Lee’s method [14],
f1 (ŷ1 , ŷ2 , . . . , ŷp , X0 ) = f1 [(β1 − α1 )ŷ1 + α1 , (β2 − α2 )ŷ2
and Table III [15] in terms of Ju, which
[15], as shown in Table II [14]
N
is defined as Ju = (1/N ) i= 1
(u∗k − ûk )2 , where u∗k is the kth + α2 , . . . , (βp − αp )ŷp + αp , X0 ]
objective output and ûk is the kth model output. Our proposed scheme
obtains better accuracy of Ju = 0.0167 while also using fewer fuzzy for Y = [ŷ1 , ŷ2 , . . . , ŷp ] ∈ [0, 1] × [0, 1] × · · · × [0, 1], then for all
rules and fewer parameters (12 fuzzy rules and 12 parameters) than Joo X = [X1 , X2 , . . . , Xp , X0 ] ∈ U = U1 × U2 × · · · × Up × U0 ,
and Lee’s method with an accuracy of Ju = 0.02 (27 fuzzy rules and 45
parameters). When more fuzzy rules are used in our proposed scheme, F (X) = f1 [f0 , 1 (X1 ), f0 , 2 (X2 ), . . . , f0 , p (Xp ), X0 ]
better accuracy is obtained, e.g., 54 fuzzy rules and parameters result in
an accuracy of Ju = 0.0035. So, this shows that our proposed scheme = f1 [(β1 − α1 )f0 , 1 (X1 )
has better performance than Joo and Lee’s method, in terms of both + α1 , (β2 − α2 )f0 , 2 (X2 )
accuracy and the number of parameters.
+ α2 , . . . , (βp − αp )f0 , p (Xp ) + αp , X0 ]
IV. CONCLUSION

f0 , 1 (X1 ) − α1
= f1 (β1 − α1 )
Hierarchical fuzzy systems have been shown by a number of papers β1 − α 1
to be an effective approach to overcoming the “curse of dimensional-
ity” in flat grid-based fuzzy systems. When applying gradient descent f0 , 2 (X2 ) − α2
+ α1 , (β2 − α2 )
methods to learn hierarchical fuzzy systems, there is a great difficulty β2 − α 2
in handling the intermediate variables introduced by the hierarchical
structures, as the intermediate variables often go outside their defini- f0 , p (Xp ) − αp
+ α2 , . . . , (βp − αp ) + αp , X0
tion domain that makes gradient descent learning invalid. In order to βp − α p
overcome this difficulty, this paper has proposed a gradient descent
learning scheme integrated with the normalization process of interme- = f1 [f0 , 1 (X1 ), f0 , 2 (X2 ), . . . , f0 , p (Xp ), X0 ] = F (X).
diate variables for general hierarchical fuzzy systems.
The main contribution of this paper is the introduction of a nor- This ends the proof.
malization process for the intermediate variables during the learning Proof of Theorem 2.1: Based on Lemma 1, which is the case for
iterations, to make the gradient descent learning method applicable to, a two-level hierarchical fuzzy system and mathematical induction,
and effective for, general hierarchical fuzzy systems. We theoretically Theorem 2.1 can be obtained immediately. Here, we omit the detailed
proved the validity of this normalization process and verified that the steps to apply mathematical induction due to the space limitation.
proposed normalization process does not damage the optimization of
the final solution. The advantages of the proposed learning scheme
B. Proof of Theorem 3.1
over existing methods [5], [11], [14], [15] are shown by benchmark
simulations in terms of better accuracy, better transparency, and fewer Proof of Theorem 3.1: We prove that the conclusion part for any
fuzzy rules and parameters. fuzzy subsystem ŷl , p = fl , p (Yl −1 , p , Xl −1 , p ) is between zero and one
(i.e., ŷl , p ∈ [0, 1]). First, we have from (4) that original variable. There are m intermediate variables and n0 original
input variables for the given subsystem.

U (ŷl −1 , p )V (xl −1 , p )
Now, define the output for the rth rule for the given fuzzy subsystem

y l,p = as or
U (ŷl −1 , p )V (xl −1 , p )
m
n0
j 1 j 2 ···j P i i ···i Q
l −1 , p 1 2 l −1 , p j 1 j 2 ···j n l , p
or = y r × µr (yp ) × µr (xq ). (C2)
p=1 q=1
j 1 j 2 ···j P i i ···i Q
l −1 , p 1 2 l −1 , p
× yl , p Then,

R
(B1) ŷ = o = or . (C3)
r=1
where U (ŷl −1 , p ) and V (xl −1 , p ) are defined as The overall error of the given fuzzy subsystem caused by normal-
ization of the intermediate variables yp (where p = 1,. . ., m) is

P l −1 , p

U (ŷl −1 , p ) =
j
µl ,kp , k (ŷl −1 , p , k ) (B2)
R
m
n0
k=1 e= y ×
r
µ (ȳp ) ×
r r
µ (xq )
r=1 p=1 q=1

Q l −1 , p
i

R
m
n0
V (xl −1 , p ) = vl ,kp , k (xl −1 , p , k ). (B3) − yr × µr (yp ) × µr (xq )
k=1 r=1 p=1 q=1
For a simple representation, we define a normalization factor as

R
n0

m
m
= y ×
r
µ (xq ) ×
r
µ (ȳp ) −
r r
µ (yp )
N (ŷl −1 , p , xl −1 , p ) = U (ŷl −1 , p )V (xl −1 , p ). (B4) r=1 q=1 p=1 p=1
j 1 j 2 ···j n l , p
(C4)
Then, the following two inequalities are obtained as shown (B5) and
where
(B6) at the bottom of this page. yp − minp
This immediately implies ŷl , p ∈ [0, 1]. ȳp = . (C5)
maxp − minp
If we define the error caused by the rth rule by the normalization of
C. Proof of Theorem 3.2
the intermediate variables as er
Proof of Theorem 3.2: Consider the hierarchical fuzzy system given
in (4) or (11). To simplify the expression, in the following, we omit the
n0

m
m
index of layers and the index of fuzzy subsystems in that layer. Then, e =y ×
r r
µ (xq ) ×
r
µ (ȳp ) −
r r
µ (yp ) . (C6)
for a given training instance, the output of any given fuzzy subsystem q=1 p=1 p=1
can be presented as
Then,

R
R m
n0
ŷ = o = yr × µr (yp ) × µr (xq ) (C1) e= er . (C7)

r=1
r=1 p=1 q=1
As (C6) and (C7) show that the overall error e is the sum of the error
where yp (p = 1, 2, . . . , m) are the associated intermediate variables, caused by each fuzzy rule, this implies that, to prove that the overall
xq (q = 1, 2, . . . , n0 ) are the original input variables, ŷ is the output, error caused by the normalization process is eliminated, it is sufficient
y r is the THEN part of the rth rule, µr (yp ) is the membership of to prove the error caused by each fuzzy rule is eliminated, which is
the pth intermediate variable, and µr (xq ) is membership of the qth given later.

U (ŷl−1,p )V (xl−1,p ) j 1 j 2 ···j P l −1 , p i 1 i 2 ···i Q l −1 , p
ŷl,p = yl,p
j 1 j 2 ···j P l −1 , p i 1 i 2 ···i Q l −1 , p
N (ŷl−1,p , xl−1,p )


≤ max yl,p =1 (B5)
N (ŷl−1,p , xl−1,p ) j 1 j 2 ···j n l , p
j 1 j 2 ···j P l −1 , p i 1 i 2 ···i Q l −1 , p

ŷl,p = yl,p
j 1 j 2 ···j P l −1 , p i 1 i 2 ···i Q l −1 , p
N (ŷl−1,p , xl−1,p )


≥ min yl,p =0 (B6)
N (ŷl−1,p , xl−1,p ) j 1 j 2 ···j n l , p
j 1 j 2 ···j P l −1 , p i 1 i 2 ···i Q l −1 , p
m m m
Consider the error caused by the rth rule as given in (C6) and define 3) If ( i = 1 µr (yp ) > 0 and i= 1
µr (ȳp ) = 0) or ( i = 1
m m
m r i = 1 µ (ȳp )m−
r r r
µ
m r(y ) > 0, µ (ȳ ) > 0, and

m
m p i= 1 p
∆= µr (ȳp ) − µr (yp ). (C8) i= 1

µ (yp ) < 0), then ∆ = i = 1 µ (ȳp ) − i = 1
i= 1 i= 1
µr (yp ) < 0. Then, there are two possible cases that are
considered next.
Then, (C6) can be rewritten as a) If y r (k) > 0, then from (C9), er (k) < 0, which
m n

n0

m
m implies η × p = 1 µ (ȳp ) × q =0 1 µr (xq ) × er (k) < 0
r
er = y r × µr (xq ) × µr (ȳp ) − µr (yp ) and y r (k + 1) > y r (k) from (C11). Given a proper small
q=1 p=1 p=1 value of the learning rate η in (C11), then y r (k)and y r (k +
1) have the same sign, i.e., y r (k + 1) > 0. Then, based

n0
n
= yr × µr (xq ) × ∆ (C9) on (C9), er (k + 1) = y r (k + 1) × q =0 1 µr (xq ) × ∆ <
n ∆e = e (k + 1) − e (k) = [y (k + 1) −
r r r r
0. Further,
q=1
y r (k)] × q =0 1 µr (xq ) × ∆ < 0 as y r (k + 1) > y r (k).
which implies Based on er (k + 1) > 0, er (k) > 0, and er (k + 1) <

m
n0 er (k), we have |er (k + 1)| < |er (k)|.
∂or b) Else if y r (k) r
= µ r
(y p ) × µr (xq ) (C10) < 0, then from n(C9), e (k) > 0, which
∂y r m
implies η × p = 1 µr (ȳp ) × q =0 1 µr (xq ) × er (k) > 0
p=1 q=1
and y r (k + 1) < y r (k) from (C11). Given a
∂or proper small value of the learning rate η in
y r (k + 1) = y r (k) − η × × er (C11), then y r (k) and y r (k + 1) have the
∂y r
same sign, i.e., y r (k + 1) <0. Then, based on

m
n0 n
(9), er (k + 1) = y r (k + 1) × q =0 1 µr (xq ) × ∆ > 0.
= y r (k) − η × µr (yp ) × µr (xq ) × er . Further, ∆e = e (k + 1) − er (k) = [y r (k + 1) −
r r
n
p=1 q=1 y r (k)] × q =0 1 µr (xq ) × ∆ > 0 as y r (k + 1) < y r (k).
(C11) Based on er (k + 1) < 0, er (k) < 0, and
e (k + 1) > e (k), we have |e (k + 1)| < |er (k)|.
r r r
Equation (C9) is the error caused by the rth rule of the given fuzzy By combining the conclusion of 1)–3), it is implied that
subsystem and (C11) is used for parameter updating. Based on (C8)– |er (k + 1)| ≤ |er (k)|, where er (k) and er (k + 1) are errors caused
(C11), we consider each of the three possible m cases as follows: by the normalization process. Then, we say the error introduced by
m r r m
1) If p=1
µ (ȳ ) = 0 and µ (yp ) = 0, or
m
p p=1 m r p = 1 the normalization process for the considered fuzzy subsystem can be
r r
µ (ȳp ) = p = 1 µ (yp ) > 0, i.e., ∆ = i = 1 µ (ȳp ) − corrected in the following training process. This conclusion can then be
m r
i= 1
µ (y p ) = 0, then based on (C9), er = 0, i.e., no applied to arbitrary fuzzy subsystems involving intermediate variables.
error is introduced by the normalization for this rule. Remark C.1: The aforementioned proof is based on that, for a given
Further, no action is taken to adjust the parameters, as training data point, the error introduced by the normalization for this
y r (k + 1) = y r (k) − η × (∂or /∂y r ) × er = y r (k) based on data point in some iteration will be corrected by the same data point
(C11).m m r m in subsequent iterations. In the practical execution of the proposed
2) If ( i = 1 µr (y p ) = 0 and i= 1
µ (ȳ p ) > 0), or (i = 1 algorithm, the error introduced by the normalization of a training data
m m m
µr (yp ) > 0, µr (ȳpm) > 0r and im= 1 µ r (ȳp ) − i = 1
r point, in fact, can occur with the same iteration, if, in the training
i= 1
µ (yp ) > 0), then ∆ = i = 1 µ (ȳp ) − i = 1 µ (yp ) > 0.
r dataset, there are input–output pairs which are very similar to the given
Then, there are two possible cases that are considered next. data point. This can be verified by using similar arguments to those
a) If y r (k) > 0, mthen rfrom (C9),
r
n 0e (k) > 0, which im- given in the aforementioned proof.
plies η × p = 1 µ (ȳp ) × q = 1 µr (xq ) × er (k) > 0
and y r (k + 1) < y r (k) from (C11). Given a ACKNOWLEDGMENT
proper small value of the learning rate η in
(C11), then y r (k) and y r (k + 1) have the The authors would like to thank the Associate Editor and reviewers
same sign, i.e., y r (k + 1) >0. Then, based on for their detailed constructive comments and valuable suggestions that
n
(C9), er (k + 1) = y r (k + 1) × q =0 1 µr (xq ) × ∆ > 0. have greatly helped improve the presentation and quality of the paper.
Further, ∆e = e (k + 1) − er (k) = [y r (k + 1) −
r r
n
y r (k)] × q =0 1 µr (xq ) × ∆ < 0 as y r (k + 1) < y r (k). REFERENCES
Based on er (k + 1) > 0, er (k) > 0, and [1] H. Ying, “Sufficient conditions on general fuzzy systems as function ap-
er (k + 1) < er (k), we have |er (k + 1)| < |er (k)|. proximators,” Automatica, vol. 30, pp. 521–525, 1994.
b) Else if y r (k) < 0, then from n(C9), er (k) < 0, which [2] X. J. Zeng and M. G. Singh, “Approximation theory of fuzzy systems—
m
implies η × p = 1 µ (ȳp ) × q = 1 µ (xq ) × er (k) < 0
r 0 r SISO case,” IEEE Trans. Fuzzy Syst., vol. 2, no. 2, pp. 162–176, May
and y r (k + 1) > y r (k) from (C11). Given a 1994.
[3] X. J. Zeng and M. G. Singh, “Approximation accuracy analysis of fuzzy
proper small value of the learning rate η in system as function approximators,” IEEE Trans. Fuzzy Syst., vol. 4, no. 1,
(C11), then y r (k) and y r (k + 1) have the pp. 44–63, Feb. 1996.
same sign, i.e., y r (k + 1) <0. Then, based on [4] J. J. Buckley, “Universal fuzzy controller,” Automatica, vol. 28, pp. 1245–
n
(9), er (k + 1) = y r (k + 1) × q =0 1 µr (xq ) × ∆ < 0. 1248, 1992.
[5] L. X. Wang, “Fuzzy systems are universal approximators,” in Proc. Conf.
Further, ∆er = er (k + 1) − er (k) = [y r (k + 1) − Fuzzy Syst., San Diego, CA, 1992, pp. 1163–1170.
n
y r (k)] × q =0 1 µr (xq ) × ∆ > 0 as y r (k + 1) > y r (k). [6] T. Takagi and M. Sugeno, “Fuzzy identification of systems and its ap-
Based on er (k + 1) < 0, er (k) < 0, and plications to modelling and control,” IEEE Trans. Syst., Man Cybern.,
e (k + 1) > e (k), we have |e (k + 1)| < |er (k)|.
r r r vol. SMC-15, no. 1, pp. 116–132, Feb. 1985.
[7] G. Tsekourasa, H. Sarimveisb, and E. K. George, “A hierarchical fuzzy- Observer-Based Relaxed H∞ Control for Fuzzy Systems
clustering approach to fuzzy modelling,” Fuzzy Sets Syst., vol. 150, Using a Multiple Lyapunov Function
pp. 245–266, 2005.
[8] Y. EI-Sonbaty and M. A. Ismail, “Fuzzy clustering for symbolic data,”
IEEE Trans. Fuzzy Syst., vol. 6, no. 2, pp. 195–204, May 1998.
Sung Hyun Kim and PooGyeon Park
[9] G. V. S. Raju and J. Zhou, “Adaptive hierarchical fuzzy controller,” IEEE
Trans. Syst., Man, Cybern., vol. 23, no. 4, pp. 973–980, Jul./Aug. 1993. Abstract—This short paper proposes a method of designing a fuzzy
[10] V. Torra, “A Review of the construction of hierarchical fuzzy systems,” observer-based H∞ controller for discrete-time Takagi–Sugeno (T–S) fuzzy
Int. J. Intell. Syst., vol. 17, pp. 531–543, 2002. systems. To enhance the applicability of the output-feedback controller and
[11] L. X. Wang, “Analysis and design of hierarchical fuzzy systems,” IEEE improve its performance, this short paper first builds a set of fuzzy control
Trans. Fuzzy Syst., vol. 7, no. 5, pp. 617–624, Oct. 1999. rules with premise variables different from those of the T–S fuzzy system,
[12] F. L Chung and J. C. Duan, “On multistage fuzzy neural network model- and sets the overall controller to be dependent on not only the current time
ing,” IEEE Trans. Fuzzy Syst., vol. 8, no. 2, pp. 125–142, Apr. 2000. but also the one-step-past information on the estimated fuzzy weighting
[13] R. J. G. B. Campello and W. C. Amaral, “Optimization of hierarchical functions. Then, based on the fuzzy control rules, this short paper estab-
neural fuzzy models,” in Proc. IEEE-INNS-ENNS Int. Joint Conf. Neural lishes a less conservative H∞ stabilization condition incorporated with
Netw., Como, Italy, Jul. 2000, vol. 5, pp. 8–13. a multiple Lyapunov function dependent on the estimated fuzzy weight-
[14] M. G. Joo and J. S. Lee, “Universal approximation by hierarchical fuzzy ing functions. Through a two-step design procedure, the H∞ stabilization
systems with constraints on the fuzzy rules,” Fuzzy Sets Syst., vol. 130, condition is formulated in terms of parameterized linear matrix equalities
pp. 175–188, 2002. (PLMIs), which are reconverted into LMIs with the help of an efficient and
[15] M. G. Joo and J. S. Lee, “A class of hierarchical fuzzy systems with effective relaxation scheme.
constrains on fuzzy rules,” IEEE Trans. Fuzzy Syst., vol. 13, no. 2, pp. 194–
203, Apr. 2005. Index Terms—Fuzzy weighting-dependent Lyapunov function
[16] X. J. Zeng and J. A. Keane, “Approximation capabilities of hierarchical (FWDLF), H∞ performance, observer-based fuzzy control, non-parallel
fuzzy systems,” IEEE Trans. Fuzzy Syst., vol. 13, no. 5, pp. 659–672, distributed compensation (non-PDC) scheme, parameterized linear matrix
Oct. 2005. inequalities (PLMIs), relaxation scheme.
[17] E. Mamdani, “Advances in the linguistic synthesis of fuzzy controller,”
Int. J. Man-Mach. Stud., vol. 8, no. 6, pp. 669–678, 1976.
[18] X. J. Zeng and M. G. Singh, “Decomposition property of fuzzy systems I. INTRODUCTION
and its applications,” IEEE Trans. Fuzzy Syst., vol. 4, no. 2, pp. 149–165,
May 1996. For a systematic control design of nonlinear systems, the Takagi–
[19] V. Duraisamy, N. Devarajan, D. Somasundareswari, S. N. Sivanan- Sugeno (T–S) fuzzy model [1], [27] has been a popular choice not only
dam. (2004). Comparative study of membership functions for de-
sign of fuzzy logic fault diagnosis system for single phase induc-
in consumer products but also in industrial processes due to its ability
tion motor. Acad. Open Internet J. [Online]. Available: http://www. to represent the nonlinear system only from input–output data without
acadjournal.com/2004/V13/Part6/p4/ complex mathematical equations. Thus, based on the T–S fuzzy model,
[20] G. Feng, “A Survey on analysis and design of model-based fuzzy control various kinds of fuzzy control methods have been developed under
systems,” IEEE Trans. Fuzzy Syst., vol. 14, no. 5, pp. 676–697, Oct. the so-called parallel distributed compensation (PDC) scheme [2], [4],
2006.
[15]. Recently and notably, one has focused on developing the fuzzy
control method associated with a fuzzy weighting-dependent (multiple)
Lyapunov function (FWDLF) [11], [20], [24], which is because, for a
large number of fuzzy rules, the use of the common quadratic Lyapunov
function (CQLF) [2], [4], [22], [25] leads to overconservative design
solutions. In spite of this trend, to the best of our knowledge, there has
been almost no results using the FWDLF approach when designing
fuzzy output-feedback controllers, except for [24].
Practically, all states are not fully measurable; hence, it is necessary
in this area to design a fuzzy output-feedback controller, such as static
output feedback [14], [22], [25], dynamic output feedback [15], [24],
and observer-based output feedback [5], [7], [9], [12], [13], [16], [21],
[23], [28]. Especially in the case where the premise variables of the T–S
fuzzy system are related to the immeasurable state, one needs to design
a fuzzy observer estimating the state. Of course, under the assumption
that the premise variables of the fuzzy controller are same as those of
the T–S fuzzy system, Ma et al. [5] and Yoneyama et al. [7] proved the
separation principle for the observer-based fuzzy controller; Xiaodong
and Qingling [13] proposed two sufficient linear matrix inequalities
(LMI) conditions guaranteeing the existence of the observer-based H∞
control; and Lo and Lin [16] proposed the method of designing an
Manuscript received February 11, 2008; revised May 30, 2008 and October
2, 2008; accepted December 6, 2008. First published December 22, 2008; cur-
rent version published April 1, 2009. This work was supported in part by the
Ministry of Knowledge Economy, Korea, under the Information Technology
Research Center (ITRC) support program supervised by the Institute of In-
formation Technology Advancement (IITA): IITA-2008-C1090-0801-0037 and
IITA-2008-C1090-0801-0004.
The authors are with the Division of Electrical and Computer Engineer-
ing, Pohang University of Science and Technology, Pohang 790-784, Korea
(ie222@postech.ac.kr; ppg@postech.ac.kr).
1063-6706/$25.00 © 2009 IEEE
[7] G. Tsekourasa, H. Sarimveisb, and E. K. George, “A hierarchical fuzzy- Observer-Based Relaxed H∞ Control for Fuzzy Systems
clustering approach to fuzzy modelling,” Fuzzy Sets Syst., vol. 150, Using a Multiple Lyapunov Function
pp. 245–266, 2005.
[8] Y. EI-Sonbaty and M. A. Ismail, “Fuzzy clustering for symbolic data,”
IEEE Trans. Fuzzy Syst., vol. 6, no. 2, pp. 195–204, May 1998.
Sung Hyun Kim and PooGyeon Park
[9] G. V. S. Raju and J. Zhou, “Adaptive hierarchical fuzzy controller,” IEEE
Trans. Syst., Man, Cybern., vol. 23, no. 4, pp. 973–980, Jul./Aug. 1993. Abstract—This short paper proposes a method of designing a fuzzy
[10] V. Torra, “A Review of the construction of hierarchical fuzzy systems,” observer-based H∞ controller for discrete-time Takagi–Sugeno (T–S) fuzzy
Int. J. Intell. Syst., vol. 17, pp. 531–543, 2002. systems. To enhance the applicability of the output-feedback controller and
[11] L. X. Wang, “Analysis and design of hierarchical fuzzy systems,” IEEE improve its performance, this short paper first builds a set of fuzzy control
Trans. Fuzzy Syst., vol. 7, no. 5, pp. 617–624, Oct. 1999. rules with premise variables different from those of the T–S fuzzy system,
[12] F. L Chung and J. C. Duan, “On multistage fuzzy neural network model- and sets the overall controller to be dependent on not only the current time
ing,” IEEE Trans. Fuzzy Syst., vol. 8, no. 2, pp. 125–142, Apr. 2000. but also the one-step-past information on the estimated fuzzy weighting
[13] R. J. G. B. Campello and W. C. Amaral, “Optimization of hierarchical functions. Then, based on the fuzzy control rules, this short paper estab-
neural fuzzy models,” in Proc. IEEE-INNS-ENNS Int. Joint Conf. Neural lishes a less conservative H∞ stabilization condition incorporated with
Netw., Como, Italy, Jul. 2000, vol. 5, pp. 8–13. a multiple Lyapunov function dependent on the estimated fuzzy weight-
[14] M. G. Joo and J. S. Lee, “Universal approximation by hierarchical fuzzy ing functions. Through a two-step design procedure, the H∞ stabilization
systems with constraints on the fuzzy rules,” Fuzzy Sets Syst., vol. 130, condition is formulated in terms of parameterized linear matrix equalities
pp. 175–188, 2002. (PLMIs), which are reconverted into LMIs with the help of an efficient and
[15] M. G. Joo and J. S. Lee, “A class of hierarchical fuzzy systems with effective relaxation scheme.
constrains on fuzzy rules,” IEEE Trans. Fuzzy Syst., vol. 13, no. 2, pp. 194–
203, Apr. 2005. Index Terms—Fuzzy weighting-dependent Lyapunov function
[16] X. J. Zeng and J. A. Keane, “Approximation capabilities of hierarchical (FWDLF), H∞ performance, observer-based fuzzy control, non-parallel
fuzzy systems,” IEEE Trans. Fuzzy Syst., vol. 13, no. 5, pp. 659–672, distributed compensation (non-PDC) scheme, parameterized linear matrix
Oct. 2005. inequalities (PLMIs), relaxation scheme.
[17] E. Mamdani, “Advances in the linguistic synthesis of fuzzy controller,”
Int. J. Man-Mach. Stud., vol. 8, no. 6, pp. 669–678, 1976.
[18] X. J. Zeng and M. G. Singh, “Decomposition property of fuzzy systems I. INTRODUCTION
and its applications,” IEEE Trans. Fuzzy Syst., vol. 4, no. 2, pp. 149–165,
May 1996. For a systematic control design of nonlinear systems, the Takagi–
[19] V. Duraisamy, N. Devarajan, D. Somasundareswari, S. N. Sivanan- Sugeno (T–S) fuzzy model [1], [27] has been a popular choice not only
dam. (2004). Comparative study of membership functions for de-
sign of fuzzy logic fault diagnosis system for single phase induc-
in consumer products but also in industrial processes due to its ability
tion motor. Acad. Open Internet J. [Online]. Available: http://www. to represent the nonlinear system only from input–output data without
acadjournal.com/2004/V13/Part6/p4/ complex mathematical equations. Thus, based on the T–S fuzzy model,
[20] G. Feng, “A Survey on analysis and design of model-based fuzzy control various kinds of fuzzy control methods have been developed under
systems,” IEEE Trans. Fuzzy Syst., vol. 14, no. 5, pp. 676–697, Oct. the so-called parallel distributed compensation (PDC) scheme [2], [4],
2006.
[15]. Recently and notably, one has focused on developing the fuzzy
control method associated with a fuzzy weighting-dependent (multiple)
Lyapunov function (FWDLF) [11], [20], [24], which is because, for a
large number of fuzzy rules, the use of the common quadratic Lyapunov
function (CQLF) [2], [4], [22], [25] leads to overconservative design
solutions. In spite of this trend, to the best of our knowledge, there has
been almost no results using the FWDLF approach when designing
fuzzy output-feedback controllers, except for [24].
Practically, all states are not fully measurable; hence, it is necessary
in this area to design a fuzzy output-feedback controller, such as static
output feedback [14], [22], [25], dynamic output feedback [15], [24],
and observer-based output feedback [5], [7], [9], [12], [13], [16], [21],
[23], [28]. Especially in the case where the premise variables of the T–S
fuzzy system are related to the immeasurable state, one needs to design
a fuzzy observer estimating the state. Of course, under the assumption
that the premise variables of the fuzzy controller are same as those of
the T–S fuzzy system, Ma et al. [5] and Yoneyama et al. [7] proved the
separation principle for the observer-based fuzzy controller; Xiaodong
and Qingling [13] proposed two sufficient linear matrix inequalities
(LMI) conditions guaranteeing the existence of the observer-based H∞
control; and Lo and Lin [16] proposed the method of designing an
Manuscript received February 11, 2008; revised May 30, 2008 and October
2, 2008; accepted December 6, 2008. First published December 22, 2008; cur-
rent version published April 1, 2009. This work was supported in part by the
Ministry of Knowledge Economy, Korea, under the Information Technology
Research Center (ITRC) support program supervised by the Institute of In-
formation Technology Advancement (IITA): IITA-2008-C1090-0801-0037 and
IITA-2008-C1090-0801-0004.
The authors are with the Division of Electrical and Computer Engineer-
ing, Pohang University of Science and Technology, Pohang 790-784, Korea
(ie222@postech.ac.kr; ppg@postech.ac.kr).
1063-6706/$25.00 © 2009 IEEE
observer-based robust H∞ control. Besides, Lam and Zhou [24] also where
applied the FWDLF approach under the assumption. However, it is  A(Θ ) B (Θ ) B (Θ ) 
k 1 k 2 k
worth noting that the assumption imposes a strict constraint on the
 C1 (Θk ) D1 1 (Θk ) D1 2 (Θk ) 
applicability of the designed controller in practical applications [12],
[21]. C2 (Θk ) D2 1 (Θk ) 0
This short paper aims at enhancing the applicability of the output- A B1 i B2 i

feedback controller while improving its H∞ performance. To this end,
r i
we first build a set of fuzzy control rules with premise variables differ- = 
θi (η(k)) C1 i D1 1 i D1 2 i  (3)
ent from those of the T–S fuzzy system and set the overall controller i= 1
C2 i D2 1 i 0
to be dependent on not only the current time but also the one-step-past
information on the estimated fuzzy weighting functions. Then, based gi (η(k))
s
θi (η(k)) = r , gi (η(k)) = fi j (ηj (k)). (4)

on the fuzzy control rules, we establish a less conservative H∞ stabi- g (η(k))
i= 1 i j=1
lization condition incorporated with a multiple Lyapunov function. In
the derivation, the H∞ stabilization conditions are formulated in terms The notation η(k) = [η1 (k), . . . , ηs (k)]T ∈ Rs , fi j (ηj (k)) denotes
of parameterized LMIs (PLMIs) through a two-step design procedure the grade of membership of ηj (k) in Fi j , and Θk ∈ Rr stands for a
similar to [12], [13], [16], and [21], which are reconverted into LMIs vector of time-varying fuzzy weighting functions θi (η(k)). Moreover,
r
with the help of an efficient and effective relaxation scheme [11]. Since let gi (η(k)) ≥ 0, for i = 1, . . . , r, and i = g (η(k)) > 0 for all time
1 i
r
the proposed stabilization condition incorporates some additional con- k. Then, θi (η(k)) ≥ 0, for i = 1, . . . , r, θ (η(k)) = 1 for all
i= 1 i
ditions on fuzzy weighting functions into the interactions among the time k. Here, if the state is directly unmeasurable, it is impossible for
fuzzy subsystems, more powerful and useful results can be obtained as a fuzzy control law to share the same premise variables as those of the
the number of the fuzzy rules involved increases. system (2) since the premise variables are generally associated with
Notation: The Lebesgue space L2 + = L2 [0, ∞) consists of square- the state of the system (2). Thus, we consider a set of fuzzy control
integrable functions on [0, ∞). The notations X ≥ Y and X > Y mean IF–THEN rules such as the following.
that X − Y is positive semidefinite and positive definite, respectively. Control Rule i: IF η̂1 (k) is Fi 1 and · · · and η̂s (k) is Fi s , THEN
The notation A ⊕ B stands for the Kronecker sum of two matrices A
and B. In symmetric block matrices, (∗) is used as an ellipsis for terms x̂k + 1 = Ai x̂k + B2 i uk + L(Θ̂k )(yk − ŷk ), ŷk = C2 i x̂k
that are induced by symmetry. uk = F (Θ̂k −1 , Θ̂k )x̂k , for i = 1, 2, . . . , r (5)
II. SYSTEM DESCRIPTION AND PRELIMINARIES where η̂1 (k) · · · η̂s (k) denote the premise variables of the controller,
x̂k ∈ Rn x and ŷk ∈ Rn y denote the estimated state and measurement
Consider a discrete-time T–S fuzzy model described by fuzzy IF– output, respectively, and the notation Θ̂k (or Θ̂k −1 ) denotes a vector
THEN rules of the following form: of time-varying fuzzy weighting functions θi (η̂(k)) [or θi (η̂(k − 1))]
Plant rule i: IF η1 (k) is Fi 1 and · · · and ηs (k) is Fi s , THEN with respect to (4). Here, L(Θ̂k ) and F (Θ̂k −1 , Θ̂k ) are not assumed to
be affine in time-varying θi (η̂(k)) and θi (η̂(k − 1)).
xk + 1 = A i xk + B 1 i w k + B 2 i u k
Remark 1: Differently from [21], the proposed observer-based fuzzy
zk = C 1 i xk + D 1 1 i w k + D 1 2 i u k controller is designed to be dependent not only on the current time esti-
mated fuzzy weighting function vector Θ̂k but also on the one-step-past
y k = C 2 i xk + D 2 1 i w k , for i = 1, 2, . . . , r (1)
estimated fuzzy weighting function vector Θ̂k −1 for time k, which is
where the consequent subsystems (1) represent linear systems in local afterward implemented by nonparallel distributed compensation (non-
operating regions, Fi j denotes a fuzzy set, η1 (k), . . . , ηs (k) denote the PDC) scheme.
premise variables of the plant, r denotes the number of IF–THEN rules, Remark 2: A reason for employing both Θ̂k −1 and Θ̂k in (5) is to
and xk ∈ Rn x , uk ∈ Rn u , yk ∈ Rn y , wk ∈ Rn w , and zk ∈ Rn z de- use the transition information existing between Θ̂k −1 and Θ̂k as well as
note the state, the control input, the measured output, the disturbance, the instant information Θ̂k when performing the control action (refer
and the performance output, respectively. For the premise variables, to [11] and [18]).
η1 (k), . . . , ηs (k), we assume that the following assumptions are valid. As in (2), the fuzzy control system (5) can be represented as follows:
A1) The premise variables do not explicitly depend on the control
x̂k + 1 = A(Θ̂k )x̂k + B2 (Θ̂k )uk + L(Θ̂k ) (yk − ŷk )
uk and the disturbance wk ∈ L2 + .
A2) The premise variables are determined by the unmeasurable state ŷk = C2 (Θ̂k )x̂k
of the plant.
uk = F (Θ̂k −1 , Θ̂k )x̂k (6)
Based on A1), the overall fuzzy model is inferred as follows:
r

r
xk + 1 = A(Θk )xk + B1 (Θk )wk + B2 (Θk )uk where A(Θ̂k ) = θ (η̂(k))Ai , B2 (Θ̂k ) =
i= 1 i
θ (η̂(k))B2 i ,
i= 1 i

r
and C2 (Θ̂k ) = θ (η̂(k))C2 i . Hence, the closed-loop fuzzy sys-
i= 1 i
zk = C1 (Θk )xk + D1 1 (Θk )wk + D1 2 (Θk )uk
tem under the control law (6) is given by (7), as shown at the bottom
yk = C2 (Θk )xk + D2 1 (Θk )wk (2) of this page.
 
  A(Θ̂k ) + B2 (Θ̂k )F (Θ̂k −1 , Θ̂k ) − L(Θ̂k )C2 (Θ̂k ) L(Θ̂k )C2 (Θk ) L(Θ̂k )D21 (Θk )  
x̂k +1   x̂k
x   B2 (Θk )F (Θ̂k −1 , Θ̂k ) A(Θk ) B1 (Θk )  
 k +1  =    xk  . (7)
 
zk D12 (Θk )F (Θ̂k −1 , Θ̂k ) C1 (Θk ) D11 (Θk ) wk
Now, let us consider an appropriate similarity transform matrix specification (DS): the closed-loop system is asymptotically stable for
yielding all admissible grades Θk , Θ̂k , and Θ̂k −1 when wk ≡ 0, i.e., this short
x̂k I 0 x̂k x̂k paper aims at solving the following optimization problem:
TS = = (8)
xk −I I xk ek min γ subject to zk 2 < γ wk 2 and DS
where ek denotes the estimation error ek = xk − x̂k . Then, the equiv-
alent representation of (7) becomes for all nonzero wk ∈ L2 + . (11)
Consider a Lyapunov function V (x̃k ) of the following form:
x̃k + 1 Ã(Θk , Θ̂k , Θ̂k −1 ) B̃(Θk , Θ̂k ) x̃k
= (9)
zk C̃(Θk , Θ̂k , Θ̂k −1 ) D1 1 (Θk ) wk V (x̃k ) = x̃Tk P (Θ̂k −1 )x̃k , P (Θ̂k −1 ) > 0 (12)
where x̃Tk = [x̂Tk eTk ] denotes the augmented state, and the element whose forward difference along the closed-loop system trajectories

matrices are given by ∆V (x̃k ) = V (x̃k + 1 ) − V (x̃k ) is given by

Ã1 1 L(Θ̂k )C2 (Θk ) ∆V (x̃k ) = x̃Tk + 1 P (Θ̂k )x̃k + 1 − x̃Tk P (Θ̂k −1 )x̃k . (13)
Ã(Θk , Θ̂k , Θ̂k −1 ) =
Ã2 1 Ã2 2 Then, the following two statements are equivalent (see [11] and [16]).

L(Θ̂k )D2 1 (Θk ) 1) The closed-loop system (9) is stable with the H∞ performance
B̃(Θk , Θ̂k ) = γ.
B1 (Θk ) − L(Θ̂k )D2 1 (Θk )
2) There exist P (Θ̂k −1 ) and P (Θ̂k ) such that
C̃(Θk , Θ̂k , Θ̂k −1 ) = [ C̃1 1 C1 (Θk ) ]  
P (Θ̂k −1 ) 0 (∗) (∗)
in which Ã1 1 = A(Θ̂k ) + B2 (Θ̂k )F (Θ̂k −1 , Θ̂k ) + L(Θ̂k ) (C2 (Θk ) − 
 0 γ2 I (∗) (∗) 

C2 (Θ̂k )), Ã2 1 = (A(Θk ) − A(Θ̂k )) + (B2 (Θk ) − B2 (Θ̂k ))F (Θ̂k −1 ,  − − − − − − − − − − − − − − − − − − − − −
Θ̂k ) − L(Θ̂k )(C2 (Θk ) − C2 (Θ̂k )), Ã2 2 = A(Θk ) − L(Θ̂k )C2 (Θk ), 
0<  . (14)
−1 
and C̃1 1 = C1 (Θk ) + D1 2 (Θk )F (Θ̂k −1 , Θ̂k ).  Ã(Θk , Θ̂k , Θ̂k −1 ) B̃(Θk , Θ̂k ) P (Θ̂k ) 0 
 C̃(Θ , Θ̂ , Θ̂ ) I 
Remark 3: From (9), we can observe that if wk ≡ 0 and Θk ≡ Θ̂k , k kD (Θ )
k −1 0 11 k
then the closed-loop system becomes x̃k + 1 = Ã(Θk , Θk −1 )x̃k , where

AF (Θk −1 , Θk ) L(Θk )C2 (Θk ) As is well known, (14) is directly derived by 0 < γ 2 wkT wk − zkT zk −
Ã(Θk , Θk −1 ) = (10) ∆V (x̃k ).
0 AL (Θk )
Remark 4: To enhance the causality between the controller

F (Θ̂k −1 , Θ̂k ) and the difference ∆V (x̃k ), we set the Lyapunov func-
where AF (Θk −1 , Θk ) = A(Θk ) + B2 (Θk )F (Θk −1 , Θk ) and

tion (12) to be dependent not on Θ̂k but on Θ̂k −1 (refer to [11] and [18]).
AL (Θk ) = A(Θk ) − L(Θk )C2 (Θk ). In this sense, Ma et al. [5]
Let us partition matrices P (Θ̂k −i ) and P̄ (Θ̂k −i ) = P −1 (Θ̂k −i ), for

proved, based on the vector comparison principle, that the design of
i = 0, 1, in the form
the controller can be carried out independently of the design of the
observer, and Sala et al. [27] also mentioned on it.

X(Θ̂k −i ) 0
Finally, note that θi (η̂(k)) is generally subject to the follow- P (Θ̂k −i ) =
0 Z(Θ̂k −i )
ing constraints
r for all time k: C1) 0 ≤ θi (η̂(k)) ≤ 1 ∀i ∈ 1, . . . , r;
C2) i = 1 θi (η̂(k)) = 1; C3) |θi (ηk ) − θi (η̂k )| ≤ δi < 1, from which X̄(Θ̂k −i ) 0
we shall derive a relaxed H∞ stabilization condition based on LMIs. P̄ (Θ̂k −i ) = (15)
0 Z̄(Θ̂k −i )
III. H∞ STABILIZATION BASED ON PLMI CONDITIONS where all element matrices are positive definite, X̄(Θ̂k −i ) =
In general, the H∞ norm boundedness of the transfer function from X −1 (Θ̂k −i ), and Z̄(Θ̂k −i ) = Z −1 (Θ̂k −i ). And define the matrices

w to z, Gz w , is denoted as Gz w ∞ < γ, i.e., zk 2 < γ wk 2 for all Q̄(Θ̂k −1 ) = diag(Q(Θ̂k −1 ), I) and R̄ = diag(I, R), where Q(Θ̂k −1 ) ∈
nonzero wk ∈ L2 + , where the upper bound γ represents the disturbance Rn x ×n x and R ∈ Rn x ×n x are positive definite. Then, pre- and post-

rejection capability [3]. This short paper focuses on how to design an multiplying T T and T = diag(Q̄(Θ̂k −1 ), I, R̄, I) > 0 on the right-
observer-based H∞ output-feedback controller (6) that achieves the hand side of (14) yields (16), as shown at the bottom of this page, which
maximal disturbance rejection while satisfying the following design is guaranteed by (17), as shown at the bottom of this page, because, for
 
QT (Θ̂k −1 )X(Θ̂k −1 )Q(Θ̂k −1 ) ⊕ Z(Θ̂k −1 ) 0 (∗) (∗)
 0 γ2 I (∗) (∗) 
 
0<  (16)
 R̄T Ã(Θk , Θ̂k , Θ̂k −1 )Q̄(Θ̂k −1 ) R̄T B̃(Θk , Θ̂k ) X̄(Θ̂k ) ⊕ RT Z̄(Θ̂k )R 0 
C̃(Θk , Θ̂k , Θ̂k −1 )Q̄(Θ̂k −1 ) D11 (Θk ) 0 I
 
(QT (Θ̂k −1 ) + Q(Θ̂k −1 ) − X̄(Θ̂k −1 )) ⊕ Z(Θ̂k −1 ) 0 (∗) (∗)

 0 γ2 I (∗) (∗) 

0<  (17)
 R̄T Ã(Θk , Θ̂k , Θ̂k −1 )Q̄(Θ̂k −1 ) R̄T B̃(Θk , Θ̂k ) X̄(Θ̂k ) ⊕ (RT + R − Z(Θ̂k )) 0 
C̃(Θk , Θ̂k , Θ̂k −1 )Q̄(Θ̂k −1 ) D11 (Θk ) 0 I
any matrix Q, it always holds that QT X Q ≥ QT + Q − X −1 , X > 0. IV. H∞ STABILIZATION CONDITIONS BASED ON LMIS
Here, note that (17) naturally implies Q(Θ̂k −1 ) > 0 and R > 0 under
To obtain a finite number of LMIs from the derived PLMIs, we spe-
the following conditions:
cially select the structures of the fuzzy weighting-dependent variables
0 < X̄(Θ̂k −1 ) and 0 < Z(Θ̂k ). (18) as follows:

r
Based on (9), the element matrices in (17) are formulated as follows: X̄(Θ̂k −i ) = θ (η̂(k − i))X̄ , for i = 0, 1
= 1
(1, 1)a L(Θ̂k )C2 (Θk )
R̄T Ã(·)Q̄(Θ̂k −1 ) =
r
(2, 1)a (2, 2)a Z(Θ̂k −i ) = θ (η̂(k − i))Z , for i = 0, 1

(1, 1)a = A(Θ̂k ) + L(Θ̂k ) C2 (Θk ) − C2 (Θ̂k ) = 1

r
r
× Q(Θ̂k −1 ) + B2 (Θ̂k )F̄ (Θ̂k −1 , Θ̂k ) Q(Θ̂k −1 ) = θ (η̂(k − 1))Q , L̄(Θ̂k ) = θ (η̂(k))L̄

(2, 1)a = R T
A(Θk ) − A(Θ̂k ) Q(Θ̂k −1 ) = 1 = 1
F̄ (Θ̂k −1 , Θ̂k ) = F̄1 (Θ̂k −1 ) + F̄2 (Θ̂k ) (21)

+ RT B2 (Θk ) − B2 (Θ̂k ) F̄ (Θ̂k −1 , Θ̂k ) r r
where F̄1 (Θ̂k −1 ) = = 1 θ (η̂(k − 1))F̄1 and F̄2 (Θ̂k ) = = 1
− L̄(Θ̂k ) C2 (Θk ) − C2 (Θ̂k ) Q(Θ̂k −1 ) θ (η̂(k))F̄2 , i.e., all variables are assumed to be affinely depen-
dent on θ (η̂(k)) or θ (η̂(k − 1)). Henceforth, for shortening the
(2, 2)a = RT A(Θk ) − L̄(Θ̂k )C2 (Θk ) expression, we use the following notations: θi = θi (η(k)), θ̂i =
θi (η̂(k)), and θ̂i− = θi (η̂(k − 1)).
L(Θ̂k )D2 1 (Θk )
R̄T B̃(Θk , Θ̂k ) =
(2, 1)b
A. Step 1: Observer Design Conditions
(2, 1)b = R B1 (Θk ) − L̄(Θ̂k )D2 1 (Θk )
T
Based on (21), the PLMI conditions (19) and 0 < Z(Θ̂k ) can be,
C̃(·)Q̄(Θ̂k −1 ) = [ (1, 1)c C1 (Θk ) ] respectively, rewritten as

r
(1, 1)c = C1 (Θk )Q(Θ̂k −1 ) + D1 2 (Θk )F̄ (Θ̂k −1 , Θ̂k ) 0< θi θ̂j θ̂− Li j (22)

i , j, = 1
where F̄ (Θ̂k −1 , Θ̂k ) = F (Θ̂k −1 , Θ̂k )Q(Θ̂k −1 ) and L̄(Θ̂k ) =
RT L(Θ̂k ). In spite of employing P (Θ̂k ) and P̄ (Θ̂k ) of the form (15),
r
the observer-based H∞ stabilization problem (17) is still nonconvex. 0< θ̂j Zj (23)
j=1
Thus, similarly to [12], [13], [16], and [21], we shall propose a proper
two-step procedure of converting the stabilization problem (17) into where
two PLMI problems, based on the fact that (17) implies
 
Z 0 (∗) (∗)
   2 
Z(Θ̂k −1 ) 0 (∗) (∗) 0 γ I (∗) (∗)
Li j =  

  (2, 2)a (2, 1)b 0 

(∗)  R + R − Zj
T
 0 γ2 I (∗) 
0< . (19) C1 i D1 1 i 0 I
 (2, 2)a (2, 1)b RT + R − Z(Θ̂k ) 0 
in which (2, 2)a = RT Ai − L̄j C2 i and (2, 1)b = RT B1 i − L̄j D2 1 i .
C1 (Θk ) D1 1 (Θk ) 0 I
The following lemma presents a set of LMI conditions for step 1 in
Proposition 1: Following is the two-step procedure. Proposition 1.
Step 1: For a prescribed γ > 0, solve (19) and 0 < Z(Θ̂k ). Then, Lemma 4.1: For a prescribed γ > 0, suppose that there exist Zi > 0,
from its solutions, we can reconstruct the observer gain L(Θ̂k ) as L̄i , for i ∈ [1, r], and R such that
follows: L(Θ̂k ) = R−T L̄(Θ̂k ).
Step 2: Substitute the obtained R and L̄(Θ̂k ) into (17), and solve 0 < Li j ∀i, j, ∈ [1, r]. (24)
the following optimization problem:
Then, the observer gain L(Θ̂k ), fork ≥ 0, can be online recon-
r
γC = min γ structed as follows: L(Θ̂k ) = R−T θ̂ L̄ . Moreover, a min-
i= 1 i i
X̄ ( Θ̂ k ), X̄ ( Θ̂ k −1 ), Z ( Θ̂ k ), Z ( Θ̂ k −1 ), Q ( Θ̂ k −1 ), F̄ ( Θ̂ k −1 , Θ̂ k )
imized H∞ performance (observer) can be obtained by γO =
subject to (17) and (18). min γ subject to (24).
Proof: Obviously, (22) and (23) are guaranteed by (24) that naturally
Then, from its solutions, we can reconstruct the control gain
implies 0 < Zj , j ∈ [1, r].
F (Θ̂k −1 , Θ̂k ) as follows:
Remark 7: In (17), the reason for using a common r R is to
F (Θ̂k −1 , Θ̂k ) = F̄ (Θ̂k −1 , Θ̂k )Q−1 (Θ̂k −1 ). (20) make L(Θ̂k ) affinely dependent on θ̂i , i.e., L(Θ̂k ) = i = 1 θ̂i Li ,

Remark 5: Solving the PLMIs in Proposition 1 is equivalent to Li = R−T L̄i . The matrices Li will be used for the LMI formulation
solving an infinite number of LMIs. Thus, to numerically solve the of step 2.
optimization problems, one needs to find a finite number of solvable
LMIs from the PLMIs by selecting an appropriate structure for each B. Step 2: Relaxed Control Design Conditions
fuzzy weighting-dependent variable, i.e., X̄(Θ̂k ), X̄(Θ̂k −1 ), Z(Θ̂k ), r
Z(Θ̂k −1 ), Q(Θ̂k −1 ), L̄(Θ̂k ), and F̄ (Θ̂k −1 , Θ̂k ). Using (21), we can rewrite (17) as follows: 0 < ,m = 1
−
Remark 6: To reduce the number of decision variables, we can put θ θ̂m G m (Θ̂k ), as shown (25), at the bottom of next page, whose
Q(Θ̂k −1 ) = X̄(Θ̂k −1 ) in (17), but it leads to a more conservative result element matrices are given by (1, 1)a = A(Θ̂k )Qm + R−T L̄(Θ̂k )
than when the slack variable Q(Θ̂k −1 ) is used (refer to [17] and [26]). (C2 − C2 (Θ̂k ))Qm + B2 (Θ̂k )(F̄1 m + F̄2 (Θ̂k )), (1, 2)a = R−T L̄
(Θ̂k )C2 , (2, 1)a = RT (A − A(Θ̂k ))Qm + RT (B2 − B2 (Θ̂k )) 0 < X̄i , 0 < S0 + S0T , 0 < Λi + ΛTi (30)
(F̄1 m + F̄2 (Θ̂k )) − L̄(Θ̂k )(C2 − C2 (Θ̂k ))Qm , (2, 2)a = RT A −
0 < Q̌i + Q̌Ti , 0 < Ǔi + ǓiT ∀i ∈ [1, r] (31)
L̄(Θ̂k )C2 , (1, 1)b = R−T L̄(Θ̂k )D2 1 , (2, 1)b = RT B1 − L̄(Θ̂k )
D2 1 , and (1, 1)c = C1 Qm + D1 2 (F̄1 m + F̄2 (Θ̂k )). Furthermore, 0 < Ξi j ∀i, j ∈ [1, r], j = i (32)
by the S-procedure, (17) subject to C3) can be written as follows:

where Λi , Si , S0 , and Ξi j are in Rn c ×n c , nc = 4nx + nw + nz

r
r
r
0< −
θ θ̂m G m (Θ̂k) + (θi − θ̂i − δi)Ui + (−δi − θi + θ̂i)Qi
( m )
( m )

r
r

,m = 1 i= 1 i= 1 Γ0 = G0 + U − Q − δi Ui − δi Qi − S0 + S0T
r (26) i= 1 i= 1
where Ui > 0 and Qi > 0. Note that from θ =
r −
r = 1
(m )
(m )

θ̂m = 1, it follows that (θi − θ̂i − δi )Ui = ∆i = Gi i + Λi + Λi T
+ Si + Si T
r
m = 1 r r i = 1
θ θ̂ − (U − i = 1 δi Ui − i = 1 θ̂i Ui ) ≤ 0 and
r
,m = 1 m
r −
r Γi
( m )
= Gi
( m )
+ Q̌i − Ǔi − Λi + S0 − Si
(−δi − θi + θ̂i )Qi = , m = 1 θ θ̂m (−Q − i = 1 δi Qi +
r
i= 1
(m )
θ̂ Q ) ≤ 0. Then, we can obtain the following condition

i= 1 i i Φi j i j + (Si + Sj ) − Ξi j − Ξj i .
= Gm
guaranteeing (26):
Then, the closed-loop fuzzy system (9) with the constraints C1)–C3) is

r

r
0 < G m (Θ̂k ) + U − Q − δi Ui − δ i Qi asymptotically stabilizable with H∞ performance γ for all admissible

i= 1 i= 1
grades Θk , Θ̂k , and Θ̂k −1 , and the corresponding fuzzy controller gain
F (Θ̂k −1 , Θ̂k ), for k ≥ 0, can be online reconstructed as follows:
r
+ θ̂i (Qi − Ui ) ∀, m ∈ [1, r] (27) r −1

r
r
i= 1 F (Θ̂k −1 , Θ̂k ) = θ̂i− F̄1 i + θ̂i F̄2 i θ̂i− Qi .
which can be rewritten
as follows: i= 1 i= 1 i= 1
( m )

r
r
(33)
0 < G0 + U − Q − δi Ui − δ i Qi Moreover, the minimized H∞ performance can be obtained by the
i= 1 i= 1 following optimization problem:
r
r
( m ) ( m )T (m ) γC = min γ subject to (29)–(32).
+ θ̂i Gi + Gi + Qi − Ui + θ̂i2 Gi i
i= 1 i= 1 Proof: By the S-procedure [3] and Finsler’s lemma [10], [19], (28)
i −1 subject to C1)–C2) can be written as follows:

r r
+
(m )
θ̂i θ̂j Gi j +
(m )T
θ̂i θ̂j Gi j (28)

r

r
i= 1 j=1 j = i+ 1 0 < G m (Θ̂k ) + U − Q − δi Ui − δ i Qi
( m ) ( m ) (m ) (m )
where G0 , Gi , Gi i , and Gi j are defined at the bottom of i= 1 i= 1
next page.
r
Based on (28), the following theorem presents a relaxed LMI con- + θ̂i (Qi − Ui ) − N (Θ̂k ) ∀, m ∈ [1, r] (34)
dition set for step 2 of Proposition 1. i= 1
Theorem 1: Let the matrices R > 0 and L̄i , for i ∈ [1, r], be given,
and separate Qi = Q̌i + Q̌Ti and Ui = Ǔi + ǓiT . Suppose that there where 0 ≤ N (Θ̂k ) is given by
exist X̄i > 0, Zi > 0, Q̌i , Ǔi , Qi , F̄1 i , F̄2 i , Λi , Si , S0 , Ξi j , and γ > 0

r
such that N (Θ̂k ) = C1 + C1T + C2 i (Λi + ΛTi )
 ( m ) 
Γ0 (∗) (∗) ··· (∗) i= 1
 − − − − − − − − − − − − − − − − −−
r
r

 ( m )  C3 i j Ξi j + ΞTij
 Γ (m )
··· (∗) 
+ (35)
 1 ∆1 (∗) 
  i = 1 j = 1 , j = i
 ( m ) . .. 
. 
T    T T  

 Γ Φ
(m )
∆
(m ) .. 
0 < L̃ m = 2 21 2  I I S0 I
     −I   S T   
 .. .. .. .. 

 1    1   θ̂1 I 
 . . . . (∗)  θ̂ I
  0 = C1 =  
 ..   ..  
 .


 . 
 . 
 Γ( m) (m )
··· Φr (r −1 ) ∆r 
(m ) (m )
 .   .   ..   . 
 r Φr 1 
θ̂r I −I SrT θ̂r I

∀, m ∈ [1, r] (29) 0 ≤ C2 i = − θ̂i2 + θ̂i 0 ≤ C3 i j = θ̂i θ̂j .
 T 
Qm + Qm − X̄m ⊕ Zm 0 (∗) (∗)
 γ2 I (∗) 
 0 (∗) 

 

Gm (Θ̂k ) =  (1, 1)a (1, 2)a (1, 1)b  (25)
 X̄(Θ̂k ) ⊕ (RT + R − Z(Θ̂k )) 0 
 (2, 1)a (2, 2)a (2, 1)b 
[ (1, 1)c C1 ] D11 0 I
Here, C1 , C2 i , and C3 i j are from C1)–C2), respectively, and the mul- where Γ0 , Γi
( m ) ( m )
, ∆i , and Φi j
(m ) (m )
are defined in Theorem 1. Con-
tiplier variables Λi , S0 , Si , and Ξi j are in Rn c ×n c and should satisfy sequently, (37) boils down to
the conditions 0 < Λi + ΛTi and 0 < Ξi j + ΞTji in (30) and (32). With
some algebraic manipulations, (35) can be represented as follows: 0 < [ I| θ̂1 I · · · θ̂r I ]L̃ m [ I| θ̂1 I · · · θ̂r I ]T ∀, m ∈ [1, r].

r

r
Meanwhile, if (29) holds, then (34) also holds, which im-
N (Θ̂k ) = N0 + θ̂i Ni + NTi + θ̂i2 Ni i
r 0 < G m
plies (Θ̂k ) because0 ≤ N (Θ̂k ) and 0 ≥ (U − Q −
r r
i= 1 i= 1 δ U − i = 1 δi Qi ) + i = 1 θ̂i (Qi − Ui ). Thus, from (25), it
i −1 i= 1 i i

r r follows that 0 < Zm , i.e., 0 < Z(Θ̂k ). And, the condition 0 < X̄i in
+ θ̂i θ̂j Ni j + θ̂i θ̂j NTij (36) (30) assures 0 < X̄(Θ̂k −1 ).
i= 1 j=1 j = i+ 1 Remark 8: In order to reduce the number of decision vari-
ables, we can follow the constraint-elimination method in [11]:
S0 = −Λ, Si = Λ, i ∈ [1, r], and Ξi j = 0, i, j ∈ [1, r], j = i. How-
where N0 = S0 + S0T , Ni = Λi − S0 + Si , Ni i = − Λi + ΛTi −
ever, the method [11] produces a more conservative result than
Si + SiT , and Ni j = − (Si + Sj ) + (Ξi j + Ξj i ). Hence, the H∞
Theorem 1.
stabilization condition (34) becomes, for all , m ∈ [1, r]
Corollary 1: A method to reduce the computational complexity of
Theorem 1 is to set Qi = X̄i together with Remark 8, i.e., to use
( m )

r
( m ) ( m )T

r
(m )
X̄(Θ̂k −1 ) instead of Q(Θ̂k −1 ).
0 < Γ0 + θ̂i Γi + Γi + θ̂i2 ∆i Now, we shall consider some possible relaxation forms for the H∞
( m )
i= 1 i= 1 stabilization condition (27), based on, as shown (Gi j ), at the bottom
i −1 −T
of next page, where (1, 1)a = Ai Qm + R L̄i (C2 − C2 j )Qm +

r r
+
(m )
θ̂i θ̂j Φi j +
(m )T
θ̂i θ̂j Φi j (37) B2 i (F̄1 m + F̄2 j ), (1, 2)a = R−T L̄i C2 , (2, 1)a = RT (A − Ai )
i= 1 j=1 j = i+ 1 Qm + RT (B2 − B2 i )(F̄1 m + F̄2 j ) − L̄i (C2 − C2 j )Qm , (2, 2)a =
 
QTm + Qm − X̄m 0 0 (∗) (∗) (∗)
 
 0 Zm 0 (∗) (∗) (∗) 
 
 0 0 γ2 I (∗) (∗) (∗) 
(m )
 
G0 =  
 0 0 0 0 0 0 
 
 T 
 R A Qm T
+ R B2 F̄1m R A T
R B1 T
0 T
R +R 0 
C1 Qm + D12 F̄1m C1 D11 0 0 I
 
0 0 0 0 0 0
 0
 0 0 0 0 0 
 
 0 0 0 0 0 0
 
(m )
 1 
Gi =  Ai Qm + R−T L̄i C2 Qm + B2i F̄1m R−T L̄i C2 R−T L̄i D21 X̄i 0 0
 2 
 
 1 
 −RT Ai Qm − RT B2i F̄1m + RT B2 F̄2i − L̄i C2 Qm −L̄i C2 −L̄i D21 0 − Zi 0
 2 
D12 F̄2i 0 0 0 0 0
 
0 0 0 (∗) (∗) 0
 
 0 0 0 0 0 0
 
 0 0 0 0 0 0
(m )
 
Gii =  
 −R L̄i C2i Qm + B2i F̄2i
−T
0 0 0 0 0
 
 
 −RT B2i F̄2i + L̄i C2i Qm 0 0 0 0 0
0 0 0 0 0 0
 
0 0 0 0 0 0
 
 0 0 0 0 0 0
 
 0 0 0 0 0 0
(m )
 
Gij =  .
 −R−T L̄i C2j Qm + B2i F̄2j − R−T L̄j C2i Qm + B2j F̄2i 0 0 0 0 0
 
 
 −RT B2i F̄2j + L̄i C2j Qm − RT B2j F̄2i + L̄j C2i Qm 0 0 0 0 0
0 0 0 0 0 0
RT A − L̄i C2 , (1, 1)b = R−T L̄i D2 1 , (2, 1)b = RT B1 − L̄i D2 1 , give asymptotic exactness of the relaxation, considering no knowledge
and (1, 1)c = C1 Qm + D1 2 (F̄1 m + F̄2 i ). on the membership functions, and the last one proposes a method that
Relaxation A [6]: For all , m, i ∈ [1, r], j ∈ (i, r], min γ subject can reduce the number of LMI conditions by taking advantage of the
( m ) ( m ) ( m )
to Gi i > 0, Gi j + Gj i > 0. redundancy of descriptor systems.
Relaxation B [8]: For all , m, i ∈ [1, r], j ∈ (i, r], min γ subject Remark 10: In [34], it has been known that the conditions in [32]
to 2Gi i
( m ) ( m ) ( m ) ( m ) ( m ) ( m )
≥ Ḡi i , Gi j + Gj i ≥ Ḡi j , 0 < [Ḡi j ]r ×r , where are more relaxed than those in [29]. And it was proved in [13] that the
( m ) ( m ) ( m )T ( m ) conditions in [13] admit a great deal more freedom than those in [6]
Ḡj i = Ḡi j and Ḡi j = Ḡi j , i ≤ j.
and [8]. Particularly for r ≥ 3, more relaxed conditions were proposed
Relaxation C [13]: For all , m, i ∈ [1, r], j ∈ (i, r], min γ
( m ) ( m ) ( m ) ( m ) ( m ) ( m ) in [22] than those in [13] and [32]. Meanwhile, Sala and Ariño [33]
subject to Gi i ≥ Ḡi i , Gi j + Gj i ≥ Ḡi j + Ḡj i , 0 < mentioned that their own method does not provide better results than
( m ) ( m ) ( m ) T
[Ḡi j ]r ×r , where Ḡj i = Ḡi j , i ≤ j. in [13] for βi j = 0.25, i.e., the method with normal partitions is useful
Relaxation D [29]: For all , m, i, j ∈ [1, r], min γ subject to only in fuzzy models with three or more rules.
( m ) ( m ) ( m ) ( m )
Gi i > 0, 1/(r − 1)Gi i + (1/2)(Gi j + Gj i ) > 0, i = j. Remark 11: The derived BMI problem can be solved by other pos-
Relaxation E [32]: For all , m, h, i ∈ [1, r], j ∈ (i, r], min γ sible iterative algorithms, such as cone complementarity linearization
subject to X̄i > 0, Ti j h ≥ 0, Mi j = MiTj algorithm or D-K iteration algorithm. However, our interest is not in
 ( m ) ( m ) ( m )  making a comparison between our algorithm and other algorithms but
G1 1 − C1 h Ḡ1 2 h ··· Ḡ1 r h in introducing a simple method of solving the BMI problem within two
 ( m ) ( m ) ( m ) 
 Ḡ2 1 h G2 2 − C2 h ··· Ḡ2 r h  steps. Of course, one can use this method as a procedure for determining
0 <
 .. .. ..

 an initial point of such iterative algorithm.
..
 . . . .  Remark 12: As a simple attempt to improve the H∞ performance,
( m )
Ḡr 1 h
( m )
Ḡr 2 h ···
( m )
G2 2 − C2 h we can add a small variation to R when incorporating R (obtained by
 Lemma 4.1) in Theorem 1.
 Mi h , if i < h Remark 13: Similar to [23] and [28], we can also convert the PLMI
Ci h = Mh i , if i > h condition into a quasi-LMI condition with some tuning parameters.
 However, since this kind of approach yields a large matrix inequality
0, if h = i
condition unsuitable for our relaxation scheme, we did not employ it
where
in our study.
 1 ( m ) ( m )

 (Gi j + Gj i ) − Ti j h

 2

 if i < j V. NUMERICAL EXAMPLE

 + (Hi j h − HiTj h ) + 1 Wi j h ,
( m ) 2 To verify the performance of the proposed observer-based H∞ fuzzy
Ḡi j h =

 1 ( m ) ( m )
(Gi j + Gj i ) − Tj i h
control method, we consider a problem of determining the duty ratio



 2 of the pulsewidth modulation (PWM) buck converter (refer to [21]):

 1
if i > j
T
+ (Hj i h − Hj i h ) + Wj i h , β 1 1
2 (x1 (t) − x1 d )2 − x2 (t) − Vl (t)
ẋ1 (t) =
L L L
Mi j , if i = h or j = h
Wi j h = 1 1
0, if i = h and j = h. + (Vin + VD − RM x1 (t))τ (t) + x2 d
L L
Relaxation F [33]: Assume 0 ≤ θ̂i θ̂j ≤ βi j . For all , m, i ∈ 1 1
( m ) ( m ) ( m ) ẋ2 (t) = x1 (t) − (x2 (t) − β(x1 (t) − x1 d )2 )
[1, r], j ∈ (i, r], min γ subject to Gi i + Yi i − U ≥ Ḡi i , Gi j + C RC
( m ) ( m ) ( m ) ( m )
Gj i + Yi j − 2U ≥ Ḡi j + Ḡj i , and 0 < [Ḡi j ]r ×r , where 1
− Il (t) y(t) = x2 (t) (38)
r ( m ) ( m ) T C
U = i = 1 i ≤j ≤r βi j Yi j , = Ḡi j Ḡj i
, and 0 < YiTj = Yi j ,
i ≤ j. where x1 and x2 denote the inductor current and the voltage of the
( m ) ( m )
Relaxation G [22]: min γ subject to Gi i > Ḡi i i for all i ∈ [1, r], capacitor, x1 d and x2 d = R x1 d denote the desired state, and the duty
( m ) ( m ) ( m ) ( m ) ( m ) ( m )T
Gi i + Gi j + Gj i ≥ Ḡi i j + Ḡi j i + Ḡi i j for all i, j ∈ ratio τ (t) corresponds to the control input (see Fig. 1). By the map-
[1, r], j = i, Gi j
( m ) ( m )
+ Gi h + Gj i
( m ) ( m ) ( m )
+ Gj h + Gh i + Gh j ≥
( m ) ping ζ1 (t) = x1 (t) − x1 d and ζ2 (t) = x2 (t) − x2 d , the system (38) is
( m ) ( m ) ( m ) ( m )T ( m )T ( m )T converted into
Ḡi j h + Ḡi h j + Ḡj i h + Ḡi j h + Ḡi h j + Ḡj i h for all i ∈
( m )
[1, r − 2], j ∈ [i + 1, r − 1], h ∈ [j + 1, r], and 0 < [Ḡi h j ]r ×r for β 1 1
ζ̇1 (t) = ζ12 − ζ2 (t) − Vl (t)
all h ∈ [1, r]. L L L
Remark 9: Apart from Relaxations A–G, various results on the re- 1
laxation have appeared in the literature [30], [31], [35]. The first two + (Vin + VD − RM (ζ1 (t) + x1 d ))τ (t)
L
 
(QTm + Qm − X̄m ) ⊕ Zm 0 (∗) (∗)
 (∗)
 (∗) (∗) (∗) 

  r r
=  + (U − Ui ) + (Qi − Q ) −
(m )
Gij  (1, 1)a (1, 2)a (1, 1)b δ U − δi Qi
0
i i
 X̄i ⊕ (RT + R − Zi ) 
 (2, 1)a (2, 2)a (2, 1)b  i=1 i=1
[ (1, 1)c C1 ] D11 0 I
condition than Relaxations C–E. Here, the reason for not using Re-
laxations F and G is that these cannot apply to the case of r < 3 (see
Remark 10).
VI. CONCLUDING REMARKS

In this short paper, we addressed the problem of designing a fuzzy
observer-based H∞ controller for discrete-time T–S fuzzy systems. By
building a set of fuzzy control rules with premise variables different
Fig. 1. Equivalent circuit of a PWM buck converter. from those of the T–S fuzzy system, we enhanced the applicability of
the output-feedback controller.
TABLE I
COMPARISON OF γC ACCORDING TO VARIOUS β
REFERENCES
[1] T. Takagi and M. Sugeno, “Fuzzy identification of systems and its ap-
plications to modeling and control,” IEEE Trans. Syst., Man., Cybern.,
vol. SMC-15, no. 1, pp. 116–132, Jan. 1985.
[2] K. Tanaka and M. Sugeno, “Stability analysis and design of fuzzy control
systems,” Fuzzy Sets Syst., vol. 45, pp. 135–156, 1992.
[3] S. Boyd, L. E. Chauoi, E. Feron, and V. Balakrishan, Linear Matrix In-
equalities in System and Control Theory. Philadelphia, PA: SIAM, 1994.
[4] K. Tanaka, T. Ikeda, and H. O. Wang, “Robust stabilization of a class of
uncertain nonlinear systems via fuzzy control: Quadratic stabilization H∞
control theory, and linear matrix inequality,” IEEE Trans. Fuzzy Syst.,
1 β 1 1 vol. 4, no. 1, pp. 1–13, Oct. 1996.
ζ̇2 (t) = + ζ1 (t) ζ1 (t) − ζ2 (t) − Il (t) [5] X.-J. Ma, Z.-Q. Sun, and Y.-Y He, “Analysis and design of fuzzy controller
C RC RC C
and fuzzy observer,” IEEE Trans. Fuzzy Syst., vol. 6, no. 1, pp. 41–51,
z(t) = ζ2 (t) + 0.4Vl (t) ȳ(t) = ζ2 (t) + 0.3Vl (t) (39) Oct. 1998.
[6] K. Tanaka, T. Ikeda, and H. O. Wang, “Fuzzy regulators and fuzzy ob-
where Vl (t) and Il (t) denote the leakage voltage and current. In (39), servers: Relaxed stability conditions and LMI-based designs,” IEEE
Trans. Fuzzy Syst., vol. 6, no. 2, pp. 250–265, May 1998.
we can naturally regard the unmeasurable state ζ1 as a premise vari- [7] J. Yoneyama, M. Nishikawa, H. Katayama, and A. Ichikawa, “Output
able. Thus, given the assumption that ζ1 (t) ∈ [ρ1 , ρ2 ], the premise
2 stabilization of Takagi–Sugeno fuzzy systems,” Fuzzy Sets Syst., vol. 111,
variable ζ1 (t) can be rewritten as ζ1 (t) = i = 1 θi (ζ1 (t))ρi , where pp. 253–266, 2000.
θ1 = (ζ1 (t) − ρ2 )/(ρ1 − ρ2 ), θ2 = (ρ1 − ζ1 (t))/(ρ1 − ρ2 ), and the [8] E. Kim and H. Lee, “New approaches to relaxed quadratic stability con-
dition of fuzzy control systems,” IEEE Trans. Fuzzy Syst., vol. 8, no. 5,
resulting system model can be described as follows: pp. 523–534, Oct. 2000.
[9] J. Yoneyama, M. Nishikawa, H. Katayama, and A. Ichikawa, “Design of

2
output feedback controllers for Takagi–Sugeno fuzzy systems,” Fuzzy
ζ̇(t) = θi (ζ1 (t))(Ai ζ(t) + B1 w(t) + B2 i τ (t)) Sets Syst., vol. 121, pp. 127–148, 2001.
i= 1 [10] M. C. de Oliveira and R. E. Skelton, “Stability tests for constrained linear
systems,” in Perspectives in Robust Control (Lecture Notes in Control and
where Information Sciences), vol. 268, S. O. R. Moheimani, Ed. New York:
 β 1
 Springer-Verlag, 2001, pp. 241–257.
ρi − [11] D. J. Choi and P. Park, “H∞ state-feedback controller design for discrete-
 L L  −1/L 0 time fuzzy systems using fuzzy weighting-dependent Lyapunov func-
Ai =   B1 =
1 β 1 0 −1/C tions,” IEEE Trans. Fuzzy Syst., vol. 11, no. 2, pp. 271–278, Apr. 2003.
+ ρi − [12] S. K. Nguang and P. Shi, “H∞ fuzzy output feedback control design for
C RC RC
1 nonlinear systems: An LMI approach,” IEEE Trans. Fuzzy Syst., vol. 11,
no. 3, pp. 331–340, Jun. 2003.
(Vin + VD − RM (ρi + x1 d )) [13] L. Xiaodong and Z. Qingling, “New approaches to H∞ controller designs
B2 i = L .
0 based on fuzzy observers for T–S fuzzy systems via LMI,” Automatica,
vol. 39, pp. 1571–1582, 2003.
After all, we can obtain a discrete-time fuzzy system for (x1 d , x2 d ) = [14] J.-C. Lo and M.-L. Lin, “Robust H∞ nonlinear control via fuzzy static
output feedback,” IEEE Trans. Circuits Syst., vol. 50, no. 11, pp. 1494–
(2, 12) from the bilinear transformation with fast sampling time Ts = 1502, Nov. 2003.
0.0003 s, where L = 0.09858 mH, C = 0.2025 mF, R = 6 Ω, VD = [15] H. D. Tuan, P. Apkarian, T. Narikiyo, and M. Kanota, “New fuzzy control
0.82 V, Vin = 30 V, ρ1 = −2, ρ2 = 2, RM = 0.5 Ω, and the current- model and dynamic output feedback parallel distributed compensation,”
dependent source parameter β is changed from 0.04 to 0.12. Based on IEEE Trans. Fuzzy Syst., vol. 12, no. 1, pp. 13–21, Feb. 2004.
[16] J.-C. Lo and M.-L. Lin, “Observer-based robust H∞ control for fuzzy
the discrete-time fuzzy system, we first obtain R and L̄i , i ∈ [1, r], by systems using two-step procedure,” IEEE Trans. Fuzzy Syst., vol. 12,
solving the LMI-based problem in Lemma 4.1, where (β, γO ) = (0.04, no. 3, pp. 350–359, Jun. 2004.
1.0975), (0.043, 1.1043), (0.046, 1.1114), (0.048, 1.1163), (0.05, [17] T. M. Guerra and L. Vermeiren, “LMI-based relaxed nonquadratic stabi-
1.1211), (0.1, 1.264), and (0.12, 1.3345). Next, based on the obtained lization conditions for nonlinear systems in the Takagi–Sugeno’s form,”
L̄i and R, we solve the LMI-based optimization problem in Theorem 1, Automatica, vol. 40, pp. 823–829, 2004.
[18] D. K. Kim, P.-G. Park, and J. W. Ko, “Output-feedback H∞ control of
and then compare its result with those of Relaxations C–E (see systems over communication networks using a deterministic switching
Table I). Of course, since the solved problem is a nonconvex opti- system approach,” Automatica, vol. 40, pp. 1205–1212, 2004.
mization problem whose global optimal solution cannot be obtained [19] L. Fang, H. Lin, and P. J. Antsaklis, “Stabilization and performance anal-
easily by our two-step algorithm, we cannot make an absolute compari- ysis for a class of switched systems,” in Proc. 43rd IEEE Conf. Decis.
Control, Paradise Island, Bahamas, Dec. 2004, pp. 3265–3270.
son from the result shown in Table I. However, from Table I, we observe [20] S. Zhou, G. Feng, J. Lam, and S. Xu, “Robust H∞ control for discrete-
that under the same BMI solver, our relaxation method (Theorem 1) time fuzzy systems via basis-dependent Lyapunov functions,” Inf. Sci.,
is more effective at increasing the feasibility of the H∞ stabilization vol. 174, no. 3/4, pp. 197–217, 2005.
[21] K.-Y. Lian and J.-J. Liou, “Output tracking control for fuzzy systems [29] H. D. Tuan, P. Apkarian, T. Narikiyo, and Y. Yamamoto, “Parameterized
via output feedback design,” IEEE Trans. Fuzzy Syst., vol. 14, no. 5, linear matrix inequality techniques in fuzzy control system design,” IEEE
pp. 628–639, Oct. 2006. Trans. Fuzzy Syst., vol. 9, no. 2, pp. 324–332, Apr. 2001.
[22] C.-H. Fang, Y.-S. Liu, S.-W. Kau, L. Hong, and C.-H. Lee, “A new LMI- [30] A. Kruszewski, A. Sala, T.-M. Guerra, and C. Arino, “Sufficient and
based approach to relaxed quadratic stabilization of T–S fuzzy control asymptotic necessary conditions for the stabilization of Takagi–Sugeno
systems,” IEEE Trans. Fuzzy Syst., vol. 14, no. 3, pp. 386–397, Jun. model,” presented at the third IFAC Workshop AFNC 2007, Valenciennes,
2006. France.
[23] T. M. Guerra, A. Kruszewski, L. Vermeiren, and H. Tirmant, “Conditions [31] A. Sala and C. Ariño, “Asymptotically necessary and sufficient condi-
of output stabilization for nonlinear models in the Takagi–Sugeno’s form,” tions for stability and performance in fuzzy control: Applications of
Fuzzy Sets Syst., vol. 157, pp. 1248–1259, 2006. Polya’s theorem,” Fuzzy Sets Syst., vol. 158, no. 24, pp. 2671–2686,
[24] J. Lam and S. Zhou, “Dynamic output feedback H∞ control of discrete- 2007.
time fuzzy systems: A fuzzy-basis-dependent Lyapunov function ap- [32] M. C. M. Teixeira, E. Assuncão, and R. G. Avellar, “On relaxed LMI-
proach,” Int. J. Syst. Sci., vol. 38, no. 1, pp. 25–37, Jan. 2007. based designs for fuzzy regulators and fuzzy observers,” IEEE Trans.
[25] S.-W. Kau, H.-J. Lee, C.-M. Yang, C.-H. Lee, L. Hong, and C.-H. Fang, Fuzzy Syst., vol. 11, no. 5, pp. 613–623, Oct. 2003.
“Robust H∞ fuzzy static output feedback control of T–S fuzzy systems [33] A. Sala and C. Ariño, “Relaxed stability and performance conditions for
with parametric uncertainties,” Fuzzy Sets Syst., vol. 158, pp. 135–146, Takagi–Sugeno fuzzy systems with knowledge on membership function
2007. overlap,” IEEE Trans. Syst., Man., Cybern. B, Cybern., vol. 37, no. 3,
[26] S. Zhou, J. Lam, and W. X. Zheng, “Control design for fuzzy systems based pp. 727–732, Jun. 2007.
on relaxed nonquadratic stability and H∞ performance conditions,” IEEE [34] M. C. M. Teixiera, M. F. R. Catharino, E. Assuncao, and E. R.
Trans. Fuzzy Syst., vol. 15, no. 2, pp. 188–199, Apr. 2007. M. Daruichi, “A comparison study between two relaxed LMI-based fuzzy
[27] A. Sala, T. M. Guerra, and R. Babuška, “Perspectives of fuzzy systems control designs,” in Proc. IEEE Int. Conf. Fuzzy Syst., 2005, pp. 969–
and control,” Fuzzy Sets Syst., vol. 156, pp. 432–444, 2005. 972.
[28] A. Kruszewski and T.-M. Guerra, “Conditions of output stabilization for [35] K. Tanaka, H. Ohtake, and H. O. Wang, “A descriptor system approach to
uncertain discrete fuzzy models,” presented at the IFAC Triennial World fuzzy control system design via fuzzy Lyapunov functions,” IEEE Trans.
Congr., Prague, Czech Republic, 2005. Fuzzy Syst., vol. 15, no. 3, pp. 333–341, Jun. 2007.

ITFS Volume 17 Issue 2 April 2009

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

ITFS Volume 17 Issue 2 April 2009

Uploaded by

Copyright:

Available Formats

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 17, NO.

2, APRIL 2009 253

Piecewise Fuzzy Anti-Windup Dynamic Output

1063-6706/$25.00 © 2009 IEEE

As shown in Fig. 1, the CSTR (46) has the following three

B. H∞ Output Feedback Control of Constrained CSTR via

Fig. 2. Open-loop response for +5 K (a) and −5 K (b) step changes in T c of C1 = C2 = C3 = C = [ 0 1 ]

tude bound ρa = 40 and the rate bound ρr = 5. The space of

Fig. 5. Piecewise fuzzy anti-windup dynamic output feedback control of

and the performance shown in Fig. 5 is much better than those

H∞ Fuzzy Control of Nonlinear Systems Under

Abstract—This paper investigates the problem of H∞ fuzzy con-

Index Terms—Basis-dependent Lyapunov function, H∞ control,

study how to design control systems with the simultaneous

C. Communication Links Then, the problem to be addressed in this paper is expressed

Based on this, we have

closed-loop fuzzy system in (5), we have = h2i ΓTii P̄ Γii − P xk + hi hj (14)

E{xTk +1 P xk +1 } − E{xTk P xk } + E{zkT zk } − γ 2 wkT wk ≤ 0. where

Furthermore, if the previous conditions have feasible solutions, Pj > 0, j = 1, . . . , r, satisfying

ΓTii P̄j Γii − Pi < 0, i = 1, . . . , r, j = 1, . . . , r   

 T E {Vk+1 | ξk} +E zkT zk  ξk − γ 2 wkT wk − ξkT hi Pi ξk ≤ 0.

basic idea in the CCL algorithm is that if the LMI [ PI LI ] ≥ 0

H∞ performance γ, if there exist a matrix P > 0, scalars εi , εij , B. Discussions

Model Rule 2 : IF xk (1) is h2 (xk (1))

Fig. 6. Closed-loop state variables when ē = 0.8.

Fig. 7. Inverted pendulum controlled by a dc motor.

Huijun Gao (M’06) was born in Heilongjiang,

An Adaptive Consensus Support Model for Group

Abstract—Different consensus models for group decision-

2) pali is the proximity between the preference values of

Fig. 3. ACSS model in a multigranular fuzzy linguistic context.

order to make his/her preferences closer to the collective

to denote each fuzzy set p̃lk i

B. Computing the Consensus Degree and Control of the

the alternatives and only the preference values in disagreement

P = {(l, k)|cplk < cp, l, k = 1, . . . , n}. # PREFECH Li ≤ # PREFECH Vi L .

i ) − cv(p̃c )) < 0, then the expert ei should in-

crease the assessment associated with the pair of alternatives

crease the assessment associated with the pair of alternatives

modify the assessment associated with the pair of alternatives

1) Experts e3 and e7 use label set A TABLE II

b2 = (0.17, 0.33, 0.5) b3 = (0.33, 0.5, 0.67)

b) According to rule DR2, the experts are required to de-

1 = {(3, 1), (3, 2)}

iii) Relation: cr = 0.756. are.

H∞ Fuzzy Filtering of Nonlinear Systems

1063-6706/$25.00 © 2009 IEEE

increasing attention has been paid to the synthesis problems.

II. PROBLEM FORMULATION A(h) = hi Ai , B(h) = hi Bi , C(h) = hi Ci

B. Filter Before proceeding further, we first introduce the following

i = 1, . . . , r. (5) Then, the problem to be addressed in this paper is expressed

where Here, (15) implies

implies that = x̄T0 (λm in (−Φ))−1 max(P (h))x̄0

E{Vk +1 } − E{Vk } + E z̄kT z̄k − γ 2 wkT wk ≤ 0.

Proof: Suppose that there exist matrices R, S, W, Āf i , B̄f i ,

a given positive constant γ, if there exist matrices Qi =

To show the effectiveness of the obtained results, we assume the

control of fuzzy systems with simultaneous consideration of

Ke Chen received the B.S. degree in mathematics in

An Interval Type-2 Fuzzy Rough Set Model

1063-6706/$25.00 © 2009 IEEE

where ★ is the minimum or algebraic product t-norm

ΓTii P̄j Γii − Pi < 0, i = 1, . . . , r, j = 1, . . . , r

T E {Vk+1 | ξk} +E zkT zk ξk − γ 2 wkT wk − ξkT hi Pi ξk ≤ 0.

Proof: Suppose that there exist matrices R, S, W, Āf i , B̄f i ,

a given positive constant γ, if there exist matrices Qi =

≡ µÃ (x) µB̃ (x), x∈U (7)

where ∧ [(1 − µD f˜(x) (y )) ∨ µD A(y )]

⇒ ∨ rR̃ (x, y) = 1. ∨ [µD f˜(x) (y ) ∧ µD A(y )]

According to Proposition 7, = ∨ [µD f˜(x) (y) ∧ µD A(y)] ∨ [1 ∧ µD A(y )]

3) R̃ is serial ⇒ ∃y ∈ W such that lR̃ (x, y ) = rR̃ (x, y ) =

(E, E ), (E, F ), (E, P ), (F, E ), (F, F ), (F, P ), (P, E ), (P, F ), (P, P )