Asian Journal of Control, Vol. 10, No. 1, pp.

74 87, January 2008

Published online in Wiley InterScience (www.interscience.wiley.com) DOI: 10.1002/asjc.008

FAULT TOLERANT CONTROL OF FEEDBACK LINEARIZABLE

SYSTEMS WITH STUCK ACTUATORS
Min Wang and Donghua Zhou

ABSTRACT

Fault tolerant control of feedback linearizable systems with stuck actua-

tors is studied in this paper. Once stuck, the faulty actuators cannot respond to
the control inputs and have fixed constant outputs. The fault considered here
is severe in the sense that the faulty system is no longer feedback linearizable.
By constructing a nonlinear transformation and extending Chen’s disturbance
observer to a multi-dimensional case, a generalized disturbance observer is
developed for estimating the unknown constant outputs of the faulty actua-
tors. Then, through integrating the generalized disturbance observer with a
controller constructed by the famous cascade design method, a fault tolerant
controller is obtained. It is proven that the fault tolerant controller ensures
not only boundedness of the state of the faulty system but also satisfactory
output performance. Finally, computer simulations are done using an inverted
pendulum model to verify the effectiveness of the proposed method.

Key Words: Feedback linearizable systems, stuck actuators, fault tolerant

control, disturbance observer.

I. INTRODUCTION Developed in the 1970s, FTC has experienced

Increasing safety and performance requirements
in modern industry lead to more and more complex
control systems, which make faults inevitable. When a
fault occurs in a control system such as a malfunction
in some actuator or sensor, the controller may not act as
expected, which usually causes performance degrada-
tion of the system or even makes the system unstable.
So it is of practical interest to design controllers that
can guarantee the system stability and maintain satis-
factory performance when the fault occurs. This is the
main objective of fault tolerant control (FTC) [1].
Manuscript received 20 January 2006; revised 15 September
2006; accepted 26 December 2006.
Min Wang and Donghua Zhou (e-mail: zdh@mail.tsinghua.
edu.cn) are with the Department of Automation, Tsinghua
University, Beijing 100084, China.
This work was supported by NSFC (Grant No. 60736026,
60574084), 863 program (Grant No. 2006AA04Z428), and Na-
tional 973 program (Grant No. 2002CB312200) of China.
great progress in the past three decades [1]. Generally,
the fault information, such as where the fault is and
how bad the fault is, can be very useful information in
designing a fault tolerant controller. According to how
this fault information is used, there are two types of
FTC methods: passive and active methods [1]. Assum-
ing the possible faults are known a priori, the passive
method takes into account of all these possible faults
in the design stage and does not change the controller
when the fault occurs [2–7]. An active method usually
uses a fault detection and isolation (FDI) unit to collect
the fault information and changes the controller accord-
ing to the fault [8–28]. Intuitively, the active method
uses more information, thus, can achieve better perfor-
mance than the passive method. One typical active FTC
system widely used in the literature contains a normal
controller, a FDI unit, and a fault estimation and con-
troller redesign module [8, 9, 11–15, 19, 20, 23–28] as
illustrated in Fig. 1. When a fault is detected, the FDI
unit determines where the fault is, i.e. fault isolation,
then the online fault estimation algorithm is activated.
Based on this estimation, the controller is redesigned

q 2008 John Wiley and Sons Asia Pte Ltd and Chinese Automatic Control Society
 M. Wang and D. Zhou: Fault Tolerant Control of Feedback Linearizable Systems with Stuck Actuators
Fault
Setpoint u y
Controller Plant
Controller Fault detection
redesign and isolation
Online fault
estimation
Fig. 1. A typical active FTC system.
to compensate for the effect of the fault. In this paper,
we also adopt this typical active FTC system.
One important fault scenario is that of stuck
actuators, which is common in flight control and chem-
ical processes. Once stuck, the faulty actuators can no
longer respond to the control inputs and have fixed
constant outputs. Thus, the faulty actuators not only
reduce some freedom in the control ability but also act as
persistent external disturbances [9]. The related FTC
problem is very challenging in the sense that the remain-
ing healthy actuators must compensate for the effect of
the faulty ones and maintain the original performance
as much as possible at the same time. Generally speak-
ing, there are mainly two difficulties in the related FTC
problem. Due to the additional cost for extra compo-
nents, in practice, the outputs of actuators are usually
not measured. Thus, the constant outputs of the stuck ac-
tuators, i.e. stuck positions, are difficult to obtain. This
is the first difficulty. The second one is: given the stuck
positions, designing a fault tolerant controller is still
not an easy task. This is especially true when the struc-
ture of the original system is dramatically changed by
the faulty actuators. Boskovic [8] and Tao [21] solved
the FTC problem by adaptive control techniques. Chen
[9] presented an iterative learning observer to esti-
mate the unknown stuck positions. All of them assume
a matching condition, i.e., the correct outputs of the
stuck actuators can be completely generated by the
remaining healthy ones, which makes the fault tolerant
controller design very easy when the stuck positions
are available. When this matching condition does not
hold, their methods cannot be applied.
Feedback linearizable systems (FLS) are a simple
class of nonlinear systems, which can be turned into lin-
ear systems by appropriate feedback control [29]. Based
on the linear systems, the conventional linear controller
design techniques can be used. Therefore, the control
problem of FLS is relatively easier compared with that
of the general nonlinear systems. Although simple, FLS
is a very important class of nonlinear systems, which is
common in physical nonlinear systems such as robotics,
q 2008 John Wiley and Sons Asia Pte Ltd and Chinese Automatic Control Society
75
electric machines, autopilot control, and flight mechan-
ics [30]. Thus, the FTC problem for FLS with stuck
actuators is of practical interest. To solve this problem,
Tao [22] presented an adaptive compensation frame-
work which required no FDI unit. Global stability is
guaranteed by adjusting the controller’s parameters on-
line. Liang [4] proposed a passive approach. Separating
the actuators into a potentially faulty part and non-faulty
part, the main idea is to design a basic controller by the
use of the non-faulty ones. When the potentially faulty
actuators are healthy, they improve the performance;
and when they are stuck, their effects can be compen-
sated for by the basic controller. Another method for
resolving the same problem using optimization tech-
niques was also investigated in [10]. However, none of
these methods considered the two above mentioned dif-
ficulties simultaneously. Reference [10] assumed that
the stuck positions were known, which made the first
difficulty trivial, while references [4] and [22] assumed
that the faulty system is still feedback linearizable, i.e.
satisfies the matching condition, making the second dif-
ficulty trivial. Therefore, further work on this problem
is necessary.
In this paper, we consider an FTC problem for
FLS with stuck actuators. The two above mentioned
difficulties are considered simultaneously, i.e., we as-
sume that the stuck positions are unknown and the faulty
system is no longer feedback linearizable. The typical
active FTC system depicted in Fig. 1 is used. Due to
the great development of fault detection and diagnosis
techniques, in this paper, we assume that the FDI unit
has been successfully designed. The aim of this paper
is to design the online fault estimation algorithm and
redesign the controller for the faulty system to com-
pensate for the effect of the stuck actuators. Since the
faulty system is not feedback linearizable, the controller
design becomes difficult. To deal with the difficulty, a
nonlinear transformation is first constructed using ge-
ometric methods, which can transform the faulty sys-
tem into a cascade system. Under an assumption of the
feasibility of the famous cascade design method [31],
a controller is then designed for the faulty system. On
the other hand, based on the nonlinear transformation,
a generalized disturbance observer (GDO) is presented
by the nonlinear observer design techniques to estimate
the unknown stuck positions. Finally, integrating the
controller with the GDO, a fault tolerant controller is
obtained. Compared with the existing work already re-
ported in the literature, the contribution of this paper is
mainly in the following three aspects:
1. A GDO is newly developed which can estimate
the stuck positions accurately. Moreover, its de-
76 Asian Journal of Control, Vol. 10, No. 1, pp. 74 87, January 2008

sign is systematic and simple. As a matter of fact, B = diag(B1 , B2 , . . . , Bm )

the proposed GDO is a multi-dimensional exten- C = diag(C1 , C2 , . . . , Cm ).
sion of Chen’s disturbance observer [32]. This
extension is nontrivial and greatly enlarges the ap- The linear subsystem (Ai , Bi , Ci ) (1≤i≤m) takes the
plication field of Chen’s result. Brunovsky canonical form [29]:
⎡ ⎤
2. The two difficulties in FTC for control systems with 0 1 0 ··· 0
stuck actuators are considered simultaneously, which ⎢0 0 1 ··· 0⎥
⎢ ⎥
makes the proposed approach less conservative and ⎢. . . . . ⎥
⎢ .
Ai = ⎢ . . .. . . . .⎥.
more practical for real systems. ⎥
3. The boundedness of the state and the satisfactory ⎢ ⎥
⎣0 0 0 ··· 1⎦
output performance of the resultant FTC system are
rigorously proved. 0 0 0 ··· 0 ri ×ri
⎡ ⎤
0
The remaining parts of this paper are organized ⎢0⎥
⎢ ⎥
as follows. Section II introduces the problem formula- ⎢.⎥
tion. Section III contains the main results of this paper Bi = ⎢ ⎥
⎢ .. ⎥
including the construction of the nonlinear transforma- ⎢ ⎥
⎣0⎦
tion, the design of the GDO, the controller design by
the cascade design method with the true stuck positions 1 ri ×1
and the fault tolerant controller design as well as the 
m
Ci = [1 0 0 · · · 0]1×ri , ri = n.
theoretical proofs of its performance. Section IV i=1
presents simulation results while Section V concludes
the paper with final remarks. The canonical form (1) can also be written in a
Notations. A T and A−1 are, respectively, the more compact form:
transpose and the inverse of a matrix A. In denotes ẋ = f (x) + g(x)u
an identity matrix of dimension n. 0m×n stands for an (2)
y = Cx
m × n zero matrix.
·
denotes the Euclidean norm
or its induced norm. col(x, y) implies that the two where f (x) = Ax + Bb(x) and g(x) = Ba(x) = (g1 (x),
vectors x and y are stacked to produce a new vector. . . . , gm (x)).
diag(A1 , . . . , Ak ) represents a block diagonal matrix Note that y is the output to be controlled. It doesn’t
with k matrices Ai (1≤i≤k) on its main diagonal. mean that only y is measurable. Throughout the present
paper, the following condition is assumed.
II. PROBLEM FORMULATION Assumption 1. The state x of system (1) is measurable.
Given the normal system (1), the normal control
2.1 A canonical form of FLS and the normal objective is to design a controller such that:
control objective
1. The state of the closed-loop system consisting of the
As is commonly known, FLS can be transformed
controller and the normal system (1) is bounded; The
to a canonical form by a standard nonlinear transforma-
controlled output of the closed-loop system satisfies
tion [29]. Thus, in this paper, we consider the canonical
limt→∞ y(t) = yc .
form as follows:
ẋ = Ax + B[b(x) + a(x)u] yc = col(y1c , . . . , ymc ) is the constant setpoint. The sub-
(1) script c means ‘constant’. The normal control objective
y = Cx
can be easily achieved by the feedback linearization
where x ∈ D0 ⊆ R n is the state, u = col(u 1 , . . . , u m ) ∈ techniques. Thus, no details of the normal controller
R m is the control input, and y = col(y1 , . . . , ym ) ∈ R m design will be given in this paper.
is the controlled output. b(x) : D0 → R m and a(x) : D0
→ R m×m are sufficiently smooth over an open set D0 2.2 The faulty system description and the FTC
with 0 ∈ D0 . b(0) = 0 and a(x) is nonsingular for all objective
x ∈ D0 . Hence, the origin x = 0 is an equilibrium of the The aim of this paper is to design the fault esti-
open-loop system (1) when u = 0. A ∈ R n×n , B ∈ R n×m mation and controller redesign module. Throughout the
and C ∈ R m×n have the following structure, paper, the FDI unit is assumed to be successfully de-
A = diag(A1 , A2 , . . . , Am ) signed. Without loss of generality, we suppose that the

q 2008 John Wiley and Sons Asia Pte Ltd and Chinese Automatic Control Society
 M. Wang and D. Zhou: Fault Tolerant Control of Feedback Linearizable Systems with Stuck Actuators
first p (1≤ p<m) actuators are stuck at the time t = t f
while the remaining (m − p) actuators are healthy. The
faulty system can be modeled as:
ẋ = f (x) + g F (x)
+ g H (x)u H
y = Cx
where
= col(
1 , . . . ,
p ) ∈
⊆ R p is an unknown
constant vector which represents the constant outputs
of the stuck actuators. g F (x) = (g1 (x), . . . , g p (x)),
g H (x) = (g p+1 (x), . . . , gm (x)) and u H = col(u p+1 ,
. . . , u m ). The subscript F means ‘faulty’ while H
means ‘healthy’.
One necessary condition of a nonlinear system
to be feedback linearizable is that the number of the
independent control inputs equals the number of the
independent controlled outputs [29]. When the first p
actuators are stuck, the number of the independent con-
trol inputs is less than m. Therefore, the faulty system is
no longer feedback linearizable with the m independent
controlled outputs. The normal control objective is no
longer achievable due to this dramatic structure change.
Only most (m − p) priority outputs can be arbitrarily
controlled while p secondary outputs have to be given
up as analyzed in [10]. Without loss of generality, we
suppose that the last (m − p) outputs of system (3) are
priority outputs.
Given the faulty system (3), the FTC objective is
to design a controller such that:
1. The state of the closed-loop system consisting of the
controller and the faulty system (3) is bounded; The
last (m − p) controlled outputs y H of the closed-
loop system satisfies limt→∞ y H (t) = y H c where
y H c = col(y( p+1)c , . . . , ymc ).
Remark 1. The number of the stuck actuators greatly
influences the feasibility of the FTC objective. Partic-
ularly, when all the actuators are stuck, the faulty sys-
tem is no longer controllable and no FTC methods will
be feasible. Setting the FTC objective as the above, we
do not mean that the objective can always be achieved.
Alternatively, we just aim to find out when the FTC
objective is achievable and how to achieve it. Con-
trollers that can achieve this FTC objective are said to
be fault tolerant.
If the FTC objective is achieved, the equilibrium
of the closed-loop system consisting of the fault toler-
ant controller and the faulty system (3) must be in the
following form:
xe (
) = col(y1 (
), 0r1 −1 , . . . , y p (
), 0r p −1 ,
(6)
y( p+1)c , 0r p+1 −1 , . . . , ymc , 0rm −1 )
q 2008 John Wiley and Sons Asia Pte Ltd and Chinese Automatic Control Society
77
where yi (
) : R p → R (1≤i≤ p) is any function such
that xe (
) is an equilibrium. If a controller is designed
such that the state of the closed-loop system consisting
of the controller and the faulty system (3) is bounded
and limt→∞ x(t) = xe (
), then the FTC objective will
(3)
be achieved. Motivated by this observation, all the ef-
forts in the sequence are devoted to achieving this goal.
III. MAIN RESULTS
3.1 Construction of a nonlinear transformation
In this subsection, under an involutive condition,
a nonlinear transformation is constructed, which can
transform the faulty system (3) to a cascade system.
The aim of this transformation is twofold. First, based
on the transformation, Chen’s disturbance observer [32]
can be extended to a multi-dimensional situation for
constant disturbance, which plays a key role in the active
FTC system. Second, the faulty system is no longer
feedback linearizable, which makes the controller de-
sign difficult. A cascade form makes the cascade design
possible, thus facilitates the controller design. To begin
with, an involutive condition on g(x) is supposed.
Assumption 2. The distribution i = span{gi (x), . . . ,
gm (x)} is involutive over D0 for all i (2≤i≤ p +1). (For
the definition of involutive, see A1.)
Lemma 1. If Assumption 2 holds, then there exists
a nonlinear transformation z = col(z 1 , z 2 ) = (x) =
col(1 (x), 2 (x)) defined on a compact set D ⊆ D0 ,
which transforms the faulty system (3) to a cascade
system:
ż 1 = f 1 (z) + g1 (z)

ż 2 = f 2 (z) + g2F (z)
+ g2H (z)u H (5)
y = h(z)
where z 1 = 1 (x) ∈ R n−m+ p and z 2 = 2 (x) ∈ R m− p .
f 1 (z) ∈ R n−m+ p , f 2 (z) ∈ R m− p , g1 (z) ∈ R (n−m+ p)× p ,
g2F (z) ∈ R (m− p)× p and g2H (z) ∈ R (m− p)×(m− p) are as
follows.

f 1 (z) *(x)
= f (x)
f 2 (z) *x x=−1 (z)
⎡ ⎤
*1 (x)
f (x)
⎢ *x ⎥
=⎢
⎣ * (x)
⎥
⎦
2
f (x)
(4) *x x=−1 (z)
78 Asian Journal of Control, Vol. 10, No. 1, pp. 74 87, January 2008


g1 (z) 0 Let x ∈ R n−m denote all the entries of x except
the last state of every subsystem (Ai , Bi , Ci ) (1≤i≤m)
g2F (z) g2H (z) of system (1). By the special structure of system (1),

we have:
*(x)

= g(x)

*x
*x x=−1 (z) [g1 (x), . . . , gm (x)] = 0. (12)
⎡ ⎤
*x
*1 (x) *1 (x)
Hence, the entries of x are (n − m) solutions to the first
g (x) g (x)

⎢ *x F
*x
H ⎥
equation in (11).
=⎣⎢ ⎥

*2 (x) *2 (x) ⎦

Next, for the second equation in (11), 2 is invo-

g F (x) g H (x)
lutive and nonsingular with rank (m −1). From Froben-
*x *x x=−1 (z)
nius theorem (see A3), the equation has (n − m + 1)
(7)
independent solutions (For independent functions, see
A3). Note that the entries of x are its (n − m) indepen-
h(z)=C x|x=−1 (z) . (8) dent solutions. One more independent solution T1 (x)
exists over a set D1 ⊆ D0 .
Moreover, g2H (z) is nonsingular for all z ∈ (D) and Continue in this way until the last equation in (11).
g1 (z) has the following form: Since i is involutive and nonsingular with rank (m −


*1 (x)
0 i + 1), the ith equation in (11) must have (n − m + i −
g1 (z) =

g F (x)
= 1) independent solutions. One more solution Ti−1 (x)
*x x=−1 (z) gT F (z) (3≤i≤ p+1), which is independent of the (n−m+i −2)
(9) solutions to the (i − 1)th equation, i.e., the entries of x
and T1 (x), . . . , Ti−2 (x), exists over a set Di−1 ⊆ Di−2 .
where gT F (z) ∈ R p× p is a low triangular nonsingular Define T (x) =: col(T1 (x), . . . , T p (x)). By con-
matrix over (D). struction, the entries of T (x) are independent over D p .
Moreover,
Proof. To prove the first part of the lemma, a nonlinear
transformation (x) satisfying the following equation *Ti (x)
g j (x) = 0 (∀i, j, 1≤i≤ p, i + 1≤ j≤m).
is constructed. *x
*1 (x) (13)
g H (x) = 0 (10)
*x On the other hand, a function S(x) : R n → R m− p
can always be found such that the derivative of
Since a(x) is nonsingular, g(x) is full column
(x) = col(x, T (x), S(x)) ∈ R n is nonsingular over a
rank. By the definition of nonsingular distributions
compact set D ⊆ D p . Define a nonlinear transforma-
(see A2; note the difference from nonsingular matri-
tion z = col(z 1 , z 2 ) = (x) = col(1 (x), 2 (x)) over
ces), i (2≤i≤ p + 1) is nonsingular. Together with
D where z 1 = 1 (x) = col(x, T (x)) ∈ R n−m+ p and
Assumption 2, we conclude that i is involutive and
z 2 = 2 (x) = S(x) ∈ R m− p . Then, (10) holds by (12)
nonsingular with rank (m − i + 1) over D0 for all
and (13). Thus, the first part of the lemma is proved.
i (2≤i≤ p + 1).
By the full column rank of g(x) as well as the
Consider the following equations with an un-
low triangular form of the left part in (7), g1 (z) and
known function T (x) : R n → R.
g2H (z) are full column rank over (D). Hence, g2H (z)
⎧
⎪ *T (x) is nonsingular. As for g1 (z), the zero submatrix of g1 (z)
⎪
⎪ [g1 (x), . . . , gm (x)] = 0
⎪
⎪ *x in (9) can be confirmed by (12). Together with the full
⎪
⎪
⎪
⎪ column rank of g1 (z), the nonsingularity of gT F (z) can
⎪
⎪ *T (x)
⎨ [g2 (x), . . . , gm (x)] = 0 be inferred. Moreover, notice that
*x (11)

⎪ *T (x)

⎪
⎪ .. gT F (z) = g F (x)

⎪
⎪ . *x
⎪
⎪ x=−1 (z)
⎪
⎪

⎪  
⎩ *T (x) [g p+1 (x), . . . , gm (x)] = 0.
⎪ *Ti (x)

*x = g j (x)
. (14)
*x p× p

x= −1
(z)
It is obvious that the solutions to the ith equation in
(11) are also solutions to the (i + 1)th equation. Its low triangularity can be verified by (13).

q 2008 John Wiley and Sons Asia Pte Ltd and Chinese Automatic Control Society
 M. Wang and D. Zhou: Fault Tolerant Control of Feedback Linearizable Systems with Stuck Actuators 79

i(i−1)
3.2 Design of the generalized disturbance observer Define gTi· F (z) =: col(gTi1F (z), . . . , gT F (z)) ∈
Chen [32] proposed a disturbance observer for R i−1 and i =: diag(1 , . . . , i ), 1≤i≤ p. Let gTi F (z)
one-dimensional disturbance. In this subsection, based (1≤i≤ p) denote the ith principal submatrix of gT F (z).
on the cascade system (5) in Section 3.1, we extend his Thus, gT F (z) and Q can be rewritten as follows:
result to a multi-dimensional situation and propose a ⎡ p−1
⎤
GDO to estimate the stuck positions of the faulty actu- gT F (z) 0
ators. This extension is nontrivial, as will be shown in gT F (z) = ⎣ ⎦
p· pp
what follows. (gT F (z))T gT F (z)
Let z T denote the vector consisting of the last p ⎡ ⎤
p·
entries of z 1 in (5). Its dynamics is: Q p−1  p gT F (z)
Q=⎣ ⎦ (20)
ż T = f T (z) + gT F (z)
(15) p·
 p (gT F (z))T
pp
2 p gT F (z)
where f T (z) is the vector consisting of the last p entries
of f 1 (z) and gT F (z) is given by (9). p−1
where Q p−1 =  p−1 gT F (z) + ( p−1 gT F (z))T .
p−1
Design the GDO as follows: Using Schur complements (see A4) for (20), we
have:
˙ = −gT F (z) − [gT F (z)z T + f T (z)] pp
(16) Q >  p I p ⇔ 2 p gT F (z)> p


ˆ =  + z T
p· p· (21)
where  ∈ R p is the state of the GDO,
ˆ ∈ R p is the gT F (z)(gT F (z))T
Q p−1 >  p I p−1 + 2p .
estimate of
and  ∈ R p× p is a constant observer pp
2 p gT F (z) −  p
gain which will be designed later. The estimation error


˜ =:
−
ˆ satisfies: Choosing  p =  p /d 0p , from (19) we have:
˙˜ = −gT F (z)


˜ (17) pp
 p gT F (z)≥ p . (22)
Lemma 2. If Assumptions 1–2 hold and the state z of
the cascade system (5) is bounded by (D), then a By  p =  p /d 0p and (22), the following sufficient con-
constant observer gain  exists such that the estimation dition can be obtained:
error of the GDO (16) exponentially converges to zero.
Q> p I p ⇐ Q p−1 > p I p−1
Proof. If a constant matrix  exists such that:
p p· p·
Q =: gT F (z) + gTT F (z)T > p I p (18) + gT F (z)(gT F (z))T . (23)
(d 0p )2
where  p >0 is a pre-chosen positive constant, then
the error system (17) will be exponentially convergent p·
From the boundedness of gT F (z) on (D), a constant
to zero. To prove Lemma 2, such a constant matrix p·
 p−1 exists such that  p−1 I p−1 > p /(d 0p )2 gT F (z)
 is constructed. In the rest of the proof, we only p·
consider the case of constant diagonal matrices, i.e., (gT F (z))T . Consequently,
 = diag(1 , 2 , . . . ,  p ).
ij Q> p I p ⇐ Q p−1 >( p−1 +  p )I p−1 . (24)
Let gT F (z) (1≤i≤ p, 1≤ j≤ p) denote the (i, j)th
element of gT F (z). Since gT F (z) is a low triangular Hence, to make (18) hold, it is sufficient to design  p−1
nonsingular matrix on (D), as pointed out in Lemma such that Q p−1 >( p−1 +  p )I p−1 .
1, its diagonal elements are nonzero on (D). Without Continue in this way until Q 1 = 21 gT11F (z) ∈ R.
loss of generality, suppose gTii F (z)>0. Thus, they can p
Taking i = (1/di0 ) k=i k (2≤i≤ p), we have:
be divided as:

gTii F (z) = di0 + di1 (z), 1≤i≤ p (19) 

p
Qi > k Ii ⇐ Q i−1
k=i
where the constant di0 >0 and di1 (z)≥0. di0 can be chosen
as the minimum of gTii F (z) on (D), which cannot be 
p
> k Ii−1 , 2≤i≤ p (25)
zero due to the compactness of (D). k=i−1

q 2008 John Wiley and Sons Asia Pte Ltd and Chinese Automatic Control Society
80 Asian Journal of Control, Vol. 10, No. 1, pp. 74 87, January 2008
where Q i = i gTi F (z) + (i gTi F (z))T . {i } (1≤i≤ p)
is a constant sequence satisfying
gTi· F (z)(gTi· F (z))T 
p
i−1 Ii−1 > × k ,
(di0 )2 k=i
From (25), it follows that Q> p I p holds if
p
Q 1 > k=1 k . The latter can always be satisfied by
choosing 1 sufficiently large. Thus, an observer gain
 has been constructed such that (18) holds, which
completes the proof.

Remark 2.
(i) It can be verified that the GDO’s estimation error
satisfies

(t)
≤

˜
(0)

˜ exp(− p t/2). A larger  p
leads to faster convergence speed. On the other
hand, it also leads to a larger i (1≤i≤ p), which
results in high sensitivity to the measurement noise.
Therefore, a trade-off should be made when choos-
ing  p .
(ii) In the proof, the existence of a feasible observer
gain is proved by constructing a diagonal one. Nev-
ertheless, better performance may be achieved by
relaxing the diagonal form restriction, which de-
serves further research.
3.3 Controller design with the true stuck positions
and stability analysis
Consider the cascade system (5). Define z e (
) =:
col(z 1e (
), z 2e (
)) = (xe (
)) where z 1e (
) ∈
R n−m+ p and z 2e (
) ∈ R m− p . Throughout this paper,
we assume that z e (
) is in the interior of (D). If the
state z of the cascade system (5) is controlled to z e (
)
and is bounded, then the state x of the faulty system
(3) is controlled to xe (
), and the FTC objective will
be achieved as analyzed in Section II. In this subsec-
tion, based on the cascade system (5), a controller is
designed to achieve this goal under an assumption of
the feasibility of the cascade design method [31]. The
true stuck positions are used which will be replaced by
their estimates in Section 3.4.
Define z̃ =: z − z e (
) = col(z̃ 1 , z̃ 2 ) where z̃ 1 =
z 1 − z 1e (
) and z̃ 2 = z 2 − z 2e (
). The transformed
faulty system (5) can be rewritten as:
z̃˙ 1 = f 1 (z̃ + z e (
)) + g1 (z̃ + z e (
))

z̃˙ 2 = f 2 (z̃ + z e (
)) + g2F (z̃ + z e (
))

+ g2H (z̃ + z e (
))u H .
q 2008 John Wiley and Sons Asia Pte Ltd and Chinese Automatic Control Society
Assumption 3. There exist a continuous function
(z̃ 1 ,
) : R n−m+ p × R p → R m− p and two continu-
ously differentiable functions W (z̃ 1 ,
) : R n−m+ p ×
R p → R and K (z̃ 1 ,
) : R n−m+ p × R p → R m− p with
2≤i≤ p
K (0,
) = 0 for all {z̃ 1 |col(z̃ 1 + z 1e (
), z̃ 2 + z 2e (
)) ∈
(D)} and all
∈
such that the following inequali-
(26) ties hold:
c1
z̃ 1
2 ≤ W (z̃ 1 ,
)≤c2
z̃ 1
2 (28)
 
 *W (z̃ 1 ,
) 
  ≤ c3
z̃ 1
(29)
 *z̃ 1 
 
 *W (z̃ 1 ,
) 
  ≤ c4
z̃ 1
(30)
*

*W (z̃ 1 ,
)
[ f 1 (z) + g1 (z)
]
*z̃ 1
≤ −c5
z̃ 1
2 +
z̃ 2 − K (z̃ 1 ,
)

(z̃ 1 ,
)

(31)
where c1 , c2 , c3 , c4 and c5 are all positive constants.
Remark 3.
(i) Equations (28), (29) and (31) in Assumption 3
mean that the subsystem z̃ 1 of (27) can be expo-
nentially stabilized to zero regarding z̃ 2 as a virtual
control
input. This assumption is widely used with a lin-
ear growth assumption of interconnection terms
as a feasibility condition of the cascade design
[31]. Nevertheless, (31) replaces the linear growth
assumption with a generalized matching condi-
tion [33]. Actually, there are many other feasibil-
ity conditions for cascade design. We point out
that how to design a better controller based on
the cascade form (5) is out of the scope of this
research; on the contrary, here, we just use the
existing methods. The main topics of this paper
are how to transform the faulty system (3) to a
cascade form, how to estimate the stuck positions
and how to use the estimates to design a fault
tolerant controller.
(ii) Equations (28), (29) and (31) are reasonable in the
sense that if subsystem z̃ 1 cannot be controlled by
z̃ 2 , then it will become uncontrollable. If this is
true, no FTC scheme is feasible. The main restric-
tion imposed by this is that z̃ 2 not only exponen-
tially stabilizes z̃ 1 but also satisfies a generalized
matching condition. Despite the conservativeness,
it covers many applications [31], e.g., (28), (29)
and (31) always hold if system (27) is linear and
(27) controllable.
(iii) Equation (30) is a condition we need to guaran-
tee the stability of the closed-loop system when
 M. Wang and D. Zhou: Fault Tolerant Control of Feedback Linearizable Systems with Stuck Actuators
replacing the true stuck positions with their esti-
mates in the controller designed by the cascade
design method. It seems restrictive. Nevertheless,
when introducing an unknown parameter to the
Lyapunov function to reduce conservativeness (see
(28)), the controller design becomes much more
difficult. (30) is meaningful for it clarifies when
the replacement is feasible.
(iv) Due to the difficulty of nonlinearity, it may be too
ambitious to give a general method of finding the
functions W (z̃ 1 ,
) and K (z̃ 1 ,
) satisfying the
assumption. However, in some cases, they can be
found by the existing nonlinear control techniques.
How to find them is beyond the scope of this re-
search.
By combing the cascade design method with Qu’s
method [33], a controller for the faulty system (27) is
designed:
u H (z̃,
)

z̃ 2 − K (z̃ 1 ,
)

(z̃ 1 ,
)
(t)
−1
=g2H (z) −(z̃ 2 − K (z̃ 1 ,
)) − f 2 (z)
*K (z̃ 1 ,
)
− g2F (z)
+ ( f 1 (z) + g1 (z)
)
*z̃ 1
 V̇ (z̃) ≤ −c5
z̃ 1
2 − (z̃ 2 − K (z̃ 1 ,
))T
(z̃ 2 − K (z̃ 1 ,
))
(z̃ 1 ,
)
2
− (32)

z̃ 2 − K (z̃ 1 ,
)

(z̃ 1 ,
)
+ (t)
where >0 is a constant and (t) = exp(−t). The term
(t) is introduced to avoid singularity [33].
For the closed-loop system (27) and (32), a Lya-
punov function candidate can be taken as:
V (z̃) = W (z̃ 1 ,
) + 12 (z̃ 2 − K (z̃ 1 ,
))T
× (z̃ 2 − K (z̃ 1 ,
)).
Define
Mb =: {z̃ + z e (
)|V (z̃)≤b, b>0, b ∈ R}
M =: max{Mb |Mb ⊆ (D)}.
Lemma 3. Suppose, for the normal system (1), the first
p (1≤ p<m) actuators are stuck at the time t f , Assump-
tions 1–3 hold and the normal controller switches to
the controller (32) at the time ts (ts ≥t f ). If  is chosen
sufficiently small and z(ts ) ∈ M, then the correspond-
ing trajectory of the closed-loop system (27) and (32)
is bounded. Moreover, z(t) → z e (
) as t → ∞.
Proof. The proof is similar to the one in Qu [33]. We
include it here for the completeness of the paper. The
q 2008 John Wiley and Sons Asia Pte Ltd and Chinese Automatic Control Society
81
derivative of V (z̃) along the trajectory of (27) and (32)
is given below.
*W (z̃ 1 ,
)
V̇ (z̃) = [ f 1 (z) + g1 (z)
]
*z̃ 1
 
*K (z̃ 1 ,
) ˙
+ (z̃ 2 − K (z̃ 1 ,
))T z̃˙ 2 − z̃ 1
*z̃ 1
(35)
By (27), (31) and (32), we have:
V̇ (z̃) ≤ −c5
z̃ 1
2 +
z̃ 2 − K (z̃ 1 ,
)

(z̃ 1 ,
)

− (z̃ 2 − K (z̃ 1 ,
))T (z̃ 2 − K (z̃ 1 ,
))
(z̃ 2 −K (z̃ 1 ,
))T (z̃ 2 −K (z̃ 1 ,
))
(z̃ 1 ,
)
2
−

z̃ 2 −K (z̃ 1 ,
)

(z̃ 1 ,
)
+(t)
= −c5
z̃ 1
2 − (z̃ 2 − K (z̃ 1 ,
))T (z̃ 2 − K (z̃ 1 ,
))
+ . (36)

z̃ 2 − K (z̃ 1 ,
)

(z̃ 1 ,
)
+ (t)
It can be verified that the last term in (36) is less than
(t). Therefore,
× (z̃ 2 − K (z̃ 1 ,
)) + (t). (37)
On the boundary of M,
z̃ 1
and
z̃ 2 − K (z̃ 1 ,
)
can-
not be zero at the same time. By choosing  sufficiently
small, V̇ (z̃)<0 can be obtained on the boundary of M.
This implies that any trajectory starting from within M
will remain in it afterwards (for more details, see [29]).
Thus, the corresponding trajectory is bounded. By (37)
(33) and Lemma 4 (in the appendix), the attractiveness prop-
erty holds, i.e., z(t) → z e (
) as t → ∞.

Remark 4. The stability result in Lemma 3 is a local
one. Actually, M is an estimate of the region of attrac-
(34)
tion. That is, at the switching time ts , if the state z(ts )
is beyond the set M, then the boundedness and attrac-
tiveness of the closed-loop system (27) and (32) may
not be guaranteed.
3.4 Fault tolerant controller design with the
estimates of the stuck positions and stability
analysis
The controller (32) cannot be implemented since
true stuck positions
are unknown. To solve this prob-
lem, we replace
with its estimate
ˆ in (32).
82 Asian Journal of Control, Vol. 10, No. 1, pp. 74 87, January 2008

Define z̃ˆ =: z − z e (
) ˆ = col(z̃ˆ 1 , z̃ˆ 2 ) where z̃ˆ 1 = z 1 ˆ
)

where P(z̃, ˆ is given as follows:
− z 1e (
) ˆ and z̃ˆ 2 = z 2 − z 2e (
).
ˆ The cascade system (5)
ˆ
)
P(z̃, ˆ
can also be rewritten as:
˙ *z 1e (
)
ˆ ˙
z̃ˆ 1 = f 1 (z) + g1 (z)
ˆ + g1 (z)
˜ −
ˆ *W (z̃ˆ 1 ,
)
ˆ

*z 1e (
)
ˆ

*
ˆ = g1 (z) − gT F (z)
˙
z̃ˆ 2 = f 2 (z) + g2F (z)
ˆ + g2F (z)
˜ (38) *z̃ˆ 1 *
ˆ

*z 2e (
)
ˆ ˙ *W (z̃ˆ 1 ,
)ˆ
+ g2H (z)u H −
.
ˆ + gT F (z) + (z̃ˆ 2 − K (z̃ˆ 1 ,
))
ˆ T
*
ˆ *
ˆ
The resultant fault tolerant controller is: 
ˆ
) ˆ
). *K (z̃ˆ 1 ,
)ˆ *z 1e (
)
ˆ
u F T C (z̃, ˆ = u H (z̃, ˆ (39) × gT F (z)
*z̃ˆ 1 *
ˆ
A Lyapunov function candidate can be taken as
follows for the closed-loop system (16), (38) and (39). *K (z̃ˆ 1 ,
)
ˆ *K (z̃ˆ 1 ,
)
ˆ
− g1 (z) − gT F (z)
1 *z̃ˆ 1 *
ˆ
ˆ
)
V1 (z̃, ˆ + (z̃ˆ 2 − K (z̃ˆ 1 ,
))
˜ = W (z̃ˆ 1 ,
) ˆ T
2
+ (z̃ˆ 2 − K (z̃ˆ 1 ,
))
ˆ T
× (z̃ˆ 2 − K (z̃ˆ 1 ,
)) ˜ T
.
ˆ + 
˜ (40)
 
*z 2e (
)
ˆ
where >0 is a constant to be chosen. Define: × g2F (z) − gT F (z) . (43)
*
ˆ
M1b =: {(z̃ˆ + z e (
),
ˆ
−
)|V ˆ
)≤b,
˜ 1 (z̃, ˜
By (29) and (30) in Assumption 3, over the
b>0, b ∈ R} (41) bounded set M1 , it follows that

M1 =: max{M1b |M1b ⊆ (D) ×
}. ˆ
)

P(z̃, ˆ


˜
The following theorem gives a stability analysis ≤c6
z̃ˆ 1




˜ + c7
of the closed-loop system consisting of the cascade sys-
tem, the GDO and the fault tolerant controller. ×
z̃ˆ 2 − K (z̃ˆ 1 ,
)


ˆ

,
˜ ˆ
)
∀(z̃, ˜ ∈ M1(44)
Theorem 1. Suppose that, for the normal system (1), for some positive constants c6 and c7 . Using Lemma 5
the first p (1≤ p<m) actuators are stuck at the time t f , (in the appendix) yields:
Assumptions 1–3 hold, the GDO is properly designed,
ˆ
) c6
and the normal controller switches to the fault toler-
P(z̃, ˜ ≤ c6 1
z̃ˆ 1
2 +



ˆ

˜ 2 + c7 2
ant controller (39) at time ts (ts ≥t f ). If  is chosen 1
sufficiently small and (z(ts ),
(t
ˆ s )) ∈ M1 , then the cor- c7
responding trajectory of the closed-loop system (16), ×
z̃ˆ 2 − K (z̃ˆ 1 ,
)

ˆ 2 +



˜ 2
2
(38) and (39) is bounded. Furthermore, z → z e (
) and (45)

(t)
ˆ →
as t → ∞.
for positive constants 1 and 2 . Choosing 1 and 2
Proof. Since Assumption 3 is supposed to hold for all
sufficiently small while  sufficiently large such that
{z̃ 1 |(z̃ 1 + z 1e (
), z̃ 2 + z 2e (
)) ∈ (D)} and all
∈
,
it necessarily holds for (z,
) ˆ ∈ M1 ⊆ (D) ×
. If the c5 1 c6 c7  p
observer gain  of the GDO is properly designed such c6 1 ≤ , c7 2 ≤ , + ≤
2 2 1 2 2
that (18) holds, by Lemma 1 and Assumption 3, the
ˆ
) we have
derivative of V1 (z̃, ˜ along the trajectory of the closed-
loop system (16), (38) and (39) will satisfy the following ˆ
) c5 1
V̇1 (z̃, ˜ ≤ −
z̃ˆ 1
2 − (z̃ˆ 2 − K (z̃ˆ 1 ,
))
ˆ T
inequality: 2 2
ˆ
)
V̇1 (z̃, ˜ ≤ −c5
z̃ˆ 1
2 − (z̃ˆ 2 − K (z̃ˆ 1 ,
))
ˆ T  p
× (z̃ˆ 2 − K (z̃ˆ 1 ,
))
ˆ − ˜T


˜ + (t)
× (z̃ˆ 2 − K (z̃ˆ 1 ,
)) ˆ
)
ˆ + P(z̃, ˆ

˜ 2
ˆ
)
∀(z̃, ˜ ∈ M1 (46)
˜T

−  p
˜ + (t) (42)

q 2008 John Wiley and Sons Asia Pte Ltd and Chinese Automatic Control Society
 M. Wang and D. Zhou: Fault Tolerant Control of Feedback Linearizable Systems with Stuck Actuators 83

If  sufficiently small, by Lemma 4 (in the ap- u1 u2

pendix), the boundedness and attractiveness properties m m
can be proven.
θ1 θ2

k
l

IV. SIMULATION STUDY

a

An inverted pendulum system [34] as shown in

p1 b p2
Fig. 2 is employed to demonstrate the effectiveness of
the proposed FTC method. It consists of two identical
Fig. 2. Inverted pendulums.
inverted pendulums with length l coupled by a spring
and subject to two distinct inputs u 1 and u 2 . The spring
constant is k and the position of the spring is a. The y1

Output 1
0.2
normal length of the spring equals the distance of the
0.1
two pivots p1 and p2 .
Choosing the states x1 =
1 , x2 =
˙ 1 , x3 =
2 , 0

x4 =
˙ 2 , the dynamics of the system can be modeled as (a)
0 5 10
Time(s)
15 20

follows in a neighborhood of zero with some approxi-

mation: Output 2
0.1
y2

ẋ1 = x2 0.05

0
g ka 2 0 5 10 15 20
ẋ2 = sin x1 + 2 (− sin x1 + sin x3 ) (b) Time(s)
l ml
cos x1 0.1 u1
+ u1
Inputs

u2
ml
(47) 0
ẋ3 = x4
-0.1
0 5 10 15 20
g ka 2
ẋ4 = sin x3 + 2 (sin x1 − sin x3 ) (c) Time(s)
l ml
cos x3 Fig. 3. Outputs and inputs of the inverted pendulum model under the
+ u2 normal controller: no fault case.
ml
where g is the gravity constant. As a numerical example,
set Fig. 3 displays the simulation results of the closed-loop
system (48) and (49).
g a 1 k 1
= 0.5, =√ , = 2, = 4. Suppose that the first actuator is stuck at t = t f ,
l l 2 m ml then
Choosing y1 = x1 and y2 = x3 the controlled outputs, u 1 (t) =
, t≥t f (50)
the numerical model is
where
is an unknown constant.
ẋ1 = x2
After the fault (50) occurs, only one output can
ẋ2 = −0.5 sin x1 + sin x3 + 4u 1 cos x1 be arbitrarily controlled. Thus, the normal control
ẋ3 = x4 (48) objective must degrade. The FTC objective here is to
control y2 to zero while maintaining the boundedness
ẋ4 = −0.5 sin x3 + sin x1 + 4u 2 cos x3 of the faulty system (48) and (50). This can be achieved
y1 = x1 , y2 = x3 . by steering its state to xe (
) = (tan−1 (8
), 0, 0, 0)T
which is the equilibrium of the faulty system. Define
The normal control objective is to control the out- x̃1 =: x1 − tan−1 (8
).
puts y1 and y2 to zero, which can be easily achieved by Notice that the faulty system is already a cascade
the following controller: system. Partition the faulty system (48) and (50) into
two subsystems (x1 , x2 , x3 ) and x4 . A Lyapunov func-
u 1 = (4 cos x1 )−1 (− sin x2 − sin x3 ) tion candidate can be taken as below for the subsys-
(49)
u 2 = (4 cos x3 )−1 (− sin x1 − sin x4 ) tem (x1 , x2 , x3 ) over a neighborhood of zero. It can be

q 2008 John Wiley and Sons Asia Pte Ltd and Chinese Automatic Control Society
84 Asian Journal of Control, Vol. 10, No. 1, pp. 74 87, January 2008

verified that the function is positive definite: 0.5

y1

Output 1
V (x̃1 , x2 , x3 ,
) 0

=(6
+ 1)(1 − cos x̃1 ) + 3x22 -0.6
0 1 5 10 15 20
(a) Time(s)
+ x2 sin x̃1 + 0.5(x2 + sin x3 )2 (51) 0.2
y2
√

Output 2
0.1
where
= (1/ 1 + 64
2 )(0.5 + 32
2 )>0.
Regarding x4 as a virtual control input, by the 0

well known backstepping technique, a virtual con- 0 1 5 10 15 20

troller K (x̃1 , x2 , x3 ,
) = −(1/cos x3 )[7x2 + 2 sin x3 + (b)

0.4
Time(s)

(1 −
) sin x̃1 ] can be designed for the subsystem u1

Inputs
(x1 , x2 , x3 ). Moreover, 0.2 u2

0
V̇ (x̃1 , x2 , x3 )
0 1 5 10 15 20
(c)
≤ −
sin2 x̃1 − 5x22 − (x2 + sin x3 )2 Time(s)

+ (x4 − K )(x2 + sin x3 ) cos x3 . (52) Fig. 4. Outputs and inputs of the inverted pendulum model under the
normal controller: the first actuator is stuck at t f = 1s.

It can be verified that the subsystem satisfies Assump-

tion 3. As a result, a controller can be designed follow- 0.4 y1
ing the steps in Section III,
Outputs

0.2 y2

1 *K 0
u2 = −0.5 sin x3 − sin x1 + x2 -0.2
4 cos x3 *x̃1 0 1 2 5 10 15 20
(a) Time(s)
*K *K 0.5
+ (−
sin x̃1 + sin x3 ) + x4 − 3x4 +3K x2
*x2 *x3
States

0.2 x4
 0
(x4 − K )
(x2 + sin x3 ) cos x3
2
− .
-0.2

x4 − K

(x2 + sin x3 ) cos x3
+ 0.2e−t 0 1 2 5 10 15 20

(b) Time(s)
(53)
1 u1
Inputs

The GDO is designed as 0.5

u2

˙ = −4 cos x1 0

0 1 2 5 10 15 20
− (4x2 cos x1 + sin x3 − 0.5 sin x1 ) (54) (c) Time(s)

ˆ =  + x2 .

Fig. 5. States and inputs of the inverted pendulum model: the first
actuator is stuck at t f = 1 s and the normal controller switches
Replacing
with its estimate
ˆ in (52), a fault to the fault tolerant controller at ts = 2 s.
tolerant controller is obtained:

u FTC
=
1
−0.5 sin x3 − sin x1 +
* K̂
x2 where x̃ˆ1 = x1 − tan−1 (8
),
ˆ K̂ = K (x̃ˆ 1 , x2 , x3 ,
)
ˆ and
4 cos x3 *x̃ˆ 1 ˆ
= (1/ 1 + 64
2
ˆ )(0.5 + 32
2
ˆ ).
For simulation, set the initial state as x(0) =
* K̂ [0.2, 0.2, 0.1, 0.1]T . Suppose that the first actuator is
+ (−
ˆ sin x̃ˆ 1 + sin x3 )
*x2 stuck at t = 1 s (
= −0.033). If the controller is not
redesigned after the fault happens, the system will be
* K̂
+ x4 − 3x4 + 3 K̂ unstable due to the non-minimum phase characteristic
*x3 of the model. Fig. 4 shows the inputs and outputs of

(x4 − K̂ )
(x2 + sin x3 ) cos x3
2 the faulty system under the normal controller (49). The
− first output y1 oscillates with large amplitude although

x4 − K̂

(x2 + sin x3 ) cos x3
+0.2e−t the second output y2 converges to zero. It implies that
(55) FTC is necessary.

q 2008 John Wiley and Sons Asia Pte Ltd and Chinese Automatic Control Society
 M. Wang and D. Zhou: Fault Tolerant Control of Feedback Linearizable Systems with Stuck Actuators 85

Suppose at ts = 2 s, FDI is accomplished. Then,
the GDO (54) is activated with (ts ) = 0 and the nor-
mal controller switches to the fault tolerant one (55).
Fig. 5 exhibits the results. The first output y1 is con-
trolled to a new work point and the second output y2 to
zero. All the states are bounded. So the effectiveness of
the proposed FTC method is demonstrated.
A3. (Frobennius Theorem [29]) Consider a nonsingu-
lar d-dimensional distribution (x) = span{ f 1 (x), . . . ,
f d (x)}, defined on an open set U of R n . The vector fields
f 1 (x), . . . , f d (x) are independent. There exist (n − d)
independent solutions defined on U ◦ ⊆ U to the follow-
ing differential equation if and only if the distribution
(x) is involutive.

*T (x)
( f 1 (x) · · · f d (x)) = 0
*x
V. CONCLUSIONS
By ‘independent’, we mean that the row vectors
An active FTC method for FLS with stuck actu-
*T1 (x) *Tn−d (x)
ators is proposed in this paper. The considered fault is ,...,
severe in the sense that the faulty system is no longer *x *x
feedback linearizable, which makes the controller are independent at each x.
design difficult. Under an involutive assumption, a non-
linear transformation is first constructed, which trans- A4. (Schur [35]) For a symmetric ma-
 complements

forms the faulty system into a cascade system. Under S S12 r ×r and S =
trix S = S11
T S where S11 = S11 ∈ R
T
22
an assumption of the feasibility of the famous cascade 1222
T ∈ R (n−r )×(n−r ) , the following inequalities are equiv-
S22
design method, a controller is designed for the faulty
system. On the other hand, a GDO is presented to alent.
estimate the unknown stuck positions. Integrating the
(i) S>0
GDO with the controller constructed by the cascade −1 T
(ii) S22 >0, S11 − S12 S22 S12 >0
design method, a fault tolerant controller is obtained.
The resultant fault tolerant controller guarantees not Consider a nonlinear system:
only boundedness of the state of the faulty system
but also satisfactory output performance. The design ẋ = f (x, t), x(t0 ) = x0
details of the nonlinear transformation and the GDO are
where f (x, t) : D × R + → R n is piecewise continuous
given. An illustrative example verifies the effectiveness
with respect to x and t. Moreover, f (0, t) = 0.
of the method. Due to the difficulty of nonlinearity,
the feasibility of the cascade design method which is Lemma 4. For the above system, suppose that, for any
supposed in Assumption 3 is crucial to the proposed bounded set D ⊂ R n , the range of function f (x, t)
method. Perhaps with further developments in non- is a bounded set in R n . If there exists a continuously
linear control, the conservativeness introduced by the differentiable positive definite function V2 (x) : D → R
assumption can be reduced. satisfying:

*V2
V̇2 (x) = f (x, t)
APPENDIX *x

≤ −W2 (x) + (t) ∀(x, t) ∈ D × R +

A1. A distribution (x) is involutive if the Lie bracket
[1 , 2 ] of any pair vector fields 1 and 2 belonging
to (x) is a vector field which belongs to (x), i.e., where W2 (x) is a positive definite continuous function
1 ∈ (x), 2 ∈ (x) ⇒ [1 , 2 ] ∈ (x) [29]. over D and (t) is a nonnegative ∞ continuous function
Lie bracket is defined as for all t ∈ [t0 , ∞) satisfying t0 (s) ds< which is
sufficiently small, and define M2b =: {x|V2 (x)≤b, b>0,
*2 *1 b ∈ R} and M2 =: max{M2b |M2b ⊆ D}, then for any ini-
[1 , 2 ] = 1 − 2 . tial value x0 ∈ M2 , the corresponding trajectory x(t) is
*x *x
bounded. Furthermore, x(t) is attractive, i.e., x(t) → 0
A2. A distribution (x), defined on an open set U , as t → ∞.
is nonsingular if there exists an integer d such that
dim((x)) = d for all x in U [29]. Note the difference Proof. The lemma can be directly deduced from corol-
from nonsingular matrices. lary 2.18 in [33]. Thus, the details are omitted.

q 2008 John Wiley and Sons Asia Pte Ltd and Chinese Automatic Control Society
86 Asian Journal of Control, Vol. 10, No. 1, pp. 74 87, January 2008

Lemma 5. (Cheng and Zhao [2]) For any two compat- with independent effects on system outputs,” Int. J.
ible real vectors or matrices X and Y , and for any pos- Contr., Vol. 75, No. 11, pp. 792–802 (2002).
itive constant , the inequality X T Y + Y T X ≤X T X + 13. Jiang, B., M. Staroswiecki, and V. Cocquempot,
(1/)Y T Y holds.
REFERENCES
1. Jiang, J., “Fault-tolerant control systems—an
introductory overview,” Acta Automatica Sinica,
Vol. 31, No. 1, pp. 161–174 (2005).
2. Cheng, C. and Q. Zhao, “Reliable control of
uncertain delayed systems with integral quadratic
constraints,” IEE Proc. Contr. Theory Applicat., Vol.
151, No. 6, pp. 790–796 (2004).
3. Fujita, M. and E. Shimemura, “Integrity against
arbitrary feedback-loop failure in linear multi-
variable control systems,” Automatica, Vol. 24, No.
6, pp. 765–772 (1988).
4. Liang, Y. W., S. D. Xu, T. C. Chu, and C. C.
Cheng, “Reliable output tracking control for a class
of nonlinear systems,” IEICE Trans. Fundamentals
Electron. Commun. Comput. Sci., Vol. E87-A,
No. 9, pp. 2314–2321 (2004).
5. Yang, G. H., J. L. Wang, and C. B. Soh, “Reliable
nonlinear control system design by using strictly
redundant control elements,” Int. J. Contr., Vol. 69,
No. 2, pp. 315–328 (1998).
6. Yang, G. H. and K. Y. Lum, “Fault-tolerant flight
tracking control with stuck faults,” Proc. Amer.
Contr. Conf., Denver, CO, pp. 521–526 (2003).
7. Zhou, K. M. and Z. Ren, “A new controller
architecture for high performance, robust, and fault
tolerant control,” IEEE Trans. Automat. Contr., Vol.
46, No. 10, pp. 1613–1618 (2001).
8. Boskovic, J. D., S. M. Li, and R. K. Mehra,
“Intelligent control of spacecraft in the
presence of actuator failures,” Proc. 38th Conf.
Decis. Contr., Phoenix, AZ, pp. 4472–4477 (1999).
9. Chen, W. and J. Jiang, “Fault-Tolerant Control
against Stuck Actuator Faults,” IEE Proc. Contr.
Theory Applicat., Vol. 152, No. 2, pp. 138–146
(2005).
10. Dardinier-Maron, V., F. Hamelin, and H. Noura,
“A fault-tolerant control design against actuator
failures: Application to a three-tank system,”
Proc. 38th Conf. Decis. Contr., Phoenix, AZ, pp.
3569–3574 (1999).
11. Diao, Y. X. and K. M. Passino, “Intelligent
fault-tolerant control using adaptive and learning
methods,” Contr. Eng. Practice, Vol. 10, No. 8, pp.
801–817 (2002).
12. Jiang, B., J. L. Wang, and Y. C. Soh, “An
adaptive technique for robust diagnosis of faults
q 2008 John Wiley and Sons Asia Pte Ltd and Chinese Automatic Control Society
“Active fault tolerant control for a class of nonlinear
systems,” Proc. IFAC Safeprocess, Washington, pp.
127–132 (2003).
14. Jiang, B. and F. N. Chowdhury, “Fault estimation
and accommodation for linear MIMO discrete-time
systems,” IEEE Trans. Contr. Syst. Technol., Vol.
13, No. 3, pp. 493–499 (2005).
15. Jiang, J. and Q. Zhao, “Reconfigurable control
based on imprecise fault identification,” Proc. Amer.
Contr. Conf., San Diego, pp. 114–118 (1999).
16. Kabore, P. and H. Wang, “Design of fault diagnosis
filters and fault tolerant control for a class of
nonlinear systems,” IEEE Trans. Automat. Contr.,
Vol. 46, No. 11, pp. 1805–1810 (2001).
17. Maki, M., J. Jiang, and K. Hagino, “A stability
guaranteed active fault-tolerant control system
against actuator failures,” Int. J. Robust Nonlin.
Contr., Vol. 14, No. 12, pp. 1061–1077 (2004).
18. Mhaskar, P., A. Gani, and P. D. Christo-
fides, “Fault-tolerant control of nonlinear
processes: Performance-based reconfiguration and
robustness,” Int. J. Robust Nonlin. Contr., Vol. 16,
No. 3, pp. 91–111 (2006).
19. Polycarpou, M. M., “Fault accommodation of a
class of multivariable nonlinear dynamical systems
using a learning approach,” IEEE Trans. Automat.
Contr., Vol. 46, No. 5, pp. 736–742 (2001).
20. Qu, Z. H., C. M. Ihlefeld, Y. F. Jin, and
A. Saengdeejing, “Robust fault-tolerant self-
recovering control of nonlinear uncertain systems,”
Automatica, Vol. 39, No. 10, pp. 1763–1771 (2003).
21. Tao, G., S. H. Shen, and S. M. Joshi, “An adaptive
actuator failure compensation controller using
output feedback,” IEEE Trans. Automat. Contr., Vol.
47, No. 3, pp. 506–511 (2002).
22. Tao, G., S. B. Chen, X. D. Tang, and S. M. Joshi,
Adaptive Control of Systems with Actuator Failures,
Springer, London (2004).
23. Thomas, S., B. C. Chang, and H. G. Kwatny,
“Controller reconfiguration for nonlinear systems
using composite observer,” Proc. Amer. Contr.
Conf., Denver, CO, pp. 4779–4784 (2003).
24. Wu, N. E. Y. Zhang, and K. Zhou, “Detection,
estimation and accommodation of loss of control
effectiveness,” Int. J. Adaptive Contr. Signal
Processing, Vol. 14, No. 7, pp. 775–795 (2000).
25. Yen, G. G. and L. W. Ho, “Online multiple-model-
based fault diagnosis and accommodation,” IEEE
Trans. Ind. Electron., Vol. 50, No. 2, pp. 296–312
(2003).
 M. Wang and D. Zhou: Fault Tolerant Control of Feedback Linearizable Systems with Stuck Actuators
26. Zhang, X. D., T. Parisini, and M. M. Polycarpou,
“Adaptive fault-tolerant control of nonlinear
uncertain systems: An information-based diagnostic
approach,” IEEE Trans. Automat. Contr., Vol. 49,
No. 8, pp. 1259–1274 (2004).
27. Zhang, Y. M. and J. Jiang, “Fault tolerant
control system design with explicit consideration of
performance degradation,” IEEE Trans. Aerospace
Electron. Syst., Vol. 39, No. 3, pp. 838–848 (2003).
28. Zhou, D. H. and P. M. Frank, “Fault diagnostics
and fault tolerant control,” IEEE Trans. Aerospace
Electron. Syst., Vol. 34, No. 2, pp. 420–427 (1998).
29. Isidori, A., Nonlinear Control Systems, 3rd Ed,
Springer, New York (1995).
30. Esfandiari, F. and H. K. Khalil, “Output feedback
stabilization of fully linearizable systems,”
Int. J. Contr., Vol. 56, No. 5, pp. 1007–1037 (1992).
31. Sepulchre, R., M. Jankovic, and P. V. Kokotovic,
Constructive Nonlinear Control, Springer, London
(1997).
32. Chen, W. H., “Disturbance observer based
control for nonlinear systems,” IEEE-ASME Trans.
Mechatron., Vol. 9, No. 4, pp. 706–710 (2004).
33. Qu, Z. H., Robust Control of Nonlinear Uncertain
Systems, Wiley Interscience, New York (1999).
34. Gavel, T. D. and D. D. Siljak, “Decentralized
adaptive control: Structural conditions for stability,”
IEEE Trans. Automat. Contr., Vol. 34, No. 4, pp.
413–426 (1989).
35. Boyd, S., L. E. Ghaoui, E. Feron, and V.
Balakrishnan, Linear Matrix Inequalities in System
and Control Theory, SIAM Studies in Applied
Mathematics, Philadelphia (1994).
q 2008 John Wiley and Sons Asia Pte Ltd and Chinese Automatic Control Society
87
Min Wang received her B.Eng.
degree in automation from Tianjin
University in 2002. Since 2002,
she has been studying in the De-
partment of Automation, Tsinghua
University, pursuing her Ph.D. de-
gree. Her current research interest
is fault tolerant control.
Dr. Donghua Zhou received his
B.Eng., M.Sci., and Ph.D. de-
grees all from the Department of
Electrical Engineering, Shanghai
Jiaotong University, in 1985,
1988 and 1990, respectively. He
was an Alexander von Humboldt
Research Fellow (1995–1996)
in the University of Duisburg,
Germany, and a visiting professo
in Yale University (July 2001–
Jan 2002). Dr. Zhou is currently
a professor in the Department
of Automation, and the Director of the Institute of Pro-
cess Control Engineering, Tsinghua University.
Dr. Zhou has published over 140 journal papers and
three monographs in the areas of process identification,
diagnosis and control. He serves the profession in many
capacities such as IFAC Technical Committee on Fault
Diagnosis and Safety, Deputy General Secretary of
Chinese Association of Automation (CAA), and Coun-
cil of CAA. He was also the NOC Chair of the 6th
IFAC Symposium on SAFEPROCESS, 2006.