Professional Documents
Culture Documents
Adaptive Neural-Networks-Based Fault Detection and Diagnosis Using Unmeasured States
Adaptive Neural-Networks-Based Fault Detection and Diagnosis Using Unmeasured States
Adaptive Neural-Networks-Based Fault Detection and Diagnosis Using Unmeasured States
org
ISSN 1751-8644
Abstract: Fault detection and diagnosis play important roles in modern engineering systems. A number of fault
diagnosis (FD) approaches for nonlinear systems have been proposed. But, most of the achievements are based on
the assumptions that the systems models are known, and states of systems are measurable. A novel FD
architecture for a class of unknown nonlinear systems with unmeasured states has been investigated. A general
radial basis function (RBF) neural network is used to approximate the model of unknown system, an adaptive RBF
neural network with on-line updated centre, and the width vector of ‘Gaussian’ function is used to approximate
the model of fault. A nonlinear state observer is designed to estimate system states that are inputs to the neural
networks. The stability analysis for the system is given, and the adaptive parameter-updating laws are derived using
Lyapunov theory. Simulation examples are used to illustrate the effectiveness of the proposed method.
1066 IET Control Theory Appl., 2008, Vol. 2, No. 12, pp. 1066 – 1076
& The Institution of Engineering and Technology 2008 doi: 10.1049/iet-cta:20070216
www.ietdl.org
nonlinear systems [12, 13]. To the authors’ knowledge, few 2 Representation of systems
neural networks FD schemes for unknown nonlinear
systems with unmeasured states have been done. Consider a class of multivariable nonlinear dynamic systems
described by the differential equation
Motivated by these considerations, this paper proposes a
novel FD approach for unknown nonlinear system with x_ ¼ f1 (x, u) þ b(t T )j(x)
unmeasured states. A general radial basis function (RBF) (1)
y ¼ Cx
neural network is used to approximate the model of
unknown system, an adaptive RBF neural network with
where x [ R n is the unmeasured state vector, u [ R m is
on-line updated centre and the width vector of ‘Gaussian’
the input vector and y [ R p is the output vector.
function is used to approximate the model of fault. A state
f1 (x, u) : Rn Rm ! Rn and j(x) : Rn ! Rn are unknown
observer is designed in order to estimate systems states that
smooth vector fields, representing the dynamics of system
are input to the neural networks. Parameters updating laws
and the nonlinear fault function, C is known. b(t T ) is a
are derived based on Lyapunov theory [14]. With the
matrix function representing the time profiles of fault,
proposed diagnosis scheme, the uniform ultimate
where T denotes the unknown fault occurrence time. The
boundness of all signals can be guaranteed, and the error
fault time profile b(.) is assumed to be a diagonal matrix of
between system fault and estimated fault model can be
the form [9]
made arbitrarily small by suitably choosing design
parameters.
b(t T ) ¼ diag(b1 (t T ), b2 (t T ), . . . , bn (t T ))
This paper is organised as follows. In Section 2, a general
representation of nonlinear systems is presented. In where bi : R ! R is a function representing the profile of a
Section 3, the design scheme of observer and analysis of fault affecting the ith state equation, i ¼ 1, . . . , n. More
algorithms are given. In Section 4, the fault detection is specifically, in this paper, the abrupt fault with time profiles
analysed. In Section 5, a design methodology for FD and is considered, because the diagnosis for slowly varying
the stability analysis are developed. In Section 6, simulation faults or fault prediction is another complex problem.
examples are given to illustrate the major procedure of
the method. Finally, some conclusions are illustrated in 1 tT
Choosing bi (t T ) ¼
Section 7. 0 t,T
ð 1 1=2
3 Observer design
kd k2 ; kd (t)k2 d t In control engineering, RBF network is usually used as a tool
0 for modelling nonlinear system because of its good capabilities
in function approximation. If the states can be available for
and measurement, then the following RBF network similar to
[15] is used to approximate the continuous unknown
dynamics f (x, u)
kd k1 ; sup kd (t)k
t0
f (x, u) ¼ B0 W0 S0 (X ) þ 11 (X ) (3)
We will say that d (t) [ L2 when kdk2 is finite. Similarly, we
will say that d (t) [ L1 when kdk1 is finite. arg is abbreviated and the output of neural network is f^ (x, u) ¼ B0 W
^ 0 S0 (X ),
from argument, K ¼ arg min J (K ) represents K which can where X ¼ [xT uT ]T [ RN (N ¼ n þ m), X [ Ad , RN ,
minimise index J(K ). Ad is a compact set, W0 [ Rnk is the ideal weight
IET Control Theory Appl., 2008, Vol. 2, No. 12, pp. 1066 – 1076 1067
doi: 10.1049/iet-cta:20070216 & The Institution of Engineering and Technology 2008
www.ietdl.org
matrix, that is satisfying the following condition Theorem 1: Consider the nonlinear system (4), the state
observer (5) and the error equation (6), assume that there
W0 ¼ arg min {supX [Ad j f (x, u) B0 W0 S0 (X )j} exists a set Ad large enough to satisfy Assumptions 1 and 2,
W0
if the following parameter updating law is adopted, then e
where S0 (x, u) [ Rk is the Gaussian function vector, Gaussian is uniformly ultimately bounded, that is e [ L1
function is chosen as si (x) ¼ exp(kx ci k2 =s2i ) with centre
vector ci and width si , k is the number of the neurons. _^ T T
W 0 ¼ g0 L0 ey S0 (^
x, u) (7)
B0 [ Rnn is a matrix to be designed, 11 (X ) is a neural
network approximation error.
where L0 [ Rpn , B0 [ Rnn satisfying PB0 ¼ C T L0 , P ¼
Assumption 1: There exists an ideal weight matrix W0 , so P T . 0, (A T P þ PA) ¼ Q . 0, ey ¼ y^ y ¼ Ce, g0 . 0
for all X [ Ad , 10 and W
that j11 (X )j 10 and kW0 k W is the parameter to be designed.
are positive constants.
Proof: The Lyapunov function can be chosen as
However, the states are immeasurable in this paper; the
real input of neural network should be estimated state
1 1 ~ T0 W
vector x^ , and its output is f^ (^x, u). V ¼ eT Pe þ tr(W ~ 0) (8)
2 2g0
Where f^ (^x, u) ¼ B0 W
^ 0 S0 (^x, u), W
^ 0 is an estimated matrix of
_~ _^ T T
W0 . By differentiating (8) and applying W 0 ¼ W 0 ¼ g0 L0 ey S0
(^x, u), we obtain
If there is no fault, from (3), (2) can be rewritten as
1
x_ ¼ A1 x þ B0 W0 S0 (x, u) þ 11 (X ) (4) V_ ¼ eT Qe þ eT P1 þ eyT L0 W ~ 0 S0 (^x, u)
2
~ 0)
tr(S0 (^x, u)eyT L0 W (9)
To estimate the system states, a Luenberger-like observer
based on linear system counterpart [16] is constructed
Since
x_^ ¼ A1 x^ þ B0 W
^ 0 S0 (^x, u) þ G(y y^ ) (5)
~ 0 S0 (^x, u) ¼ tr(eyT L0 W
eyT L0 W ~ 0 S0 (^x, u))
where x^ is an estimate of x, y^ ¼ C x^ .
~ 0)
¼ tr(S0 (^x, u)eyT L0 W (10)
Defining e ¼ x^ x. From (4) and (5), the following state
error equation is obtained then
^ 0 S0 (^x, u) B0 W0 S0 (x, u) 11 (X )
e_ ¼ Ae þ B0 W 1 1
(6) V_ ¼ eT Qe þ eT P1 l1 eT e þ eT P1
~ 0 S0 (^x, u) þ
¼ Ae þ B0 W B0 W0 S~ 0 (x, x^ , u) 11 (X ) 2 2
1 T 1 T 4 T
¼ l1 e e l1 e e e P1 (11)
or 4 4 l1
~ 0 S0 (^x, u) þ 1
e_ ¼ Ae þ B0 W
where l1 is the minimum eigenvalue of Q.
where W ~0¼W ^ 0 W0 , 1 ¼ B0 W0 S~ 0 (x, x^ , u) 11 (X ),
~S 0 ¼ S0 (^x, u) S0 (x, u), A ¼ A1 GC. Using the fact that eT e (4=l1 )eT P1 (4= l21 )1T P T P1,
the following inequality holds
According to the linear system theory, A can be designed
as a Hurwitz matrix, which is guaranteed by the condition 1 1 1
that A1 , C is observable. V_ l1 eT e þ 1T PP1 l1 kek2
4 l1 4
1068 IET Control Theory Appl., 2008, Vol. 2, No. 12, pp. 1066 – 1076
& The Institution of Engineering and Technology 2008 doi: 10.1049/iet-cta:20070216
www.ietdl.org
BW S(^x, d , s ) þ BW S~ d (20)
The outputs f^ (^x, u), ĵ (^x) of two RBF networks are
used to approximate f (x, u), j(x), respectively, where
f^ (^x, u) ¼ B0 W
^ 0 S0 (^x, u), ĵ (^x) ¼ BWS(^
^ x). where S~ ¼ S(^x, d , s ) S(x, d , s ).
The approximating property for the nonlinear model The Taylor’s series of S(^x, d , s ) is expanded at (d^ , s^ ),
depends on the centre vector and width vector of ‘Gaussian’ that is
function. It is difficult to choose centre vectors and width
vectors of two ‘Gaussian’ functions by trial and error.
S(^x, d , s ) ¼ S(^x, d^ , s^ ) Sd0^ d~ Ss0^ s~ þ O()
Hence, an adaptive RBF network is used to approximate
the system fault, that is the parameter vectors d and s are
on-line updated. where d~ ¼ d^ d , s~ ¼ s^ s , Sd0^ ¼ @S=@djd ¼d^ [ Rqk0 ,
IET Control Theory Appl., 2008, Vol. 2, No. 12, pp. 1066 – 1076 1069
doi: 10.1049/iet-cta:20070216 & The Institution of Engineering and Technology 2008
www.ietdl.org
Sŝ0 ¼ @S=@sjs¼s^ [ Rqq . Then, we obtain system states should remain bounded after the occurrence
of a fault.
BW S(^x, d , s ) ¼ BW S(^x, d^ , s^ ) BW Sd0^ d~
3. The knowledge of fault-free system is not used at all in the
^ x, d^ , s^ )
BW Ss0^ s~ þ BW O() ¼ BWS(^ parameter-updating laws, so the FD approach proposed in
~ x, d, ^ 0^ d~ þ BWS
^ s^ ) BWS ~ 0^ d~ the paper can be applied to the system with unknown model.
BWS(^ d d
^ s0^ s~ þ BWS
BWS ~ s0^ s~ þ BW O() (21) Lemma 2: Parameter-updating laws (23) – (25) guarantee
that the following inequalities hold
Substituting (21) into (20), the error equation can be written
as tr W ~ T 1W _^
þF 0 (26)
g
~ 0 S0 (^x, u) þ BWS(^
e_ ¼ Ae þ B0 W ~ x, d^ , s^ )
~d T 1 d_^ þ Fd 0 (27)
^ 0^ d~ þ BWS
þ BWS ^ s0^ s~ þ n (22) gd
d
T 1 _
where W ~ ¼ W ^ W and n ¼ BW S~ þ B0 W0 S~ 0 d þ s~ s^ þ Fs 0 (28)
gs
~ S ^ d~ þ BW
(BW 0 ~ Ss0^ s~ þ BW O()).
d
Proof: See the Appendix.
Lemma 1: If the following parameter-updating laws (23) – Theorem 2: Consider the nonlinear fault system (17), state
(25) are adopted, then n is bounded in the set Ad observer (18) and error equation (22), suppose that there
8 exists a set Ad large enough to satisfy Assumptions 1 – 3.
>
< ^ k2 b)tr(FT W
(kW ^) ^ k2 . b and If parameter-updating laws (7) and (23) – (25) are
_^ g F þ g ^ , kW
W
W ¼ kW^k 2 tr(FT W^ ),0 adopted, and B0 , B, L0 , L satisfy the conditions
>
:
gF, other PB0 ¼ C T L0 , PB ¼ C T L, then e is uniformly ultimately
bounded, that is e [ L1 , where P is the same one as in
(23) Theorem 1.
8
>
< (kd^ k2 bd )FTd d^ ^ kd^ k2 . bd and
_
d^ ¼
gd Fd þ gd d, Proof: The Lyapunov function can be chosen as
> kd^ k2 FTd d^ , 0
:
g d F d , other 1 1
V ¼ eT Pe þ tr(W ~ 0 ) þ 1 tr(W
~ T0 W ~ TW
~)
(24) 2 2g0 2g
8 1 ~T ~ 1 T
< gs Fsi (1 þ s^ i ai ), s^ i , ai and Fsi , 0 þ d dþ s̃ s̃ (29)
s_^ i ¼ gs Fsi (1 s^ i þ bi ), s^ i . bi and Fsi . 0 2gd 2gs
: gs Fsi , other
Matrix Q is the same one as in Theorem 1.
(25)
_^ _~
~_ ¼ W
Applying PB0 ¼ C T L0 , PB ¼ C T L and W , W0 ¼
where F ¼ LT ey S T (^x, d^ , s^ ) [ Rnq , Fd ¼ Sd0^ T W
^ T LT ey [ _^
W 0 , we have
k0 T
R , Fs ¼ (ey LW ^ Ss^ ) [ R , L [ R , and b, bd , ai , bi ,
0 T q pn
1070 IET Control Theory Appl., 2008, Vol. 2, No. 12, pp. 1066 – 1076
& The Institution of Engineering and Technology 2008 doi: 10.1049/iet-cta:20070216
www.ietdl.org
IET Control Theory Appl., 2008, Vol. 2, No. 12, pp. 1066 – 1076 1071
doi: 10.1049/iet-cta:20070216 & The Institution of Engineering and Technology 2008
www.ietdl.org
network is used as a fault approximator. When the fault model conclude from Figs. 5 and 6 that even though the fault is
is the function of time, that is j(t) ¼ 0.1, (t 6 s), the input modelled as an external additive input (function of time), it
of the fault neural networks is also x̂, the learning capability of can also be approximated by neural network. It is because that
the FD architecture is also illustrated in Fig. 5. This figure the input of the neural network is the implicit function of time.
shows that the neural network output provides a good
approximation of the fault function. Fig. 6 gives a result with We also compare the FD results in Fig. 2 with the condition
the fault function j(t) ¼ 0.1sin(0.5t), (t 15 s). Hence, we where the system full state variables can be measurable, since
Figure 3 States estimation errors and estimated fault with noise N(0, 0.01)
1072 IET Control Theory Appl., 2008, Vol. 2, No. 12, pp. 1066 – 1076
& The Institution of Engineering and Technology 2008 doi: 10.1049/iet-cta:20070216
www.ietdl.org
Figure 4 States estimation errors and estimated fault with neural networks nodes (k ¼ 7)
there are nearly no difference between these two situations, the the estimation accuracy is mainly related to the number of
later is neglected. neurons of networks and g, g0 , gd , gs .
Remark 3: The simulation studies show that the system Example 2: This is a multiple faults example. The
stability is related to Matrix B, B0 to a great extent, whereas dynamical model of a planar, two-link, articulated robotic
Figure 6 Simulation results with fault function j(t) ¼ 0.1 sin 0.5t
IET Control Theory Appl., 2008, Vol. 2, No. 12, pp. 1066 – 1076 1073
doi: 10.1049/iet-cta:20070216 & The Institution of Engineering and Technology 2008
www.ietdl.org
manipulator [17] is nonlinear with strong coupling between Let the reference inputs of the robotic manipulator are
the two degrees of freedom. The equations of motion in yg ¼ [yg1 yg2 ]T .
terms of the generalised coordinates q1 and q2 , representing
the angular positions of joints 1 and 2 and applied torques The controller is designed as [18]
t1 and t2 at these joints is given by
t1
u¼
q€ 1 h_q2 h(_q 1 þ q_ 2 ) q_ 1 t1 t2
¼ M 1 þ M 1
q€ 2 h_q1 0 q_ 2 t2 x4 h(^x3 þ x^ 4 ) x^ 3
1 h^
¼ M K e^ þ y€ r þ M
þ w(q1 , q_ 1 , q2 , q_ 2 ) þ b(t T )j(q1 , q_ 1 , q2 , q_ 2 ) h^x3 0 x^ 4
where j(q1 , q_ 1 , q2 , q_ 2 ) [ R2 is an unknown fault function We consider a fault that occurs because of a tangle of
vector. w(q1 , q_ 1 , q2 , q_ 2 ) [ R2 is a vector containing the complex factors. This fault is assumed to manifest itself as
unknown static and dynamic terms. a nonlinear change (in the robotic system dynamics)
described by
Let x ¼ [x1 x2 x3 x4 ]T ¼ [q1 q2 q_ 1 q_ 2 ]T , u¼ 2 3
[ t1 t2 ] T 0
6 5x1 sin x3 0:4 sin x3 7
j(x) ¼ 6
4
7
5
Equation 37 can be rewritten as 0
x1 sin (0:5x4 ) þ x4
x_ 3 f11 (x) f12 (x) x3
¼ þ w(x) The numerical values used for simulation purposes are
x_ 4 f21 (x) f22 (x) x4
m1 ¼ 1, l1 ¼ 1, me ¼ 2, de ¼ 308, I1 ¼ 0:12, lc1 ¼ 0:5, Ie ¼
g11 (x) g12 (x) 0:25, lce ¼ 0:6. The reference inputs are yg1 ¼ cos t, yg2 ¼
þ u þ b(t T )j(x) (38)
g21 (x) g22 (x) sin t and the initial position is q(0) ¼ [0:5 0:5]T , the
initial estimate states are
e^(0) ¼0, and
or 13 0 4 0 12 0
K ¼ , A¼ , G¼
0 24 0 10 0 8
x_ 3 4 0 8 0 0:333 0
¼ f (x) þ g(x)u þ b(t T )j(x) (39) , Q¼ , P¼ , B0 ¼
x_ 4 5 12 0 10 0 0:625
2 5 2:5 11 0:667 1:667
, B¼ , L0 ¼ ,
T 1 0 x1 3 3:8 4 6 1:875 2:375
y ¼ [ y1 y2 ] ¼
0 1 x2 0:7667 2
L¼ .
6:25 9:375
From (37), it is can be seen that g(x) [ R22 are known,
f (x), j(x) are unknown which can be approximated by two Initial centre and width of the basis function
neural networks proposed in this paper. Besides, the states are dij [ [1:5, 1:5], s1 [ [0:4, 0:6], i ¼ 1; . . . ; k, j ¼
x1 , x2 are measurable, the states x3 , x4 are immeasurable. 1, 2, W^ (0) ¼ 0, k ¼ q ¼ 15. The other values are b ¼ 1:5,
1074 IET Control Theory Appl., 2008, Vol. 2, No. 12, pp. 1066 – 1076
& The Institution of Engineering and Technology 2008 doi: 10.1049/iet-cta:20070216
www.ietdl.org
Remark 4: The fault models provided by the neural [3] KABORE P., WANG H.: ‘Design of fault diagnosis filters
network can be used for further identifying the failure mode. and fault-tolerant control for a class of nonlinear
systems’, IEEE Trans. Autom. Control, 2001, 46, (11),
Remark 5: From the simulation results in the two examples, pp. 1805 – 1810
it can be seen that the neural network can provide a good
approximation of the fault function. But when fault model [4] PERHINSCHI M.G., MARCELLO R., CAMPA G., ET AL .: ‘An adaptive
is a step function, the output of fault approximator has a threshold approach for the design of an actuator failure
large dynamic error in a short time. So, the dynamic time detection and identification scheme’, IEEE Trans. Control
of fault estimate should also be considered in the practical Syst. Technol., 2006, 14, (3), pp. 556– 566
system, in order to avoid false fault detection.
[5] JIANG B., CHOWDHURY F.: ‘Fault estimation and
accommodation for linear MIMO discrete-time systems’,
7 Conclusion IEEE Trans. Control Syst. Technol., 2005, 13, (3), pp. 493– 499
In this paper, the FD scheme based on neural networks is
presented for a class of unknown nonlinear systems with [6] WANG H., HUANG Z.J., DALEY S.: ‘On the use of adaptive
unmeasured states. Two neural networks are used to updating rules for actuator and sensor fault diagnosis’,
approximate nonlinear dynamics and nonlinear fault Automatica, 1997, 33, (2), pp. 217 – 225
function which can be modelled as a function of system
states and input or a function of time. A nonlinear state [7] HAMMOURI H., KINNAERT M., YAAGOUBI E.H.EI.: ‘Observer-based
observer is designed to estimate the system states. The approach to fault detection and Isolation for nonlinear
main feature of the proposed scheme is that it needs systems’, IEEE Trans. Autom. Control, 1999, 44, (10),
neither a system model being exactly known nor system pp. 1879 – 1884
states being fully measurable, because for a physical system,
it is difficult to measure all its states. Further research work [8] WANG H., DALEY S.: ‘Actuator fault diagnosis: an adaptive
includes extension of the proposed approach to networked observer-based technique’, IEEE Trans. Autom. Control,
control systems with unknown nonlinear plants. 1996, 41, (7), pp. 1073– 1078
IET Control Theory Appl., 2008, Vol. 2, No. 12, pp. 1066 – 1076 1075
doi: 10.1049/iet-cta:20070216 & The Institution of Engineering and Technology 2008
www.ietdl.org
[12] SESHAGIRI S.S., KHALIL H.K.: ‘Output feedback control of Proof of Lemma 2:
nonlinear systems using RBF neural networks’, IEEE Trans.
Neural Netw., 2000, 11, (1), pp. 69– 79 " #
^ 2 T ^
~ T 1 _^ ~ (kW k b)tr(F W ) W
T
^
[13] HOVAKIMYAN N., CALISE A.J., KIM N.: ‘Adaptive output feedback tr W W þF ¼ tr W
g kW^ k2
control of multi-input multi-output systems using neural
networks’, Int. J. Control, 2004, 77, (15), pp. 1318 – 1329 ^ k2 b)tr(FT W
(kW ^)
¼ ~ TW
tr(W ^)
kW^k 2
[14] SLOTINE J.-J.E., LI W.: ‘Applied nonlinear control’ (Prentice
Hall, 1991)
^ k2 . b, kW
When kW ^ k . kW k,
[15] VARGAS J.A.R. , HEMERLY E.M.: ‘Robust neural adaptive
observer for general nonlinear systems’. [A]. Proc. of
that is
American Control Conf. [C]. Chicago, IL, USA, NJ, IEEE
Press, 2000, vol. 1, pp. 708 – 712
^ ) ¼ 1 kW
~ TW
tr(W ^ k2 þ 1 kW
~ k2 1 kW k2 0
2 2 2
[16] CICCARELLA G., DALLA MORA M., GERMAN A.: ‘A Luenberger-like
observer for nonlinear systems’, Int. J. Control, 1993, 57, Therefore the following conclusions are obtained
(3), pp. 537– 556
T
i. when kW^ k2 . b and ^ ) , 0,
tr(FT W ~ (1=g)
tr[W
[17] XU H., IOANNOU P.A.: ‘Robust adaptive control for a class of _^
W þ F)] 0,
MIMO nonlinear systems with guaranteed error bounds’,
IEEE Trans. Autom. Control, 2003, 48, (5), pp. 728 – 742 ii. when the conditions kW ^ k2 . b and tr(FT W
^ ) , 0 are
T
~ (1=g)W_
^ þ F)] ¼ 0.
not satisfied, tr[W
[18] ZHANG S., HU S.: ‘Output feedback tracking control for a
class of MIMO nonlinear minimum phase systems based The proofs of inequalities (27) and (28) are similar
on RBF neural networks’, Int. J. Innov. Comput., Inf. to (26). A
Control, 2008, 4, (4), pp. 803– 812
1076 IET Control Theory Appl., 2008, Vol. 2, No. 12, pp. 1066 – 1076
& The Institution of Engineering and Technology 2008 doi: 10.1049/iet-cta:20070216