Neurocomputing 74 (2011) 37533759

Contents lists available at SciVerse ScienceDirect

Neurocomputing
journal homepage: www.elsevier.com/locate/neucom

Adaptive neural control for uncertain systems subject to actuator failure

Tinggang Jia a,b, Yuanyuan Zou a,n, Yugang Niu a
a
b

Key Laboratory of Advanced Control and Optimization for Chemical Processes (East China University of Science and Technology), Ministry of Education, Shanghai 200237, P.R. China
Central Academe, Shanghai Electric Group Co., Ltd., Shanghai, PR China.

a r t i c l e i n f o

abstract

Article history:
Received 3 May 2011
Received in revised form
16 June 2011
Accepted 20 June 2011
Communicated by Z. Wang
Available online 25 August 2011

In this paper, the design of controller based on neural network is investigated for a class of uncertain
systems subject to actuator failures. An adaptive neural controller is designed by utilizing the
approximation technique of neural network. The key feature in this work is to remove the requirement
on the boundedness of unknown nonlinear functions that is usually encountered in the existing works.
Moreover, sufcient conditions are derived such that the closed-loop system is robustly stable. Finally,
numerical simulation results are given.
& 2011 Elsevier B.V. All rights reserved.

Keywords:
Adaptive control
Neural network
Actuator failure
Unknown nonlinear function

1. Introduction
Due to their powerful approximation capability, neural-network-based control strategy has shown great promises in delivering satisfactory control performance for complex nonlinear
systems. In the past decades, a wealth of the literature focussing
on adaptive neural control schemes has been provided in
[11,18,5,7,15,8] and the references therein. However, it is worthy
of noting that there is little work investigating the neural-network-based control for systems subject to actuator failures.
In practical applications, the failures of control components
often occur and may result in unsatisfactory performance or even
instabilities. The study on the design of reliable control has
recently attracted persistent attention and various control methodology has been presented, see, e.g., [13,17,6,12,16,9]. These
control systems may be roughly classied as active [3,17,1] and
passive [2,6,12,16,9]. In the passive case, a suitable controller is
designed such that the resultant closed-loop system can attain
desired performance not only when all control components are
operational, but also in the case of some admissible actuator
failures. However, in the aforementioned works, the nonlinear
uncertainties are usually assumed to satisfy certain normbounded condition. Apparently, these assumption may result in
the limited application of these control schemes. Hence, how to
reduce the requirement on the boundedness of unknown nonlinear function is interesting and has practically signicant. That
is just the objective of the present work.

Corresponding author.
E-mail address: [email protected] (Y. Zou).

0925-2312/$ - see front matter & 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.neucom.2011.06.030

In this work, the problem of adaptive neural control is

considered for a class of uncertain systems subject to actuator
failures. The unknown nonlinear function in the controlled system
is not required to be bound. By utilizing neural network technique,
the approximation model of unknown nonlinear functions is
rstly established in this work, which removes the requirement
on the boundedness of unknown nonlinear function. The control
matrix is further decomposed according to the model of actuators
failures. Then, an adaptive neural controller is designed such that
the closed-loop system is robustly stable for admissible uncertainties and actuator failures among a prescribed subset of
actuators. Moreover, the available information of delayed states
is utilized, which may further reduce the conservatism, compared
with the memoryless controller.
Throughout, J  J denotes the Euclidean norm of a vector or the
spectral norm of a matrix, J  JF denotes the Frobenius norm of a
matrix, lmin  and lmax  denote the minimum and maximum
eigenvalue of a matrix, respectively. For a real symmetric matrix,
M 4 0 o 0 means that M is positive denite (negative denite).
I denotes an identity matrix of appropriate dimensions. Matrices,
if their dimensions are not stated, will be assumed to have
compatible dimensions.

2. Problem description
Consider the following uncertain system:
_ A DAtxt Ad DAd txtd But f xt,xtd,
xt
1

3754

T. Jia et al. / Neurocomputing 74 (2011) 37533759

zt Cxt,
xt jt,

2
t A d,0,
n

3
m

where xt A R is the state, ut A R is the control input, zt A Rq

is the controlled output, matrices A, Ad, B and C are known, d is the
delay time, jt is a continuous vector-valued initial function.
DAt and DAd t are parameter uncertainties satisfying
DAt DAd t HFtE Ed ,

where H, E and Ed are known constant matrices, F(t) is an

unknown time-varying matrix satisfying F T tFt rI.
In the system (1), the unknown nonlinear function f xt,xtd A
Rm denotes external disturbance. In general, the unknown nonlinear
function f  is usually assumed to have a known or unknown
constant bound for which an adaptive estimation may be made.
However, in practical application, it is often difcult to decide the
function form of bound for some complex nonlinear function in, e.g.,
chemical process, social nance, etc. Hence, it is necessary to
investigate the kind of systems without the information on the
bound. To this end, a neural-network-method will be proposed to
approximate the unknown function in Section 3.1 later, which
provides an effective method to deal with the unknown nonlinear
functions.
On the other hand, as discussed in Section 1, the actuators may
fail or performance degradation in actual implementation. Generally, the outputs of faulty actuators may have arbitrary signals
different from normal controller outputs, and these signals will
act on the controlled system as unexpected control inputs. In this
work, we consider the case that the actuator failures are modeled
as outage, that is, the outputs of faulty actuators are assumed to
be zero. Moreover, the set of actuators that are susceptible to
failures is denoted as O Df1; 2, . . . ,mg. The other set of actuators
that is not subject to failures and essential to stabilize a given
system is therefore denoted as O D f1; 2, . . . ,mgO. Accordingly,
the control matrix B may be decomposed as follows:
B BO BO ,

where BO and BO are the control matrices associated with the set
O and O , respectively, i.e., BO and BO are generated by zeroing out
the columns corresponding to O and O. Furthermore, let L D O
correspond to a particular subset that actually experiences failures, and the decomposition similar to (5) be as follows:
B BL BL ,

where BL and BL have meanings analogous to those of BO and BO :

Remark 1. In practical plants, not only reliable components need
to be utilized, but also some redundancies components, including
actuator, sensor, etc., are often adopted so as to design a truly
reliable control systems. In this work, this set O of actuators is
usually redundant when considering system stabilization. However, it is just the existence of these redundant actuators that
make the reliable control possible. It should also be pointed out
that these redundant actuators are also useful for improving
control system performance [14].
3. The design of adaptive neural control
3.1. Neural network approximation
In this work, an RBF neural network will be employed to
approximate the unknown nonlinear function f xt,xtd. Thus,
for an arbitrary constant m 4 0, f  can be modeled as
nT

f Z W fZ ot,

where Z xT t,xT tdT , W n A Rln is the ideal weight, ot is

the approximation error satisfying JotJ r m, and fZ A Rl is the
basis function vector, whose ith element is a Gaussian type of
function dened as fi Z exp JZci J2 =ri2 i 1, . . . ,l with
ci A Rm and ri A R representing the center and the spread of the ith
basis function, respectively. Both the parameters ci and ri may be
chosen via some methods as in [13,4].
Substituting (7) into (1)(3) yields
^ T fZ
_ A DAtxt Ad DAd txtd But W
xt
T

~ fZ ot,
W

zt Cxt,
xt jt,

9
t A d,0:

10
n

In general, the ideal weight W is unknown and needs to be

^ be the estimate of Wn,
estimated in the controller design. Let W
~ W
^ W n . In the
and the corresponding estimation error be W
design later, we will develop an adaptive law for the weight
^ , and then, the adaptive neural control method will be
estimate W
provided.
Remark 2. From the viewpoint of mathematics, a multilayer
neural network with enough hide layers can approximate any
well-behaved nonlinear function with arbitrary accuracy. However, in practical application, the limited hidden neurons will
inevitably yield the error of network approximation. In this work,
the H1 technique will be utilized to attenuate the error effect of
neural approximation.
3.2. Neural control without actuator failures
In this subsection, we will rstly investigate the problem of
robust stability for the system (8)(10) without actuator failures.
To this end, the following controller based on delayed states is
designed:
ut Kxt Kd xtd ur t,

11

where the control gains K and Kd will be designed later, and the
robust term ur(t) is given as
ur t 

BT Pxt
JBT PxtJ2 d

^ T fZ,
xT tPW

12

with d 40 a small constant, and the matrix P 40 satises some

sufcient conditions later.
Substituting (11) into (8) yields the closed-loop system as
follows:
_ A BK DAtxt Ad BK d DAd txtd Bur t
xt
^ T fZW
~ T fZ ot:
W

13

Theorem 1. Suppose that the controller gain K and Kd are given. If

there exist matrix X 40, and scalars ei 4 0 (i1,2,3) satisfying linear
matrix inequality (LMI):
0
1
Y1 XET
X
XETd
B
C
B EX e1 I
0
0 C
B
C o 0,
14
B X
0 C
0
e2 I
@
A
0
0
e3 I
Ed X
with

Y1 A BKX XA BKT e1 e3 HHT e2 A BK d A BK d T ,

T. Jia et al. / Neurocomputing 74 (2011) 37533759

^ is updated according to the

and the estimate of network weight W
following adaptive rule:
^_ GfZxT tPJxtJJPJW
^ ,
W
T

15

1

with G G 4 0 and P X , then the control law (11)(12) can

guarantee the system (8)(10) to be uniformly ultimately bounded.
Proof. Choose Lyapunov function candidate as
Z t
~ T G1 W
~
xT tRxtdt trW
Vt xT tPxt

16

td
1 T
with R e1
2 I e3 Ed Ed 40.
Then, along the state trajectories of the closed-loop system (13),
one has for any scalars ei 40 (i1,2,3)

V_ t rxT tPA BK A BKT P P DAt DAT tP e1

2 I

3755

8T Z0, the closed-loop system can be ensured to be robust

stability for given disturbance attenuation level g (with bias s,
s is a positive constant), i.e.
Z T
Z T
JztJ2 dt r g2
JotJ2 dt s for 8T 4 0:
23
0

Theorem 2. Given g 4 0: If there exist matrix X 4 0, and scalars

e1 4 0 (i 1,2,3) satisfying LMI:
0
1
S1 XC T XET
X
XETd
B CX
I
0
0
0 C
B
C
B
C
C o 0,
B EX
e
I
0
0
0

24
1
B
C
C
B
0 A
0
0
e2 I
@ X
0
0
0
e3 I
Ed X

T
T
e1
3 Ed Ed xt 2x tPAd BK d DAd txtd

with

^ T fZ 2xT tP ot2xT tP W
~ T fZ
2x tPBur t 2x tP W

S1 A BKX XA BKT e1 e3 HHT e2 A BK d A BK d T g2 I,

~_ xT tde1 I e1 ET Ed xtd

~ T G1 W
2trW
2
3
d
rxT tY2 xt

2JBT PxtJ2
JBT PxtJ2 d

then, the adaptive control law (11), (12) and (15) can guarantee the
uncertain system (8)(10) to be uniformly ultimately bounded with
the performance (23) for admissible parameter uncertainties.

^ fZ 2xtP ot
xT tP W

Proof. it follows from (17), (19), and (20) that

^ T fZ2JxtJJPJtrW
~ TW
^
2xT tP W
T

rx tY2 xt

2d

T
1
V_ t rxT tPA BK A BKT P e1 PHHT P e1
1 E E e2 I

^ T fZ
x tP W
T

:BT Pxt: d

^ ,
~ TW
2xtPot2JxtJJPJtrW

T
2
PPxt
e2 PAd BK d Ad BK d T P e3 PHHT P e1
3 Ed Ed g

T
1
Y2 PA BK A BKT P e1 PHHT P e1
1 E E e2 I

18

2d

Hz, o r xT tS2 xt JxtJJPJJfZJ2 JW n J2F ,

^ T fZ r2JxtJJPJJW
^ JF JfZJ
xT tPW

^ J2 JfZJ2
r JxtJJPJJW
F

Hz, o V_ t ztT ztg2 otT ot,

S2 PA BK A BKT P e1 PHHT P e2 PAd BK d Ad BK d T P

T
1
1 T
T
2
e3 PHHT P e1
PP:
1 E E e2 I e3 Ed Ed C C g

~
2trW

^ Z JW
^ J2 JW n J2 :
W
F
F

It is noted that if there exist matrix P 4 0 and scalars e1 40

(i 1,2,3) satisfying S2 o 0, we get from (27)
21

Hence, if there exist matrix P 40, and positive scalars

e1 (i1,2,3) satisfying Y2 o 0, we get from (21)
V_ t rlmin Y2 JxtJ2 JxtJJPJ2m 99fZJ2 JW n J2F
rl1 JxtJr1 JxtJ,

28

20

Then, it follows from (17) to (20) that

V_ t rxT tY2 xt JxtJJPJ2m JfZJ2 JW n J2F :

27

with
19

and
T

26

we have

Note JotJ r m, and the facts

JBT PxtJ2 d

25

where this expression 2xT tPot r g2 xT tPPxt g2 otT ot is

utilized.
Dene

with

T
e2 PAd BK d Ad BK d T P e3 PHHT P e1
3 Ed Ed :

g2 otT ot JxtJJPJJfZJ2 JW n J2F ,

17

Hz, o r l2 JxtJr2 JxtJ,

29

n 2
2 JPJl JW JF .

Thus, we obtain
with l2 lmin S2 and r
Hz, o r 0 for JxtJ Z r2 =l2 . And then, it follows from (26) that
when S2 o 0 and JxtJ Z r2 =l2 , we have
V_ t r g2 otT otztT zt:

30

22

with l1 lmin Y2 , and r1 JPJ2m l JW n J2F , where the scalar l is the dimension of basis function vector fZ in (7). Thus, we
have V_ t r 0 as JxtJ Z r1 =l1 . This implies that all signals x(t)
~ are bounded, and x(t) is uniformly ultimately bounded.
and W
Furthermore, by Schurs complement, it can be shown that
Y2 o 0 holds if there exist matrix X 40, and scalars ei 4 0
(i1,2,3) satisfying LMI (14). This concludes the proof. &
In the above, it is shown the adaptive control law (11), (12),
and (15) can ensure the closed-loop system to be uniformly
ultimately bounded. In the following, we shall further show that
RT
if the approximation error ot A L2e , i.e., 0 JotJ2 dt o 1 (for

Integrating expression (30) from t 0 to t T yields

Z T
Z T
JztJ2 dt g2
JotJ2 dt:
VTV0 r 

31

Since VT Z0, it follows from (31) that for JxtJ Z r2 =l2 ,

Z T
Z T
JztJ2 dt r g2
JotJ2 dt V0 for 8T Z0:

32

It is easily shown that when JxtJ o r2 =l2 , we have

Z
0

JztJ2 dt o TJCJ2 r22 =l2

for 8T Z 0:

33

3756

T. Jia et al. / Neurocomputing 74 (2011) 37533759

Substituting (11) and (39) into (35) yields

Hence, it is obtained from (32) and (33) that

Z T
Z T
JztJ2 dt r g2
JotJ2 dt s for 8T Z 0
0

34

_ A DAt0:5BL BT Pxt Ad DAd tBL BT QPxtd

xt
L
L
^ T fZW
~ T fZ ot:
BL ur t W

2
2
2 =l2 g.

with s maxfV0, TJCJ r

This means that if there exist matrix P 4 0, and scalars e1 4 0
(i1,2,3) satisfying S2 o 0, the resultant closed-loop system
composed of (8)(10), (11), (12) and (15) has L2gain r g (with
bias b for admissible uncertainties.
It is easily shown that S2 o 0 holds if there exist matrix X 40,
and positive scalars ei (i 1,2,3) satisfying LMI (24). Moreover, it is
further noted from (18) to (28) that S2 Y2 C T C g2 PP. This
implies that S2 o0 results in Y2 o0. So, it follows from Theorem 1
~ are bounded, and x(t) is
that when S2 o0, all signals x(t) and W
uniformly ultimately bounded. &

40

1
Choose Lyapunov function candidate as (16) with R e1
2 I e3
e
we have

T
ETd Ed 1
4 PQBL BL QP 4 0,

V_ t r xT tPA AT PPBL BTL P e1 PHHT P e2 PAd ATd P e3 PHHT P

T
1
1 T
1
T
e1
1 E E e2 I e3 Ed Ed e4 PQBL BL QPxt
T
T
2xT tPBL BTL QPxtde1
4 x tdPQBL BL QPxtd
T

^ fZ2xT tP W
~ fZ
2xT tPBL ur t 2xT tP W
T 1 _
T
T
~ r x tPA AT PPB BT P
~ G W
2x tPot 2trW
L L
T
1
1 T
e1 PHHT P e2 PAd ATd P e3 PHHT P e1
1 E E e2 I e3 Ed Ed
T
T
T
e1
4 PQBL BL QP e4 PBL BL Pxt 2x tPBL ur t

Remark 3. In Theorems 1 and 2, the form of the pre-xed control

gain K and Kd is not explicitly given. Hence, this provides some
more freedom degree for designing controller with certain performance guarantee on the controlled system. In the sequel, a
method to decide these control gains will be further given.

^ fZ2xT tP W
~ fZ 2xT tP ot
2xT tPW
T 1 _
T
~
~
2trW G W r x tPA AT PPBO BTO P e1 PHHT P

3.3. Neural control subject to actuator failures

e4 PBBT P g2 PPxt g2 otT ot 2xT tPBL ur t

T
T
1
1 T
1
e2 PAd ATd P e3 PHHT P e1
1 E E e2 I e3 Ed Ed e4 PQBB QP

^ :
^ T fsZ2JxtJJPJtrW
~ TW
2xT tPW
In the sequel, we shall consider enforcing the reliability
requirement on the systems (8)(9), in which actuator failures
may happen within a prescribed subset O of the control channel.
By means of decomposition (6), the systems (8)(10) can be
rewritten as

Note that
^ T fZ 2xT tPB ur t
2xT tP W
L
^ T fZ
2xT tP W

^ fZ
_ A DAtxt Ad DAd txtd BL ut W
xt
~ T fZ ot,
W

35

zt Cxt,
xt jt,

36
t A d,0:

37

The following theorem provides a sufcient condition ensuring

the robust stability of uncertain systems (1)(3) for all admissible
actuator failures corresponding to any L D O.
Theorem 3. For a given constant g 4 0, if there exist matrices X 40,
Q 40, and scalars e1 40 (i1,2,4) satisfying LMI:
0
1
X1 XC T XET
X
XETd
QB
B
C
B CX
I
0
0
0
0 C
B
C
B EX
0
0
0 C
0
e1 I
B
C
38
B
C o0,
B X
0
0 C
0
0
e2 I
B
C
BE X
0
0
0
e3 I
0 C
@ d
A
0
0
0
0
e4 I
BT Q

X1 AX XA BO BO e1 e3 HH e

T
T
2 Ad Ad 4 BB

2

I,

and these weights are adaptively updated as (15), then the controller
(11)(12) with control gains as
K 0:5BT X 1 ,

Kd BT QX 1 ,

39

can guarantee that the closed-loop system is uniformly ultimately

bounded with L2gain r g despite actuators failures corresponding
to L D O.
Proof. From decomposition (6) and BL BO BOL , it can be shown
that BL K 0:5BL BT X 1 , BL Kd BL BT QX 1 and BO BT rBL BT .
L

2JBT PxtJ2

L
^ T fZ
xT tP W
JBT PxtJ2 d

^ T fZ
r 2xT tP W

2JBT PxtJ2

O
^ T fZ
xT tPW
JBT PxtJ2 d

2JBTO PxtJ2 d

^ fZ
xT tP W
JBT PxtJ2 d
^ JF JfZJ,
r 2JxtJJPJJW

42

and the expression (19), we get from (41)

V_ t r xT tPA AT PPBO BTO P e1 PHHT P e2 PAd ATd P e3 PHHT P
T
T
T
1
1 T
1
2
e1
PPxt
1 E E e2 I e3 Ed Ed e4 PQBB QP e4 PBB P g

g2 otT ot JxtJJPJJfZJ2 JW n J2F r xT tX2 xt

JxtJJPJJfZJ2 JW n J2F g2 otT otztT zt,

43

with
T
X2 PA AT PPBO BTO P e1 e3 PHHT P e2 PAd ATd P e1
1 E E
T
T
1 T
1
2
e1
PP C T C:
2 I e3 Ed Ed e4 PQBB QP e4 PBB P g

44
If there exist matrices P 4 0, Q 4 0, and scalars e1 40 (i 1,2,3,4)
satisfying X2 o 0, we have

with
T

41

V_ t r l3 JxtJr2 JxtJ g2 otT otztT zt,

45

n 2

where l3 lmin X2 , and r2 JPJl JW JF as in Theorem 2.

Thus, when JxtJ Z r2 =l3 , one can obtain
V_ t r g2 otT otztT zt,
Hence, by utilizing the arguments similar to Theorem 2, it can
be shown that the closed-loop systems (8)(10) with (11) is
uniformly ultimately bounded with L2gain r g for all admissible
uncertainties and actuators failures corresponding to L D O.

T. Jia et al. / Neurocomputing 74 (2011) 37533759

Moreover, it can be shown that X2 o 0 holds if there exist positive

denite matrices X and Q, and positive scalars e1 (i 1,2,3,4)
satisfying LMI (38). This concludes the proof. &
Remark 4. In the above design, the delayed state is assumed to
be available. In general, the performance of a memoryless controller cannot be better than a delayed feedback controller which
utilizes the available information of delayed states [10]. Moreover, the delayed state feedback controller has less conservatism
than the memoryless controller.
Actually, for the case that the information of delayed state is
not available, a memoryless control law may be designed as
follows:
ut Kxt ur t,

46

with K 0:5BT X 1 and robust term ur(t) as (12).

Substituting the controller (46) into systems (35)(37), we get
_ A DAt0:5BS BT Pxt Ad DAd txtd BL ur t
xt
S
T

^ fZW
~ fZ ot,
W
zt Cxt,
xt jt,

47
48

t A d,0:

49

By adopting similar arguments to Theorem 3, the following

sufcient condition for the robust stability of the systems
(47)(49) can be easily obtained from Theorem 3.
Corollary 1. Given a constant g 4 0, if there exist matrix X 4 0, and
scalars e1 4 0, e2 4 0, and e3 40 satisfying the following LMI:
0
1
P1 XC T XET
X
XETd
B CX
I
0
0
0 C
C
B
B
C
B EX
0
0 C
0
e1 I
50
C o0,
B
C
B
0 A
0
0
e2 I
@ X
Ed X

e3 I

with

P1 AX XAT BO BTO e1 e3 HHT e2 Ad ATd g1 I,

and the updating rule of network weight given as (15), then the
memoryless control law (46) with (12) can guarantee that the
closed-loop system is uniformly ultimately bounded with L2-gain
r g, for all admissible uncertainties as well as all actuators failures
corresponding to L D O.
Proof. By adopting similar arguments to the proof procedure of
Theorem 3, for Lyapunov function V(t) as (16), it is easily obtained
that
V_ t rxT tP2 xt r99xt99 g2 otT otztT zt,

51

where

P2 PA AT PPBO BTO P e1 e3 PHHT P e2 PAd ATd P

T
1
1 T
2
e1
PP C T C:
1 E E e2 I e3 Ed Ed g

52

Thus, if there exist positive denite matrix P 4 0, and positive

constants e1 , e2 , and e3 satisfying P2 o0, we obtain
V_ t rl4 JxtJr2 JxtJ g2 otT otztT zt,

53

with l4 lmin P2 . This implies that V_ t r g2 otT otztT zt

for JxtJ Z r2 =l4 . Hence, the result can be concluded by using
similar arguments as in Theorem 3. &
Remark 5. From (44) and (52), one has X2 P2 e1
4
PQBBT QP e4 PBBT P. This implies that P2 o0 holds if X2 o0,

3757

i.e., lmin X2 o lmin P2 . Thus, we have r2 =l3 Z r2 =l4 . Hence,
it is shown from Theorem 3 and Corollary 1 that if the delayed
state feedback controller (11) guarantees the closed-loop system
(40) to be robust H1 reliable stable, the memoryless state feedback
controller (46) also ensures the closed-loop system (47) robust
H1 reliable stable. Moreover, the convergent zone of system
states with (11) is larger than the one with the controller (46).
4. Numerical simulation
Consider the uncertain system
follows:
0
1
0
0:2
1 1 0
B
C
B
0 A, Ad @ 0:2
A@1 1
0 0 3
0


0:2 0
0
C
,
0 0 0:2
0:2 T ,

H 0:1

0:2

0:2

C
0 A,
1
0

5
B
B@ 0
1

E0

0:2

(1)(2) with parameters as

0:1
0:3
0:2

0:1 ,

0:2

C
0 A,
0

E0

0:2

0:1 ,

O f3g,

1
x1 tx2 t
!C
B
1
B
C
C
x texp 
f xt,xtd B
2 t C:
B 3
1
x
3
@
A
1 x1 tdx3 td

Apparently, it is difcult to decide the bound of nonlinear

function f xt,xtd. Here, the nonlinear function f xt,xtd
will be approximated by means of neural network (7). Hence, the
objective for the above system is to design a neural-networkbased H1 reliable control law such that the closed-loop system is
robustly stable with disturbance attenuation g for all admissible
actuators failures corresponding to L D O. For g 0:8, the desired
control gain matrices in (39) can be derived by solving LMI (38):
0
1
5:4880
18:0134 0:0211
B
C
K @ 3:6029 12:9559 0:0010 A,
0:2180 0:7196
0:0287
0
1
0:3313
1:0825 0:0352
B
C
Kd @ 0:2473 0:8893 0:0003 A,
0:0887
0:2947
0:0449
and the updating rule of network weights in (15) is given as
0
1
_
2:1958 7:2058
0:0032
B
^ :
^ GfZxT t@ 7:2058 25:9118 0:0020 C
W
A27:9295GJxtJW
0:0032

0:0020

0:0581

The initial value of state x(t) is ( 1,0,1), time delay d 1, and

the RBF neural network (7) is selected with 15 hidden layer
^ 0 to be zero. The simulation
neurons with the initial values W
results are shown in Figs. 13 in which the third actuator is failure
at t 2. It is shown that the present reliable controller can
effectively cope with the effects of actuators failures and ensures
the asymptotic stability of controlled systems.

5. Conclusion
In this work, the control design subject to actuator failure has
been considered for a class of uncertain systems with unknown
nonlinear function. The key feature of the present work is to
approximate unknown nonlinear functions such that the requirement on the bound of the nonlinear unknown function has been

3758

T. Jia et al. / Neurocomputing 74 (2011) 37533759

removed. Although the present work has been only considered for
the case with stated delay, it should be pointed out that the
proposed method may be extended to some more complex
systems, e.g., stochastic system, networked control system, etc.

2.5
2
1.5

Acknowledgments

The research was partially supported by the National Natural

Science Foundation of China (61004062, 61074041) and Shanghai
Leading Academic Discipline Project (B504), Shanghai Educational
Development Foundation.

0.5
0
0.5

References

1
1.5

Fig. 1. The trajectory of state x1 t.

1.2
1
0.8
0.6
0.4
0.2
0
0.2
0.4

Fig. 2. The trajectory of state x2 t.

1.5

0.5

Tinggang Jia is now pursuing his PhD degree in the

School of Information Science and Engineering, East
China University of Science and Technology. His current research interests include automation control and
application, robust control, networked control systems.

0.5

[1] H. Alwi, C. Edwards, Fault tolerant control using sliding modes with on-line
control allocation, Automatica 44 (2008) 18591866.
[2] J. Gao, B. Huang, Z. Wang, D.G. Fisher, Robust reliable control for a class of
uncertain nonlinear systems with time-varying multistate time delays, Int.
J. Syst. Sci. 32 (2001) 817824.
[3] Z. Gao, S.X. Ding, Actuator fault robust estimation and fault-tolerant control
for a class of nonlinear descriptor systems, Automatica 43 (2007) 912920.
[4] T. Holcomb, M. Morari, Local training of radial basis function networks:
toward sloving the hidden unit problem, in: Procedings of the American
Control Conference, 1991, pp. 23312336.
[5] C.-H. Hong, K.-C. Choi, B.-S. Kim, Applications of adaptive neural network
control to an unmanned airship, Int. J. Control Autom. Syst. 7 (2009)
911917.
[6] Y.-W. Liang, S.-D. Xu, C.-L. Tsai, Study of vsc reliable designs with application
to spacecraft attitude stabilization, IEEE Trans. Control Syst. Technol. 15
(2007) 332338.
[7] Y. Liu, Z. Wang, J. Liang, et al., Stability and synchronization of discrete-time
markovian jumping neural networks with mixed mode-dependent timedelays, IEEE Trans. Neural Network 20 (2009) 11021116.
[8] Y. Liu, Z. Wang, X. Liu, Asymptotic staility for neural networks with mixed
time-delays: the discrete-time case, Neural Network 22 (2009) 6774.
[9] Y. Liu, Z. Wang, W. Wang, Robust reliable control for discrete-time-delay
systems with stochastic nonlinearities and multiplicative noises, Optim.
Control Appl. Methods 32 (2011) 285297.
[10] Y.S. Moon, P. Park, W.H. Kwon, Y.S. Lee, Delay-dependent robust stabilization
of uncertain state-delayed systems, Int. J. Control 74 (2001) 14471455.
[11] Y. Niu, J. Lam, X. Wang, D.W.C. Ho, Neural adaptive sliding mode control for a
class of nonlinear neutral delay systems, J. Dyn. Syst. Meas. Control 130
(2008) 17.
[12] Y. Niu, X. Wang, Sliding mode control design for uncertain delay systems
with partial actuator degradation, Int. J. Syst. Sci. 40 (2009) 403409.
[13] R.M. Sanner, J.E. Slotine, Gaussian networks for direct adaptive control, IEEE
Trans. Neural Network 3 (1992) 837863.
[14] C.-J. Seo, B.K. Kim, Robust and reliable H1 control for linear systems with
parameter uncertainty and actuator failure, Automatica 32 (1996) 456467.
[15] Z. Wang, Y. Liu, X. Liu, State estimation for jumping recurrent neural
networks with discrete and distributed delays, Neural Network 22 (2009)
4148.
[16] G.-H. Yang, D. Ye, Reliable H1 control of linear systems with adaptive
mechanism, IEEE Trans. Autom. Control 55 (2010) 242247.
[17] Y. Zhang, L. Guo, H. Yu, K. Zhao, Fault tolerant control based on stochastic
distributions via mlp neural networks, Neurocomputing 70 (2007) 867874.
[18] Q. Zhu, S. Fei, T. Zhang, T. Li, Adaptive rbf neural-networks control for a class
of time-delay nonlinear systems, Neurocomputing 71 (2008) 36173624.

Fig. 3. The trajectory of state x3 t.

T. Jia et al. / Neurocomputing 74 (2011) 37533759

Yuanyuan Zou is a lecturer in Control Science and

Engineering, East China University of Science Technology. Her research interests include nonlinear control,
hybrid control systems, sliding mode control and
networked control systems.

3759

Yugang Niu is a professor in Control Science and

Engineering, East China University of Science Technology. He is now serving as an Associate Editor for
International Journal of Systems Science. His research
interests include nonlinear control, stochastic control
systems, sliding mode control and network congestion
control.