IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 24, NO.

1, JANUARY 2013 71

Observer-Based Adaptive Neural Network Control

for Nonlinear Stochastic Systems With Time Delay
Qi Zhou, Peng Shi, Senior Member, IEEE, Shengyuan Xu, and Hongyi Li

Abstract— This paper considers the problem of observer-based output-feedback stabilization for a class of nonlinear stochastic
adaptive neural network (NN) control for a class of single- systems. However, there still exists the issue of explosion of
input single-output strict-feedback nonlinear stochastic systems complexity in the backstepping design procedure because the
with unknown time delays. Dynamic surface control is used to
avoid the so-called explosion of complexity in the backstepping backstepping design requires repeated differentiation of virtual
design process. Radial basis function NNs are directly utilized controllers, which makes the complexity of controller grow
to approximate the unknown and desired control input signals dramatically as the order of the system increases. In order
instead of the unknown nonlinear functions. The proposed to avoid the above-mentioned problem, the dynamic surface
adaptive NN output feedback controller can guarantee all the control (DSC) technique was first introduced in [17] for a class
signals in the closed-loop system to be mean square semi-globally
uniformly ultimately bounded. Simulation results are provided to of strict-feedback nonlinear system with unknown functions.
demonstrate the effectiveness of the proposed methods. Then, this approach was extended to solve a class of nonlinear
systems with periodic disturbances in [18]. More recently, the
Index Terms— Adaptive control, backstepping, dynamic
surface control, fuzzy control, nonlinear systems. DSC approach was further extended to solve the problem
of output-feedback adaptive control of stochastic nonlinear
systems in [19].
I. I NTRODUCTION
On the other hand, it is well known that the method of

I N THE past decade, research on stochastic systems has

received considerable attention due to the fact that sto-
chastic disturbance exists in many practical systems and is
fuzzy logic control or NN control is useful to approximate
unknown nonlinear functions in systems (see [20]–[35] and the
references therein). Recently, the NN approximation approach
often the source of instability [1]–[5]. On the other hand, has been utilized to handle the nonlinear stochastic systems
controller design for nonlinear systems has always been an control design problems [36]–[39]. Since the general nonlinear
important topic since practical stochastic system models are function f (x) can be approximated by an NN as f (x) =
largely nonlinear. Many nonlinear stochastic control results θ T ϕ(x)+ε(x), the method can also be used in stochastic non-
and methods, such as sliding mode control [6] and T-S fuzzy linear systems. However, it should be noted that the number
control [7], have been reported in the literature. In recent of adaptation laws depends on the number of the NN nodes,
years, the well-known backstepping method has been applied and therefore the number of parameters to be estimated will
to solve the problem of nonlinear stochastic systems (see [8]). increase when the NN nodes increase, which will lead to an
The backstepping design for nonlinear stochastic single-input unacceptably large learning time. This problem was considered
single-output (SISO) systems was investigated in [9] and [10] in [40] by considering the norm of ideal weighting vector in
by introducing quartic Lyapunov functions. More recently, the fuzzy logic systems as the estimation parameter instead of the
results have been extended to output feedback control design elements of weighting vector, and was also studied in [41] by
for stochastic nonminimum-phase nonlinear systems [11], developing a direct adaptive method for a class of nonlinear
large-scale stochastic nonlinear systems [12], and high-order systems.
nonlinear stochastic systems [13]–[15]. In [16], the results It is well known that time delays are frequently encountered
have also been extended to adaptive neural network (NN) in engineering systems, and they usually become the sources
Manuscript received May 21, 2011; revised September 18, 2012; accepted of instability and degrade the performance of the systems
September 26, 2012. Date of publication November 29, 2012; date of current [42]–[47]. Therefore, how to control nonlinear stochastic time-
version December 18, 2012. This work was supported in part by the National delay systems is still an important and practical topic. In [48]
Key Basic Research Program, China, under Grant 2012CB215202, the Qing
Lan Project, the 111 Project under Grant B12018, the National Natural Science and [49], the observer-based control problem for nonlinear
Foundation of China under Grant 61174058, Grant 60974071, and Grant stochastic systems with time delay was solved. However, there
61203002, and the Engineering and Physical Sciences Research Council, U.K., exists a common weakness in the work of [48] and [49] in that
under Grant EP/F029195.
Q. Zhou and S. Xu are with the School of Automation, Nanjing the time delays are only present in the system output. Time
University of Science and Technology, Nanjing 210094, China (e-mail: delays in the state variables were not considered, which affect
[email protected]; [email protected]). the stability and performance of the systems.
P. Shi is with the School of Electrical and Electronic Engineering, The
University of Adelaide, Adelaide 5005, Australia, and also with the School Motivated by the above observation, in this paper the prob-
of Engineering and Science, Victoria University, Melbourne 3052, Australia lem of observer-based adaptive control for nonlinear stochastic
(e-mail: [email protected]). time-delay systems is investigated. The main contributions of
H. Li is with the College of Information Science and Technology, Bohai
University, Jinzhou 121013, China (e-mail: [email protected]). this paper are summarized as follows. 1) The DSC approach
Digital Object Identifier 10.1109/TNNLS.2012.2223824 is successfully applied to nonlinear stochastic systems, which
2162–237X/$31.00 © 2012 IEEE
72 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 24, NO. 1, JANUARY 2013

avoids repeated differentiating the virtual controller by intro- Remark 1: Some restrictions on the delay terms are pre-
ducing a first-order filter in each step of backstepping design sented in this assumption, which are essential for the controller
procedure. 2) A direct adaptive radial basis function NNs design in the sequel. In addition, compared with [48] and [49],
(RBF NNs) control method is proposed, where the number the time delays considered in this paper exist not only in the
of adaptation parameters is only one, which significantly system output but also in the state variables.
reduces the computation burden. Furthermore, the developed
controller design is independent of any prior knowledge of B. Stochastic Stability
NNs. 3) The time delays exist not only in the system output Consider the stochastic system
but also in the state variables, which makes the developed
results in this paper more applicable. It is proven that the dx (t) = f (x (t)) dt + g (x (t)) dw (2)
controller designed in this paper guarantees all the signals in where x ∈ R n is the system state, w is an r -dimentional
the closed-loop to be mean square semi-globally uniformly standard Wiener process, and f : R n → R n , g : R n → R n
ultimately bounded (M-SGUUB). Finally, simulation results are locally Lipschitz functions and satisfy f (0) = g(0) = 0.
are given to illustrate the effectiveness of the proposed control Definition 1: For any given V (x) ∈ C 2 , associated with the
algorithm. stochastic system (2), the infinitesimal generator L is defined
The rest of this paper is organized as follows. The prob- as follows:
lem formulation and preliminaries, stochastic stability, NN  
∂V 1 ∂2V
approximation, and output feedback DSC design procedure for LV (x) = f (x) + Tr g (x)T g (x)
SISO strict-feedback nonlinear stochastic time-delay systems ∂x 2 ∂x2
are presented in Section II. A simulation example is provided where Tr(A) is the trace of a matrix A.
in Section III to demonstrate the effectiveness of the proposed Definition 2 [36]: The solution process {x(t), t ≥ t0 } of
design scheme. The conclusions are drawn in Section IV. stochastic system (2) with initial condition x 0 ∈ S0 (some
compact set containing the origin) is said to be M-SGUUB
II. P ROBLEM F ORMULATION AND P RELIMINARIES if for any desired escape risk ε (0 < ε < 1), it is bounded
with probability 1 − ε in some compact set S(ε) ⊃ S0 , i.e.,
In this section, some preliminaries and assumptions are
inf x0∈S0 P{τ S(ε) = ∞} ≥ 1−ε. The hitting time τ S(ε) is defined
first formulated in Section II-A, and then the definition of
stochastic stability is presented in Section II-B. In order to as the first time that the trajectories of the state variable reach
the boundary of S(ε).
approximate the unknown nonlinear functions, RBF NN are
Lemma 2 [36]: Consider the stochastic system (2). If there
presented in Section II-C. By using the backstepping approach,
exists a positive-definite radially unbounded twice continu-
an output feedback DSC design procedure is formulated in
ously differentiable Lyapunov function V : R n → R, and
Section II-D.
constants c1 > 0, c2 > 0, and r0 > (c2 /c1 ) such that for some
ε (0 < ε < 1) and x 0 ∈ S0 := {x ∈ R n |V (x) ≤ r0 }
A. Problem Formulation   r0 

LV (x) ≤ −c1 V (x)+c2 , ∇x ∈ S (ε) = x ∈ R n V (x) ≤
Consider the nonlinear stochastic system ε
dx i (t) = (x i+1 (t) + f i (x̄ i (t)) + h i (x̄ i (t − τi ))) dt holds, then, for t ∈ [t0 , τ S(ε)], there is a unique solution of (2)
and the system is bounded with probability 1 − ε in S(ε).
+gi (y)T dw, 1≤i ≤n
dx n (t) = (u + f n (x (t)) + h n (x (t − τn ))) dt C. NN Approximation
+gn (y)T dw In this paper, RBF NN will be used to approximate the
y (t) = x 1 (t) (1) unknown smooth nonlinear functions. For any continuous
unknown smooth nonlinear function f (Z ) over a compact set
where x̄ i (t) = [x 1 (t), x 2 (t), . . . , x i (t)]T ∈ R i , i =
 Z ⊂ R q , there exists a RBF NN W ∗T S(Z ) such that for a
1, 2, . . . , n − 1, and x(t) = [x 1(t), x 2 (t), . . . , x n (t)]T ∈ R n
desired level of accuracy ε
denote state vectors of the system; u ∈ R and y ∈ R are
the input and output of the system, respectively. f i (.) (i = f (Z ) = W ∗T S (Z ) + δ (Z ) , |δ (Z )| ≤ ε (3)
1, 2, . . . , n) stands for the unknown smooth system function
where W∗ is the ideal constant weight vector and defined by
with f i (0) = 0. h i (.) is the unknown smooth nonlinear time-  
delay functions, τi is the unknown constant delay, and τm is the  
∗  
W = arg min sup  f (Z ) − W S (Z ) .
T
upper bound of τi . gi (.) is the unknown vector-valued smooth W ∈R N Z ∈ Z
functions with gi (0) = 0. w is an independent r -dimentional
standard Wiener process. δ(Z ) is the approximation error, W = [w1 , . . . , w N ]T is the
Lemma 1 [50] (Young’s Inequality): For ∀(x, y) ∈ R 2 , the weight vector and S(Z ) = [s1 (Z ), . . . , s N (Z )]T is the basis
following inequality holds: function vector with N being the number of the RBF NN nodes
and N > 1. RBF si (Z ) = exp[(−(Z − μi )T (Z − μi )/ηi2 )],
εp 1
xy ≤ |x| p + q |y|q i = 1, 2, . . . , N, where μi = [μi1 , μi2 , . . . , μin ]T is the
p qε center of the receptive field and ηi is the width of the Gaussian
where ε > 0, p > 1, q > 1, and ( p − 1)(q − 1) = 1. function.
ZHOU et al.: ADAPTIVE NN CONTROL FOR NONLINEAR STOCHASTIC SYSTEMS WITH TIME DELAY 73

D. Output Feedback DSC Design where z̄ l = [z 1 , . . . , z l ]T , ᾱl f = [α1 f , α2 f , . . . , αl f ]T ,

In this subsection, an observer-based controller is designed α1 f = 0.
to guarantee all the signals in the closed-loop system to be Remark 2: Since g(y(t)) is a smooth function and g(0) =
SGUUB. First, we design the following observer to estimate 0, g(y(t)) can be expressed as
the unmeasured states:
g (y (t)) = y ḡ (y) = [y ḡ1 (y) , . . . , y ḡn (y)] .
.

x̂ i = x̂ i+1 + li y − x̂ 1 i = 1, . . . , n (4)
The backstepping design procedure contains n steps. In each
where x̂ n+1 = u. step, a virtual control function α̂i should be developed using
Let x̃ = x − x̂, where x̂ = [x̂ 1 , . . . , x̂ n ], be the observer an appropriate Lyapunov function Vi , and then the real control
error, which satisfies the equation law u will be designed. To begin with the backstepping design
procedure, let us define a constant
d x̃ (t) = (A x̃ (t) + f (x̄ (t)) + h (x̄ (t − τ ))) dt+ g (y (t))T dw    
θ = max Ni Wi∗  : i = 0, 1, 2, . . . , n .
2
where
⎡ ⎤
−l1 From the definition, we know θ is an unknown constant, and
⎢ ⎥ we define θ̂ as the estimate of θ . And the feasible virtual
A = ⎣ ... In−1 ⎦
control signal is designed as
−ln 0 . . . 0
 T 1 3
f (x̄ (t)) = f 1 (x 1 (t)) · · · f n (x n (t)) αi (X i ) = − z i θ̂ , i = 1, . . . , n − 1 (6)
 T 2ai2
h (x̄ (t − τ )) = h 1 (x 1 (t − τ1 )) · · · h n (x n (t − τn ))
 T where X 1 = x 1 , X i = (x̃ 1 , x̂¯i , ᾱi f , α̇i f )T , i = 2, . . . , n − 1,
g (y) = g1 (y (t)) · · · gn (y (t))
with x̂¯i = (x̂ 1 , x̂ 2 , . . . , x̂ i )T .
and A is a strict Hurwitz matrix. So there exists a matrix Remark 3: From Assumption 1, we have
P > 0 satisfying the following equation:
z 13 h 1 (x 1 (t − τ1 ))
A P + P A = −I.
T
(5) 1 3 4
≤ 4 z 14 (t − τ1 ) 11
4
(z l (t − τ1 )) + c33 z 14 (7)
4c3 4
Then the entire system can be expressed as
where c3 is a positive constant.
d x̃ (t) = (A x̃ (t)+ f (x̄ (t))+h (x̄ (t − τ ))) dt+g (y (t))T dw

Now we are ready to present the main result of this paper.
dy (t) = x̂ 2 (t) + x̃ 2 (t) + f 1 (y (t)) + h 1 (y (t − τ1 )) dt Theorem 1: Consider the stochastic nonlinear time-delay
+g1 (y (t))T dw system in (1) with (4). If a control law is chosen as

dx̂ 2 (t) = x̂ 3 (t) + l1 x̃ 1 (t) dt 1 3

u=− z θ̂
.. 2an2 n
.
dx̂ n (t) = (u + ln x̃ 1 (t)) dt with the intermediate virtual control signals αi described as
(6) and the adaptive law defined as
where y and x̂ i , (i = 1, . . . , n) are available for the controller
 r
n
design. θ̂˙ = z 6 − k0 θ̂
2 i
(8)
In order to avoid the problem of explosion of complexity, 2a i
i=1
DSC approach is introduced in this part, and the backstepping
design is, based on the change of coordinates, as follows: where positive constants ai (i = 1, . . . , n), r , and k0 are design
parameters, then the closed-loop system can be guaranteed to
z1 = y be M-SGUUB with probability 1 − ε in (ε).
z i = x̂ i − αi f , i = 2, 3, . . . , n Proof:
Step 1: Consider the Lyapunov function
where αi f is the output of the first-order filter with αi−1 as
the input. Ito’s differentiation rule yeilds a T 2 1 1
VZ1 = x̃ P x̃ + z 14 + θ̃ 2

2 4 2r
dz 1 = x̂ 2 (t) + x̃ 2 (t) + f 1 (y (t)) + h 1 (y (t − τ1 )) dt t
1
+g1 (y (t))T dw VQ 1 = 4 z 14 (τ ) l14 (z 1 (τ )) dτ

4c3
dz i = x̂ i+1 (t) + li x̃ 1 (t) − α̇i f dt, i = 2, . . . , n. t −τ1
i t
a4  
n
Assumption 1: For 1 ≤ i ≤ n, there exist positive unknown

+ z l4 (τ ) il4 z̄ l (τ ) + ᾱl f (τ ) dτ
smooth functions il (z̄ l + ᾱl f ) such that 2
i=1 l=1 t −τi

i

V1 = V Z 1 + VQ 1
|h i (x̄ i )| ≤ |z l | il z̄ l + ᾱl f
l=1 where a > 0 and θ̃ = θ − θ̂ .
74 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 24, NO. 1, JANUARY 2013

According to Definition 1, one has Using Lemma 1, the following inequalities can be obtained:
     
LV1 = −a x̃ T P x̃  x̃2 + a x̃ T P x̃ 2 x̃ T P ( f + h) 2aTr g (z 1 ) 2P x̃ x̃ T P + x̃ T P x̃ P g (z 1 )T
3    
+ z 12 g1 (z 1 )T g1 (z 1 )  
2 ≤ 2an g (z 1 ) 2P x̃ x̃ T P + x̃ T P x̃ P g (z 1 )T 
    √
+2aTr g (z 1 ) 2P x̃ x̃ T P + x̃ T P x̃ P g (z 1 )T ≤ 6an nz 12 g (z 1 )2 P2 x̃2

√
3a 2 n n 4 √
+z 13 x̂ 2 + x̃ 2 + f1 (z 1 ) + h 1 (z 1 (t − τ1 )) ≤ z 1 g (z 1 )4 + 3n nc12 P4 x̃4 (12)
2
c1
− θ̃ θ̂˙ + 4 z 14 l14 (z 1 )
1 1
r 3 4 1 3 4 1
4c3 z 13 x̃ 2 ≤ c23 z 14 + 4 x̃ 24 ≤ c23 z 14 + 4 x̃4 . (13)
1 4 4 4c2 4 4c2
− 4 z 1 (t − τ1 ) 11 4
(z 1 (t − τ1 ))
4c3 Applying inequalities (10)–(13) and (7)–(9), we have
a4 
n 
i

+ z l4 il4 z̄ l + ᾱl f
1
2 LV1 ≤ − x̃4 + z 13 x̂ 2 + f¯1 − θ̃ θ̂˙
i=1 l=1 r
a4 
n 
i
a 4 
n i

− z l4 (t − τi ) + z l4 il4 z̄ l + ᾱl f
2 2
i=1 l=1 i=2 l=2

×il4 z̄ l (t − τi ) + ᾱl f (t − τi ) . (9) a4 a4 3

+ P8 θ 2 + P8 ε04 − z 14
2 2 4
As f  ( f1 (x̄) , . . . , fn (x̄))T , and f i (x̄) , i = 1, 2, . . . , n is
an unknown function, by Lemma 1, for any given εi0 > 0, where
there exists RBF NN Wi0 ∗T S (X ) such that
i0 0 3n (n + 1) 8
∗T  = aλmin (P) − 3 − P 3
f i (X 0 ) = Wi0 S0 (X 0 ) + δi0 (X 0 ) 4
√ 2 1
|δi0 (X 0 )| ≤ εi0 −3n nc1 P − 4 4
4c2
where X 0 = x, X 0 ∈  X 0 = {X 0 |x ∈ x } and x is defined  √
¯ 3 34 3 34 3a 2 n n
as a compact set through which the state trajectories may f 1 (X 1 ) = f 1 (z 1 ) + c + c3 +  ḡ (z 1 )4
travel. Therefore 4 2 4 c12
1 3
f (X 0 ) = W0∗T S0 (X 0 ) + δ0 (X 0 ) + 4 11 4
(z 1 ) + ḡ1T (z 1 ) ḡ1 (z 1 )
4c3 2
δ0 (X 0 ) ≤ ε0 . 
a 4 n
3
4
As S0T S0 ≤ N0 is used and N0 is the dimension of S0 , and + 11 (z 1 ) + z1.
2 4
according to the definition of θ , we know W0∗ 4 S04 ≤ θ 2 . i=1
Therefore, the following inequality holds when X 0 ∈  X 0 :
  Now, take the intermediate control signal α̂1 (X 1 ) as
2a x̃ T P x̃ x̃ T P f = 2a x̃ T P x̃ x̃ T P W0∗T S0 (X 0 ) + δ0 (X 0 )

α̂1 (X 1 ) = − k1 z 1 + f¯1
3  4 a4
 x̃4 +
≤ P8 W0∗  S04
2 2 where k1 > 0. Then, we have
3 a4
+ x̃ + 4
P8 δ04

LV1 ≤ − x̃4 + z 13 x̂ 2 − α̂1 − k1 z 14 − θ̃ θ̂˙

1
2 2
a4 a4 r
= 3 x̃4 + P8 θ 2 + P8 ε04 . (10) 3 4 a4 a4
2 2 − z1 + P θ +
8 2
P ε0
8 4
4 2 2
According to Assumption 1, the following inequality holds:
a4   4 4
n i

2a x̃ T P x̃ x̃ T Ph ≤ 2a x̃3 P2 h + z l il z̄ l + ᾱl f . (14)
2
i=2 l=2
n i
≤ 2a x̃3 P2 z l (t − τi )
However, α̂1 (X 1 ) is an unknown nonlinear function as it
i=1 l=1

contains f 1 (z 1 ), which cannot be implemented in practice.
×il z̄ l (t − τi ) + ᾱl f (t − τi ) Therefore, according to (3), for any given constant ε1 > 0,
a4   4
n i there exists RBF NN W1∗T S1 (X 1 ) such that
≤ z l (t − τi )
2
i=1 l=1

α̂1 (X 1 ) = W1∗T S1 (X 1 ) + δ1 (X 1 )
×il4 z̄ l (t − τi ) + ᾱl f (t − τi ) |δ1 (X 1 )| ≤ ε1
3n (n + 1) 8
+ P 3 x̃4 . (11)
4 where X 1 ∈  X 1 = {X 1 |x ∈ x }.
ZHOU et al.: ADAPTIVE NN CONTROL FOR NONLINEAR STOCHASTIC SYSTEMS WITH TIME DELAY 75

From the definition of θ and α1 , we have Similarly, we have

−z 13 α̂1 = −z 13 W1∗T S1 (X 1 ) − z 13 δ1 (X 1 )

m−1 
m−1 
m−1
N1 6  
W ∗ 2 + 1 a 2 + 3 z 4 + 1 ε4 LVm ≤ − x̃4 + z i3 z i+1 + z i3 χi+1 − ki z i4
≤ z 1 1 (15)
2a12 2 1 4 1 4 1 i=1 i=1 i=1
m−1 
≤
1 6 1 3
z 1 θ + a12 + z 14 + ε14
1
(16) 1  r ˙
2a12 2 4 4 + θ̃ z 6 − θ̂ + m−1
2 i
(19)
r 2a i
i=1
1
z 13 α1 = − 2 z 16 θ̂ (17)  
 χi+1
m−1 4
2a1 − − χi+1 Bi+1 (X i )
3
κi+1
where the inequality S1T S1 ≤ N1 is used and N1 is the i=1
dimension of S1 . Then, substituting (16) and (17) into (14) a4 
n 
i

yields + z l4 il4 z̄ l + ᾱl f

2

i= m+1 l=m+1
LV1 ≤ − x̃4 + z 13 x̂ 2 − α1 − k1 z 14
3 4
  +z m
3
x̂ m+1 + f¯m − z m (20)
1 r 6 ˙ 4
+ θ̃ z 1 − θ̂ + 1
r 2a12 where
a4 
n 
i

 
+ z l4 il4 z̄ l + ᾱl f (18) 3 a4  4
n

2
i=2 l=2 f¯m (X m ) = lm x̃ 1 − α̇m f + z m + lm z̄ m + ᾱm f .
4 2
i=m
where
a4 a4 1 1 Take the intermediate control signal α̂m (X m ) as
1 = P8 θ 2 + P8 ε04 + a12 + ε14.
2 2 2 4

By the definition of x̂ 2 = z 2 + α2 f , (18) can be rewritten as α̂m = − km z m + f¯m

LV1 ≤ − x̃4 + z 13 z 2 + α2 f − α1 − k1 z 14
  where km > 0. Then, adding and subtracting α̂m (X m ) in (20)
1 r 6 ˙ yields
+ θ̃ z 1 − θ̂ + 1
r 2a12

m−1 
m−1
a4   4 4
n i

LVm ≤ − x̃ + 4
z i3 z i+1 + z i3 χi+1
+ z l il z̄ l + ᾱl f .
2 i=1 i=1
i=2 l=2 m−1 

m
1  r ˙
In order to avoid repeatedly differentiating α1 , a new state − ki z i + θ̃
4
z i − θ̂
6
variable α2 f is introduced, and let α1 pass through a first- i=1
r
i=1
2ai2
order filter with time constant κ2 to obtain α2 f as  
 χi+1
m−1 4
3
κ2 α̇2 f + α2 f = α1 , α2 f (0) = α1 (0). − − χi+1 Bi+1 (X i ) + m−1
κi+1
i=1
Let χ2 = α2 f − α1 be the output error of this filter; then one a4 
n 
i

has α̇2 f = −(χ2 /κ2 ) and + z l4 il4 z̄ l + ᾱl f

2
χ2 i= m+1 l=m+1
χ̇2 = α̇2 f − α̇1 = − + B2 (X 1 )
3 4
κ2 +z m
3
x̂ m+1 − α̂m − z m . (21)
4
where
B2 (X 1 ) =
3 2 1 ˙
z 1 ż 1 θ̂ + 2 z 13 θ̂. Similarly, α̂m (X m ) can be approximated by the RBF NN
2
2a1 2a1 Wm∗T Sm (X m ) as
Then, it implies
  α̂m (X m ) = Wm∗T Sm (X m ) + δm (X m )
4 1 r 6 ˙
LV1 ≤ − x̃ + z 13 z 2 + z 13 χ2 + θ̃ z 1 − θ̂ |δm (X m )| ≤ εm
r 2a12
a4   4 4
n i

where X m ∈  X m = {X m |x ∈ x }. And we can obtain
−k1 z 14 + z l il z̄ l + ᾱl f + 1 .
2
i=2 l=2
1 6 1 2 3 4 1 4
Step m: (2 ≤ m ≤ n − 1). Choose the Lyapunov function −z m
3
α̂m ≤ 2
z m θ + am + zm + εm (22)
2am 2 4 4
candidate 1
1 4 1 3
zm αm = − 2 z m 6
θ̂ . (23)
Vm = Vm−1 + z m + χm4 . 2am
4 4
76 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 24, NO. 1, JANUARY 2013

Then, by substituting (22) and (23) into (21), we have where

 
  3 a4 4

f¯n (X n ) = ln x̃ 1 − α̇n f + z n
m−1 m−1
+ nn z̄ l + ᾱl f .
LVm ≤ − x̃4 + z i3 z i+1 + z i3 χi+1 4 2
i=1
 m i=1
 Take the intermediate control signal α̂n (X n ) as α̂n = −(kn z n +

m
1  r ˙ f¯n ), where kn > 0; then, adding and subtracting α̂n (X n ) in
4 6
− ki z i + θ̃ z i − θ̂
i=1
r
i=1
2ai2 (26) yields
  
n−1 
n−1 
n
 χi+1
m−1 4
− − χi+1 Bi+1 (X i )
3 LVn ≤ − x̃4 + z i3 z i+1 + z i3 χi+1 − ki z i4
κi+1 i=1 i=1 i=1
i=1

n−1 
+m + z m 3
x̂ m+1 − αm 1  r ˙
3
+ θ̃ 2
z i − θ̂ + z n3 u − α̂n − z n4
6

a4 
n 
i

r
i=1
2ai 4
+ z l4 il4 z̄ l + ᾱl f (24)  
2
i= m+1 l=m+1

n−1
χi+1
4
− − χi+13
Bi+1 (X i ) + n−1 . (27)
where κi+1
i=1

1 1
m m
a4 a4 Similar to the above steps, α̂n (X n ) can be approximated by
m = P8 θ 2 + P8 ε04 + ai2 + εi4 . the RBF NN Wn∗T Sn (X n ) as
2 2 2 4
i=1 i=1
α̂n (X n ) = Wn∗T Sn (X n ) + δn (X n )
Next, introduce a new variable αm+1 f , and let αm pass through
a first-order filter with the constant κm+1 to obtain αm+1 f |δn (X n )| ≤ εn

κm+1 α̇m+1 f + αm+1 f = αm , αm+1 f (0) = αm (0) . (25) where X n ∈  X n = {X n |x ∈ x }.

Now, define the closed-loop state variables x c =
Then define χm+1 = αm+1 f − αm as the output error of this (x̃, z̄ n , χ̄n , θ̃ , ᾱn f )T , where χ̄n = (χ1 , . . . , χn ) and the initial
filter. We have α̇m+1 f = −(χm+1 /κm+1 ) and condition is x 0c = (x̃(t0 ), z̄ n (t0 ), χ̄n (t0 ), θ̃ (t0 ), ᾱn f (t0 ))T . For
χ some constant c0 > ( ¯ n /c̄), where  ¯ n and c̄ are given below,
χ̇m+1 = α̇m+1 f − α̇m = − m+1 + Bm+1 (X m ) let 0 = {x c |Vn ≤ c0 }, (ε) = {x c |Vn ≤ (c0 /ε)}, and
κm+1
the initial value of Lyapunov function Vn (t0 ) ≤ c0 . Then
where define the approximation region with respect to parameter ε,
3 2 1 3 ˙
Bm+1 (X m ) = 2
z m ż m θ̂ + 2 z m θ̂. (0 < ε < 1) as
2am 2am   n  t
1 
Substituting (25) into (24) yields  1
 X n (ε) = X n  z i4 + 4 z 14 (τ ) l14 (z 1 (τ )) dτ
4 4c 3 t −τ1

m 
m i=1

LVm ≤ − x̃4 + z i3 z i+1 + z i3 χi+1 
a4   t 4
n i

c0
i=1 i=1 + z l (τ ) il z̄ l (τ ) + ᾱl f (τ ) dτ ≤
4
.
 m  2 t −τi ε

m
1  r 6 ˙ i=1 l=1
4
− ki z i + θ̃ z − θ̂ Therefore, X n (t0 ) ∈  X n (ε) and x c (t0 ) ∈ 0 ⊂  (ε). If
r 2ai2 i
i=1

i=1
 x c (t) ∈  (ε), then X n (t) ∈  X n (ε), which means τε ≤ τ X n ,
 χi+1
m−1 4
where τε is the first time x c reaches the boundary of  (ε)
− − χi+1
3
Bi+1 (X i ) + m and τ X n is the first time X n reaches the boundary of  X n (ε) .
κi+1
i=1
Following a similar procedure and by the definition of u,
a4 
n 
i

we have
+ z l4 il4 z̄ l + ᾱl f .
2 1 6 1 3 1
i= m+1 l=m+1 −z n3 α̂n ≤ z θ + an2 + z n4 + εn4 (28)
2an2 n 2 4 4
Step n: Consider the following Lyapunov function: 1
1 1 z n u = − 2 z n6 θ̂ . (29)
Vn = Vn−1 + z n4 + χn4 . 2an
4 4 Then, by substituting (28) and (29) into (27), when x c (t) ∈
Similarly, we obtain  (ε), for any t ∈ [t0 , τε ]

n−1 
n−1 
n−1

n−1 
n−1
LVn = − x̃4 + z i3 z i+1 + z i3 χi+1 − ki z i4 LVn ≤ − x̃4 + z i3 z i+1 + z i3 χi+1
i=1 i=1 i=1 i=1 i=1
n−1   n 
1  r ˙

n
1  r ˙
+ θ̃ 2
z i − θ̂ + n−1 + z n3 u + f¯n
6
− ki z i + θ̃
4
z i − θ̂
6
r
i=1
2ai i=1
r
i=1
2ai2
   

n−1
χi+1
4
3 
n−1
χi+1
4
− − χi+1 Bi+1 (X i ) − z n4
3
(26) − − χi+1 Bi+1 (X i ) + n
3
κi+1 4 κi+1
i=1 i=1
ZHOU et al.: ADAPTIVE NN CONTROL FOR NONLINEAR STOCHASTIC SYSTEMS WITH TIME DELAY 77

where 3
x1
1 1
n n
a4 a4
n = P8 θ 2 + P8 ε04 + ai2 + εi4 . 2
2 2 2 4
i=1 i=1

By the definition of θ̂˙ , we can get

1


n−1 
n−1 
n
0
LVn ≤ − x̃4 + z i3 z i+1 + z i3 χi+1 − ki z i4 + n
i=1 i=1 i=1
  −1
k0 
n−1
χi+1
4
+ θ̃ θ̂ − − χi+1
3
Bi+1 (X i ) . (30)
r κi+1 −2
i=1
By utilizing Lemma 1, we have
−3
3 1 4 0 10 20 30 40 50 60
z i3 z i+1
≤ z i4 + z i+1 Time(sec)
4 4
3 3 4 1 4 Fig. 1. Trajectory of system output y.
z i χi+1 ≤ z i + χi+1
4 4
 
 3  3 4 43 4 1 filter with αi as input. Therefore, the repeated differentiation
χi+1 Bi+1  ≤ π Bi+1 χi+1 +
3
4
 
4π 4 of αi is avoided.
1 1 Remark 5: By using the method of direct adaptive NN
θ̃ θ̂ = θ̃ θ − θ̃ ≤ − θ̃ 2 + θ 2 (31)
2 2 control, there is only one adaptive parameter θ̂ in the virtual
where π > 0 is a design constant and Bi+1 is a continuous control αi and true control u. Therefore, the computation
function. Therefore, there exists a positive constant Ni+1 such burden is greatly reduced. In addition, it is not necessary to
that |Bi+1 | ≤ Ni+1 . Substituting (31) into (30), one has a priori know the centers of the receptive field and the width
 n   of the Gaussian functions, which makes the obtained theory
7 4 more useful in practice.
LVn ≤ − x̃4 − ki − z
4 i
i=1
III. S IMULATION R ESULTS
n−1 
 
1 1 3 4 43
+ − − π 3 Bi+1 χi+1
4 In this section, a simulation example is presented to demon-
ki+1 4 4 strate the effectiveness of the proposed adaptive NN control
i=1
k0 2 method.
− ¯n
θ̃ + 
2r Consider the following stochastic nonlinear time-delay
k 1 system:
¯ n = n + 0 θ 2 +
 .  
2r 4π 4 x 3 (t − τ )
1
Let dx 1 = x 2 − 0.3x 1 + x 12 + 1
dt
1 + x 12 (t − τ1 )
>0
+x 1 sin (x 1 ) dw
   
and denote 0.2 x 14 (t − τ2 ) sin (x 2 )
   dx 2 = u + 0.1x 2 sin + dt
2 7 1 + x 12 1 + x 12 (t − τ2 )
c = min , 4 ki − ,
aλ2max (P) 4 x1
   + dw
4
1 1 3 4 43
− − π Bi+1 , k0
3 i = 1, . . . , n. 1 + x 12
ki+1 4 4 y = x1
Then one has where the nonlinear functions are f 1
(x 1 (t)) = −10x 1 + x 12 ,
  f 2 (x̄ 2 (t)) = 0.1x 2 sin 0.2/(1 + x 12 ) , nonlinear time-delay
¯ n.
LVn ≤ −c Vn − VQ 1 + 
terms are defined as h 1 (x̄ 1 (t − τ1 )) = (x 13 (t − τ1 )/
Then, there exists a positive constant c̄ such that 1 + x 12 (t − τ1 )), and h 2 (x̄ 2 (t − τ2 )) = (x 14 (t − τ2 ) sin (x 2 )/
1 + x 12 (t − τ2 )), the initial states are chosen as x 1 (0) = 0.2,
¯ n , ∇x c ∈  (ε).
LVn ≤ −c̄Vn +  x 2 (0) = 0.5, and the observer is designed as
Considering c0 > ( ¯ n /c̄) and τε ≤ τ X n , it can be concluded .

x̂ 1 = x̂ 2 + l1 y − x̂ 1
that all the signals in the closed-loop system are M-SGUUB .

with probability 1 − ε in  (ε), i.e., inf x0c ∈0 P {τε = ∞} ≥ x̂ 2 = u + l2 y − x̂ 1 .

1−ε, which are the desired results and the proof is completed. According to Theorem 1, the virtual control function α1 and
Remark 4: In the process of controller design, the DSC the true control law u are chosen, respectively, as
technique is employed to make the control scheme simple.
1 1
That is because the repeated differentiation of virtual control α1 = − 2 z 13 θ̂ , u = − 2 z 23 θ̂ (32)
αi is replaced by α̇i+1 f and αi+1 f is defined by a first-order 2a1 2a2
78 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 24, NO. 1, JANUARY 2013

80 1000
x2 θ̂
60 900
800
40
700
20 600
0 500

−20 400
300
−40
200
−60 100
−80 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time(sec) Time(sec)

Fig. 2. Trajectory of system state x2 . Fig. 5. Adaptive parameter θ̂ .

80
α
2f
In the simulation, the time delays are chosen as τ1 = τ2 = 0.5,
60
therefore the upper bound of time delay is chosen as τm = 0.5,
40 and the design parameters are chosen as l1 = l2 = 100,
a1 = a2 = 0.21, r = 22.5, k0 = 0.01, and κ2 = 0.1. The
20
simulation results are illustrated in Figs. 1–5, respectively.
0 Figs. 1 and 2 show the system output y and state variable
x 2 . Fig. 3 illustrates a new state variable of a first-order filter.
−20 Fig. 4 depicts the trajectory of input u, while Fig. 5 illustrates
−40
the trajectory of adaptive parameter θ̂ .

−60
IV. C ONCLUSION
−80
0 10 20 30 40 50 60 In this paper, the problem of adaptive NN DSC was
Time(sec) investigated for a class of nonlinear stochastic strict-feedback
systems with unknown time delays. By using the backstepping
Fig. 3. Trajectory of a state variable of a first-order filter α2 f .
approach, a direct adaptive NN control scheme was proposed.
3000 In addition, by combining the method of output-feedback
u
control and DSC, the results obtained could be applied to solve
2000 the control problem for systems with unmeasurable states, and
the problem of explosion of complexity could be avoided.
1000 In addition, as the number of the online adaptive parameters
was only one, the computation burden could be reduced
accordingly. Therefore, it is convenient to implement this
0
algorithm in practical systems. Finally, a simulation example
was presented to illustrate the effectiveness of the method
−1000
proposed.

−2000
R EFERENCES
−3000
0 10 20 30 40 50 60 [1] M. Basin, J. Rodriguez-Gonzalez, L. Fridman, and P. Acosta, “Integral
Time(sec) sliding mode design for robust filtering and control of linear stochastic
time-delay systems,” Int. J. Robust Nonlinear Control, vol. 15, no. 9,
pp. 407–421, 2005.
Fig. 4. Trajectory of control input u.
[2] L. Wu and D. Ho, “Fuzzy filter design for Itô stochastic systems with
application to sensor fault detection,” IEEE Trans. Fuzzy Syst., vol. 17,
no. 1, pp. 233–242, Jan. 2009.
where z 1 = y, z 2 = x̂ 2 − α2 f . The adaptive law is given as [3] P. Shi, Y. Xia, G. Liu, and D. Rees, “On designing of sliding-mode
control for stochastic jump systems,” IEEE Trans. Autom. Control,
vol. 51, no. 1, pp. 97–103, Jan. 2006.
 r
2
θ̂˙ =
[4] Y. Chen, W. Zheng, and A. Xue, “A new result on stability analysis for
z 6 − k0 θ̂ .
2 i
(33) stochastic neutral systems,” Automatica, vol. 46, no. 12, pp. 2100–2104,
i=1
2a i 2010.
ZHOU et al.: ADAPTIVE NN CONTROL FOR NONLINEAR STOCHASTIC SYSTEMS WITH TIME DELAY 79

[5] Y. Chen, A. Xue, W. Zheng, and S. Zhou, “Robust exponential stability [28] B. Chen, X. Liu, K. Liu, and C. Lin, “Novel adaptive neural control
conditions for retarded systems with Lipschitz nonlinear stochastic design for nonlinear MIMO time-delay systems,” Automatica, vol. 45,
perturbations,” Int. J. Robust Nonlinear Control, vol. 20, no. 18, pp. no. 6, pp. 1554–1560, 2009.
2057–2076, 2010. [29] Z. Zhang, S. Xu, Y. Guo, and Y. Chu, “Robust adaptive output-feedback
[6] L. Wu and D. Ho, “Sliding mode control of singular stochastic hybrid control for a class of nonlinear systems with time-varying actuator
systems,” Automatica, vol. 46, no. 4, pp. 779–783, 2010. faults,” Int. J. Adapt. Control Signal Process., vol. 24, no. 9, pp. 743–
[7] L. Wu and W. Zheng, “L 2 –L ∞ control of nonlinear fuzzy Itô stochastic 759, Sep. 2010.
delay systems via dynamic output feedback,” IEEE Trans. Syst., Man, [30] Q. Zhou, P. Shi, J. Lu, and S. Xu, “Adaptive output feedback fuzzy
Cybern. B, Cybern., vol. 39, no. 5, pp. 1308–1315, Oct. 2009. tracking control for a class of nonlinear systems,” IEEE Trans. Fuzzy
[8] Z. Pan and T. Basar, “Adaptive controller design for tracking and Syst., vol. 19, no. 5, pp. 972–982, Oct. 2011.
disturbance attenuation in parametric strict-feedback nonlinear systems,” [31] M. Wang, C. Wang, and S. Ge, “A novel ISS-modular adaptive neural
IEEE Trans. Autom. Control, vol. 43, no. 8, pp. 1066–1083, Aug. 1998. control of pure-feedback nonlinear systems,” Int. J. Innov. Comput., Inf.
[9] H. Deng and M. Krstic, “Output-feedback stochastic nonlinear stabi- Control, vol. 7, no. 11, pp. 6559–6570, 2011.
lization,” IEEE Trans. Autom. Control, vol. 44, no. 2, pp. 328–333, Feb. [32] R. Wang and J. Li, “Adaptive neural control design for a class of
1999. perturbed nonlinear time-varying delay systems,” Int. J. Innov. Comput.,
[10] H. Deng, M. Krstic, and R. Williams, “Stabilization of stochastic Inf. Control, vol. 8, no. 5, pp. 3727–3740, 2012.
nonlinear systems driven by noise of unknown covariance,” IEEE Trans. [33] J. Yu, Y. Ma, B. Chen, H. Yu, and S. Pan, “Adaptive neural position
Autom. Control, vol. 46, no. 8, pp. 1237–1253, Aug. 2001. tracking control for induction motors via backstepping,” Int. J. Innov.
[11] S. Liu, Z. Jiang, and J. Zhang, “Global output-feedback stabilization for Comput., Inf. Control, vol. 7, no. 7, pp. 4503–4516, Jul. 2011.
a class of stochastic non-minimum-phase nonlinear systems,” Automat- [34] Q. Zhou, P. Shi, S. Xu, and H. Li, “Adaptive output feedback control for
ica, vol. 44, no. 8, pp. 1944–1957, 2008. nonlinear time-delay systems by fuzzy approximation approach,” IEEE
[12] S. Liu, J. Zhang, and Z. Jiang, “Decentralized adaptive output-feedback Trans. Fuzzy Syst., DOI:10.1109/TFUZZ.2012.2213259.
stabilization for large-scale stochastic nonlinear systems,” Automatica, [35] H. Li, H. Liu, H. Gao, and P. Shi, “Reliable fuzzy control for
vol. 43, no. 2, pp. 238–251, 2007. active suspension systems with uncertainty of actuator delay and
[13] X. Xie and J. Tian, “State-feedback stabilization for high-order stochas- fault,” IEEE Trans. Fuzzy Syst., vol. 20, no. 2, pp. 342–357, Apr.
tic nonlinear systems with stochastic inverse dynamics,” Int. J. Robust 2012.
Nonlinear Control, vol. 17, no. 14, pp. 1343–1362, 2007. [36] H. Psillakis and A. Alexandridis, “NN-based adaptive tracking con-
[14] X. J. Xie and J. Tian, “Adaptive state-feedback stabilization of high- trol of uncertain nonlinear systems disturbed by unknown covariance
order stochastic systems with nonlinear parameterization,” Automatica, noise,” IEEE Trans. Neural Netw., vol. 18, no. 6, pp. 1830–1835, Jun.
vol. 45, no. 1, pp. 126–133, 2009. 2007.
[15] X. Xie and N. Duan, “Output tracking of high-order stochastic nonlinear [37] X. Xie, N. Duan, and X. Yu, “State-feedback control of high-order
systems with application to benchmark mechanical system,” IEEE Trans. stochastic nonlinear systems with SiISS inverse dynamics,” IEEE Trans.
Autom. Control, vol. 55, no. 5, pp. 1197–1202, May 2010. Autom. Control, vol. 56, no. 8, pp. 1921–1926, Aug. 2011.
[16] W. Chen, J. Wu, and L. Jiao, “State-feedback stabilization for a class [38] L. Liu and X. Xie, “Output-feedback stabilization for stochastic high-
of stochastic time-delay nonlinear systems,” Int. J. Robust Nonlinear order nonlinear systems with time-varying delay,” Automatica, vol. 47,
Control, vol. 48, no. 2, pp. 282–286, Feb. 2003. no. 12, pp. 2772–2779, 2011.
[17] D. Wang and J. Huang, “Neural network-based adaptive dynamic surface [39] W. Li, X. Xie, and S. Zhang, “Output-feedback stabilization of stochastic
control for a class of uncertain nonlinear systems in strict-feedback high-order nonlinear systems under weaker conditions,” SIAM J. Control
form,” IEEE Trans. Neural Netw., vol. 16, no. 1, pp. 195–202, Jan. 2005. Optim., vol. 49, no. 3, pp. 1262–1282, 2011.
[18] W. Chen, “Adaptive backstepping dynamic surface control for systems [40] Y. Yang, G. Feng, and J. Ren, “A combined backstepping and small-gain
with periodic disturbances using neural networks,” IET Control Theory approach to robust adaptive fuzzy control for strict-feedback nonlinear
Appl., vol. 3, no. 10, pp. 1383–1394, Oct. 2009. systems,” IEEE Trans. Syst., Man, Cybern. A, Syst., Humans, vol. 34,
[19] W. Chen, L. Jiao, and Z. Du, “Output-feedback adaptive dynamic surface no. 3, pp. 406–420, Mar. 2004.
control of stochastic non-linear systems using neural network,” IET [41] B. Chen, X. Liu, K. Liu, and C. Lin, “Direct adaptive fuzzy control of
Control Theory Appl., vol. 4, no. 12, pp. 3012–3021, Dec. 2010. nonlinear strict-feedback systems,” Automatica, vol. 45, no. 6, pp. 1530–
[20] S. Tong, C. Liu, and Y. Li, “Fuzzy-adaptive decentralized output- 1535, 2009.
feedback control for large-scale nonlinear systems with dynamical [42] C. Hua, X. Guan, and P. Shi, “Robust backstepping control for a class
uncertainties,” IEEE Trans. Fuzzy Syst., vol. 18, no. 5, pp. 845–861, of time delayed systems,” IEEE Trans. Autom. Control, vol. 50, no. 6,
Oct. 2010. pp. 894–899, Jun. 2005.
[21] S. Tong, X. He, and H. Zhang, “A combined backstepping and small- [43] Q. Zhou and P. Shi, “A new approach to network-based H∞ control for
gain approach to robust adaptive fuzzy output feedback control,” IEEE stochastic systems,” Int. J. Robust Nonlinear Control, vol. 22, no. 9, pp.
Trans. Fuzzy Syst., vol. 17, no. 5, pp. 1059–1069, Oct. 2009. 1036–1059, 2012.
[22] Y. Liu, W. Wang, S. Tong, and Y. Liu, “Robust adaptive tracking [44] H. Li, H. Gao, and P. Shi, “New passivity analysis for neural networks
control for nonlinear systems based on bounds of fuzzy approximation with discrete and distributed delays,” IEEE Trans. Neural Netw., vol. 21,
parameters,” IEEE Trans. Syst., Man, Cybern. A, Syst. Humans, vol. 40, no. 11, pp. 1842–1847, Nov. 2010.
no. 1, pp. 170–184, Jan. 2010. [45] X. Liu and J. Cao, “Robust state estimation for neural networks with
[23] Y. Liu, C. Chen, G. Wen, and S. Tong, “Adaptive neural output feedback discontinuous activations,” IEEE Trans. Syst., Man, Cybern. B, Cybern.,
tracking control for a class of uncertain discrete-time nonlinear systems,” vol. 40, no. 6, pp. 1425–1437, Jun. 2010.
IEEE Trans. Neural Netw., vol. 22, no. 7, pp. 1162–1167, Jul. 2011. [46] Z. Feng and J. Lam, “Stability and dissipativity analysis of distributed
[24] Y. Li, S. Tong, and T. Li, “Adaptive fuzzy output feedback control of delay cellular neural networks,” IEEE Trans. Neural Netw., vol. 22, no. 6,
uncertain nonlinear systems with unknown backlash-like hysteresis,” Inf. pp. 976–981, Jun. 2011.
Sci., vol. 198, no. 1, pp. 130–146, Sep. 2012. [47] T. Li, L. Guo, C. Sun, and C. Lin, “Further results on delay-dependent
[25] H. Dong, Z. Wang, D. Ho, and H. Gao, “Robust H∞ fuzzy output- stability criteria of neural networks with time-varying delays,” IEEE
feedback control with multiple probabilistic delays and multiple missing Trans. Neural Netw., vol. 19, no. 4, pp. 726–730, Apr. 2008.
measurements,” IEEE Trans. Fuzzy Syst., vol. 18, no. 4, pp. 712–725, [48] Y. Fu, Z. Tian, and S. Shi, “Output feedback stabilization for a class of
Aug. 2010. stochastic time-delay nonlinear systems,” IEEE Trans. Autom. Control,
[26] B. Chen and X. Liu, “Fuzzy approximate disturbance decoupling of vol. 50, no. 6, pp. 847–851, Jun. 2005.
MIMO nonlinear systems by backstepping and application to chemical [49] W. Chen, L. Jiao, J. Li, and R. Li, “Adaptive NN backstepping output-
processes,” IEEE Trans. Fuzzy Syst., vol. 13, no. 6, pp. 832–847, Dec. feedback control for stochastic nonlinear strict-feedback systems with
2005. time-varying delays,” IEEE Trans. Syst., Man, Cybern. B, Cybern.,
[27] B. Chen, X. Liu, K. Liu, P. Shi, and C. Lin, “Direct adaptive fuzzy vol. 40, no. 3, pp. 939–950, Mar. 2010.
control for nonlinear systems with time-varying delays,” Inf. Sci., [50] M. Krstic, I. Kanellakopoulos, and P. Kokotovic, Nonlinear and Adaptive
vol. 180, no. 5, pp. 776–792, Mar. 2010. Control Design. New York: Wiley, 1995.
80 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 24, NO. 1, JANUARY 2013

Qi Zhou received the B.S. and M.S. degrees in Shengyuan Xu received the B.Sc. degree from the
mathematics from Bohai University, Jinzhou, China, Hangzhou Normal University, Hangzhou, China, the
in 2006 and 2009, respectively. She is currently M.Sc. degree from Qufu Normal University, Qufu,
pursuing the Ph.D. degree with the international China, and the Ph.D. degree from the Nanjing Uni-
collaboration program, University of Portsmouth, versity of Science and Technology, Nanjing, China,
Portsmouth, U.K., and the Nanjing University of in 1990, 1996, and 1999, respectively.
Science and Technology, Nanjing, China. He was a Research Associate with the Depart-
Her current research interests include fuzzy con- ment of Mechanical Engineering, University of
trol, stochastic control, and robust control. Hong Kong, Hong Kong, from 1999 to 2000. From
2000 to 2002, he was a Post-Doctoral Researcher
in CESAME with the Université catholique de
Louvain, Louvain-La-Neuve, Belgium, and the Department of Electrical
and Computer Engineering, University of Alberta, Edmonton, AB, Canada,
respectively. From 2002 to 2004, he was a William Mong Young Researcher
and an Honorary Associate Professor with the Department of Mechanical
Engineering, University of Hong Kong. Since 2002, he has been with the
School of Automation, Nanjing University of Science and Technology as a
Peng Shi (M’95–SM’98) received the B.Sc. degree Professor. His current research interests include robust filtering and control,
in mathematics from the Harbin Institute of Technol- singular systems, time-delay systems, neural networks, and multidimensional
ogy, Harbin, China, the M.Eng. degree in systems systems and nonlinear systems.
engineering from Harbin Engineering University, Dr. Xu was a recipient of the National Excellent Doctoral Dissertation Award
Harbin, the Ph.D. degree in electrical engineering from the Ministry of Education of China in 2002. He received a grant from
from the University of Newcastle, Callaghan, Aus- the National Science Foundation for Distinguished Young Scholars of China,
tralia, the Ph.D. degree in mathematics from the Uni- in 2006. He was awarded a Cheung Kong Professorship from the Ministry of
versity of South Australia, Adelaide, Australia, and Education of China in 2008. He is a member of the Editorial Boards of the
the D.Sc. degree from the University of Glamorgan, Multidimensional Systems and Signal Processing, and the Circuits Systems
Pontypridd, U.K. and Signal Processing.
He was a Lecturer with the University of South
Australia and a Senior Scientist with the Defence Science and Technology
Organisation, Edinburgh, Australia. He was a Professor with the University
Hongyi Li received the B.S. and M.S. degrees
of Glamorgan, Victoria University, Melbourne, Australia, and is currently with
in mathematics from Bohai University, Jinzhou,
the University of Adelaide, Adelaide. He has authored or co-authored widely
China, in 2006 and 2009, respectively, and the Ph.D.
in these areas. His current research interests include system and control theory,
degree in intelligent control from the University of
computational and intelligent systems, and operational research.
Portsmouth, Portsmouth, U.K., in 2012.
Dr. Shi is a fellow of the Institution of Engineering and Technology, U.K.,
He is currently with the College of Information
and the Institute of Mathematics and its Applications, U.K. He is on the
Science and Technology, Bohai University. His cur-
editorial board of a number of international journals, such as the IEEE
rent research interests include fuzzy control, robust
T RANSACTIONS ON AUTOMATIC C ONTROL, the IEEE T RANSACTIONS ON
control, and their applications in suspension systems.
S YSTEMS , M AN AND C YBERNETICS -PART B, and the IEEE T RANSACTIONS
ON F UZZY S YSTEMS .