You are on page 1of 11

IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 5, NO.

5, SEPTEMBER 2018 923

Adaptive Neural Network-Based Control for a


Class of Nonlinear Pure-Feedback Systems With
Time-Varying Full State Constraints
Tingting Gao, Yan-Jun Liu, Senior Member, IEEE, Lei Liu, and Dapeng Li

Abstract—In this paper, an adaptive neural network (NN) In [1], the adaptive dynamic programming algorithm was
control approach is proposed for nonlinear pure-feedback sys- successfully applied to a discrete-time optimal control scheme.
tems with time-varying full state constraints. The pure-feedback The backstepping technique was used in [2] and [3] to achieve
systems of this paper are assumed to possess nonlinear function
uncertainties. By using the mean value theorem, pure-feedback a significant breakthrough in the adaptive control of a nonlin-
systems can be transformed into strict feedback forms. For the ear time-delay system and nonlinear stochastic jump system.
newly generated systems, NNs are employed to approximate On the other hand, the backstepping method was also applied
unknown items. Based on the adaptive control scheme and to nonlinear strict feedback systems [4]−[8] with unknown
backstepping algorithm, an intelligent controller is designed. At items. After obtaining a series of research results in strict
the same time, time-varying Barrier Lyapunov functions (BLFs)
with error variables are adopted to avoid violating full state feedback systems, researchers began to study corresponding
constraints in every step of the backstepping design. All closed- nonlinear pure-feedback systems. In [9], an uncertain pure-
loop signals are uniformly ultimately bounded and the output feedback nonlinear system is transformed into a non-affine
tracking error converges to the neighborhood of zero, which form, and an adaptive control method was designed to solve
can be verified by using the Lyapunov stability theorem. Two the global asymptotic state stabilization problem. In modern
simulation examples reveal the performance of the adaptive NN
control approach. control theory, neural networks (NNs) have been developed as
approximators of unknown nonlinear functions. The work in
Index Terms—Adaptive control, neural networks (NNs), non-
[10] and [11] developed an adaptive dynamic programming
linear pure-feedback systems, time-varying constraints.
algorithms with a value iteration and local value iteration
for optimal control of discrete-time nonlinear systems, while
I. I NTRODUCTION NNs were proposed to solve the problems faced by unknown

I N recent decades, the control problems of various non-


linear systems have received wide attention. In the initial
stage, scholars often used matching conditions to derive the
items. For a nonlinear pure-feedback system [12], based on
the idea of backstepping, the author used the implicit func-
tion theorem to assert the existence of continuous desired
asymptotic stability of systems. With the continuous progress virtual controls, and NNs were adopted to approximate the
of nonlinear systems control technology, the limitations of controllers. Adaptive NN control schemes were used on a
matching conditions have restricted further research of systems class of output feedback nonlinear systems [13]−[16] with
control. To overcome this problem, the backstepping technique unknown functions. On account of backstepping algorithm and
has been introduced. This approach reduces the difficulty of the Lyapunov method, adaptive NN approximation techniques
the problem with decomposition of the systems to achieve were proposed in [17]−[19] for the ocean vessels, novel coax-
control purposes. Over the past two decades, there has been ial eight-rotor unmanned aerial vehicles and networked multi-
an increased interest in adaptive control technology because it agent systems, while their effectiveness have been proven. In
is a very effective tool for handling systems with uncertainties. addition, [20]−[33] also proposed an adaptive NN control
design method for nonlinear systems. In fact, due to the
Manuscript received March 20, 2018; accepted June 13, 2018. This work
was supported in part by the National Natural Science Foundation of China physical limitations of the controller as well as the mechanical
(61622303, 61603164, 61773188), the Program for Liaoning Innovative design and manufacture in the actual systems, various types
Research Team in University (LT2016006), the Fundamental Research Funds of constraint characteristics exist in many typical industrial
for the Universities of Liaoning Province (JZL201715402), and the Program
for Distinguished Professor of Liaoning Province. Recommended by Associate control systems. However, the system’s constraints are ignored
Editor Qinglai Wei. (Corresponding author: Yan-Jun Liu.) in these papers.
Citation: T. T. Gao, Y. J. Liu, L. Liu, and D. P. Li, “Adaptive neural Under the influence of the actual background, the theoretical
network-based control for a class of nonlinear pure-feedback systems with
time-varying full state constraints,” IEEE/CAA J. of Autom. Sinica, vol. 5, no. study of constraints becomes more and more important. In
5, pp. 923−933, Sep. 2018. nonlinear control system [34] based on predictive control,
T. T. Gao, Y. J. Liu, and L. Liu are with the College of Science, Liaoning the problem of satisfying pointwise-in time input and/or state
University of Technology, Jinzhou 121000, China (e-mail: gaotthh@163.com;
liuyanjun@live.com; liuleill@live.cn). hard constraints was studied, and predictive control has also
D. P. Li is with the School of Electrical Engineering, Liaoning University been applied in [35]. The control strategies of non-holonomic
of Technology, Jinzhou 121001, China (e-mail: li dapengsir@163.com). mechanical systems under various constraints were studied in
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. [36]. Constraints can be roughly divided into the following
Digital Object Identifier 10.1109/JAS.2018.7511195 types: output constraints, partial state constraints, full state
924 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 5, NO. 5, SEPTEMBER 2018


constraints, etc. In many control applications, there are strong 
ẋi = fi (x̄i , xi+1 ) , i = 1, . . . , n − 1
constraints on the control and state variables. For constrained ẋn = fn (x̄n , u) (1)


systems, designing controllers based on unconstrained condi- y = x1
tions will deteriorate the stability of the systems, produce some
undesired responses such as oscillations, and even destroy the where x̄i = [x1 , x2 , . . . , xi ]T ∈ Ri , i = 1, . . . , n are the state
hardware equipment of systems. BLF is now a widely used vectors, u ∈ R and y ∈ R are the input and the output of the
approach to control constraints in design. In [37]−[46], the systems, respectively. fi (·) : Ri+1 → R are uncertain smooth
adaptive control method for nonlinear systems with output functions. For the systems (1), the state variables xi , i = 1,
constraints and state constraints were investigated based on . . ., n are required to remain in the set |xi | < kci (t), ∀t ≥ 0,
BLF. The adaptive control design based on a BLF framework where kci (t) : R+ → R.
not only can deal with the uncertainty of parameters, but also The control objective of this paper is to construct an
can deal with the uncertainty of functions by using NNs. The adaptive feedback controller u which ensures that the system
adaptive NN control design in [47]−[53] were applied to the output y is tracks the reference signal yd , while all the closed-
BLF-based nonlinear systems with output or state constraints. loop signals are uniformly ultimately bounded and the time-
The actual systems, state parameters and control parameters varying full state constraints are not violated.
are time-varying and determined by the physical characteristics Remark 1: In this paper, nonlinear pure-feedback systems
of the restraint systems, so the model cannot be described with with state constraints are researched. Compared with the
the usual parameters. To deal with the aforementioned control nonpure-feedback systems [40]−[46], pure-feedback systems
problems of the systems, both the systems time-varying model have a wider range of applications as a class of more general
and systems constraints must be taken into account. Motivated lower triangular nonlinear systems. Moreover, in contrast to
by these advancements in the constraints, many time-varying the control of nonlinear systems [40]−[46], [48]−[50] with
output constraints have been studied for various nonlinear constant state constraints, time-varying state constraints can re-
systems [54]−[56]. In view of adaptive NN tracking control lax the conditions of constant ones. At the same time, although
technology, state constraints are considered in nonlinear time- [62] also addressed the pure-feedback systems with time-
varying systems [57]−[59]. Based on the above theories, the varying state constraints, linear parameterization is necessary.
authors applied time-varying state constraints to a robot system This paper is more general in the case that the function is
[60] and DC motor system [61], respectively. However, there is completely unknown.
no work to deal with the adaptive NN control of pure-feedback Assumption 1 [60]: There exist the constants Kcji , i = 1,
systems with time-varying constraints, which motivates our (j)
. . . , n, j = 0, 1, . . . , n such that kci (t) ≤ Kc0i and kci (t)
study. (j)
Motivated by the aforementioned observations, we pay close ≤ Kcji , ∀t ≥ 0, where kci (t) denotes jth time derivative of
attention to the problem of adaptive NN tracking control kci (t).
scheme for uncertain nonlinear pure-feedback systems with Assumption 2 [55]: It is assumed that yd (t) satisfies |yd (t)|
(i)
time-varying full state constraints in this paper. The main ≤ Y0 (t) < kci (t) and its ith time derivative yd (t), satisfy
(i)
contributions can be described as follows: |yd (t)| ≤ Yi , i = 1, . . . , n, ∀t ≥ 0, where Y0 (t) : R+ → R+
1) Many actual systems can be described as pure-feedback and Y1 , . . . , Yn are positive constants.
systems. Meanwhile, control parameters are time-varying and Lemma 1 [12]: For ∀(x, u) ∈ Rn ×R, assume that f (x, u) :
n
determined by the physical characteristics in the actual sys- R × R → R is continuous differentiable, and there is a
tems. However, as far as we know, the control of pure- constant d such that ∂f ∂u (x,u)
> d > 0. Then there exists a
feedback systems with time-varying state constraints are rarely smooth function u = u(x) such that f (x, u∗ ) = 0.

mentioned in the literature. In this paper, for the systems (1), define unknown nonlinear
2) For the sake of control of pure-feedback systems, the functions gi (x̄i , xi+1 ) = ∂fi ∂x (x̄i ,xi+1 )
i+1
, i = 1, . . . , n − 1 and
implicit function theorem is employed to assert the existence gn (x̄n , u) = ∂fn ∂u
(x̄n ,u)
.
of the smooth ideal control input. At the same time, NNs Assumption 3 [12]: The signs of functions gi (·), i = 1, . . . ,
are applied to estimate both the virtual and actual controllers. n are bounded, i.e., there exist the constants ḡi ≥ gi > 0 such
Then, the mean value theorem is used to convert pure-feedback that gi ≤ |gi (·)| ≤ ḡi . Without loss of generality, this paper
systems to strict feedback systems. assumes that gi ≤ gi (·) ≤ ḡi .
3) Aimed at transformed systems, the BLFs are employed
In [12], the radial basis function (RBF) NNs are employed
to ensure that the full state constraints were not violated in
to approximate the unknown nonlinear function Ψ(Y ) : Rl →
any step of backstepping design.
R as follows:
Furthermore, to obtain the stability of the pure-feedback
systems, a significant theorem is given and proved. The Ψnn (Y ) = W T S (Y ) (2)
feasibility of the designed controller can be proved though
two simulation examples. where Y ∈ ΩY ⊂ Rl , W = [w1 , w2 , . . . , wm ]T ∈ Rm , and m
are the input vector, the weight vector and the node number m
II. S YSTEM D ESCRIPTIONS > 1 of the NNs, respectively; while S(Y ) = [s1 (Y ), s2 (Y ),
Consider nonlinear pure-feedback systems with the follow- . . . , sm (Y )]T , where si (Y ) being defined as the commonly
ing form: used Gaussian functions, the form defined as
GAO et al.: ADAPTIVE NN-BASED CONTROL FOR A CLASS OF NONLINEAR PURE-FEEDBACK SYSTEMS 925

à !
2
− kY − τi k where kbi will be given later.
si (Y ) = exp , i = 1, 2, . . . , m (3)
ε2i Proof: For any zi that satisfies |zi | < kbi , the term log(kb2i /
(kbi − zi2 )) in (10) can be rewritten as
2
where εi is the width of the Gaussian function and τi = [τi1 , Ã !
τi2 , . . . , τil ]T is the center of the receptive field. kb2i zi2
It has been shown that any smooth function in a compact log 2 = log 1 + 2
kbi − zi2 kbi − zi2
set ΩY ⊂ Rl can be approximated by (2) to arbitrary accuracy
as  µ ¶n 
zi2
³ ´
∗T l  z2 X∞
kb2 −zi2 
Ψ(Y ) = W S(Y ) + η ∀Y ∈ ΩY ⊂ R (4)  
≤ log 1 + 2 i 2 + i

where η is the approximation error and W ∗ is the optimal  kbi − zi n=2
n! 
weight vector which is constructed as the value of W that
minimizes |η| for all Y ∈ ΩY ⊂ Rl , i.e., Ã zi2
!
½ ¾ k2 −z 2
¯ ¯ = log e bi i
W ∗ = arg minm sup ¯Ψ(Y ) − W T S(Y )¯ . (5)
W ∈R Y ∈ΩY

According to Assumption 3, we have ∂fi ∂x (x̄i ,xi+1 )


> gi > 0 zi2
i+1 = . (11)
for all x̄i+1 ∈ Ri+1 . Introduce the function φi as kb2i − zi2
 ¥

 φ1 = −ẏd + k1 z1

 Young’s Inequality: It is assumed that if x ∈ Rn , y ∈ Rn ,
φ2 = −α̇1 + k2 z2
(6) p, q > 1 and a is an arbitrary positive constant, and 1/p +
 ..

 . 1/q = 1, then


φn = −α̇n−1 + kn zn 1 p aq q
xT y ≤ p
kxk + kyk . (12)
where ki , i = 1, 2, . . . , n are positive constants, and define the pa q
tracking error z1 = x1 − yd and the variables zi = xi − αi−1 , Especially, if p = q = 2, a is an arbitrary positive constant,
i = 2, . . . , n, where αi−1 is a virtual controller to be designed then
∂φi
by Step i−1 of the backstepping method. Because ∂x = 0,
i+1 1 2 a2 2
the following inequality holds: xT y ≤ kxk + kyk . (13)
2a2 2
∂ (fi (x̄i , xi+1 ) + φi )
> gi > 0. (7)
∂xi+1 III. T HE C ONTROLLER D ESIGN AND S TABILITY A NALYSIS
Using Lemma 1 which is an implicit function theorem In this paper, based on the backstepping method, an adaptive
obtained from [12], with xi+1 , i = 1, . . . , n − 1 and u as controller is designed for the nonlinear pure-feedback systems
virtual control input, for every value of x̄i , i = 1, . . . , n and with time-varying full state constraints. The specific process
φi , i = 1, . . . , n, there exists the smooth ideal control inputs is as follows.
αi∗ (x̄i , φi ), i = 1, . . . , n, such that Step 1: Based on the systems (1), we know that differenti-
 ating z1 with respect to time yields
fi (x̄i , αi∗ ) + φi = 0, i = 1, . . . , n − 1
(8) ż1 = ẋ1 − ẏd = f1 (x1 , x2 ) − ẏd . (14)
f (x̄ , α∗ ) + φ = 0.
n n n n
Substituting (6), (8) and (9) into (14) yields
Combining the mean value theorem, there exist λi ∈ (0, 1)
that makes ż1 = −k1 z1 + g1λ1 (x2 − α1∗ ) . (15)

 ∗ ∗
fi (x̄i , xi+1 ) = fi (x̄i , αi ) + giλi (xi+1 − αi ) , By defining the error variable z2 = x2 − α1 , we express
i = 1, . . . , n − 1 (9)

 ∗ ∗ ż1 = −k1 z1 + g1λ1 (z2 + α1 − α1∗ ) . (16)
fn (x̄n , u) = fn (x̄n , αn ) + gnλn (u − αn )
∂fi (x̄i ,x
(i+1)λi ) Consider the following positive definite time-varying Lya-
where giλi = gi (x̄i , x(i+1)λi ) = ∂x(i+1)λi , i = 1, . . . , punov function
n are the defined unknown nonlinear functions with x(i+1)λi à !
= λi xi+1 + (1 − λi )αi∗ . At the same time, we suppose that 1 kb21 (t) 1
V1 = log + W̃1T Γ-11 W̃1 (17)
there exist a constant gidi > 0 such that |ġiλi (·)| ≤ gidi , and 2g1λ1 kb21 (t) − z12 2
similar to Assumption 3, we have g iλ ≤ giλi (·) ≤ ḡiλi , where
i
constants g iλ , ḡiλi > 0. where log(∆) represents the natural logarithm of ∆, W̃1 =
i
Lemma 2 [61]: For any kbi and zi , i = 1, 2, . . . , n, Ŵ1 − W1∗ , Γ1 = ΓT1 > 0 is a constant gain matrix, and kb1 (t)
satisfying |zi | < kbi , then there is = kc1 (t) − yd (t) is a barrier function.
Based on the definition of kb1 (t), and using Assumptions 1
kb2i z2
log 2 2 ≤ 2 i 2 (10) and 2, we can obtain kb1 (t) ≤ Kb1 with Kb1 being a constant.
kbi − zi kbi − zi Substituting (16) into the time derivative of V1 leads to
926 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 5, NO. 5, SEPTEMBER 2018

à ! à !
z1 k̇b1 (t) ġ1λ kb21 (t) ˙
V̇1 = ż1 − z1 − 2 1 log + W̃1T Γ-11 Ŵ1 . (23)
g1λ1 (kb21 (t) − z12 ) kb1 (t) 2g1λ1 kb21 (t) − z12
à !
ġ1λ kb21 (t) ˙ The adaptive law is designed as follows:
− 2 1 log + W̃1T Γ-11 Ŵ1 Ã !
2g1λ1 kb21 (t) − z12
˙ z1
à Ŵ1 =Γ1 − 2 S1 (Y1 ) − σ1 Ŵ1 (24)
z1 kb1 (t) − z12
= −k1 z1 + g1λ1 (z2 + α1 − α1∗ )
g1λ1 (kb21 (t) − z12 ) where σ1 > 0 is a design constant, and σ1 -modification term
! Ã ! σ1 Ŵ1 is used to improve the robustness when there is an NN
k̇b1 (t) ġ1λ1 kb21 (t)
− z1 − 2 log approximation error η1 .
kb1 (t) 2g1λ1 kb21 (t) − z12 Using Lemma 2, the following inequality is obtained
˙
+ W̃1T Γ-11 Ŵ1 . (18) k1 z12 z1 z 2 ρ1 z12
V̇1 ≤ − ¡ ¢ + −
g1λ1 kb21 (t) − z12 kb21 (t) − z12 kb21 (t) − z12
By using an NN W1T S1 (Y1 ) to approximate α1∗ (x1 , φ1 ),
where Y1 = [x1 , ẏd , yd ]T , α1∗ can be defined as η1∗2 g1d z2
+ + 2 1 2 1 2 − σ1 W̃1T Ŵ1 . (25)
α1∗ = W1∗T S1 (Y1 ) + η1 (19) 2 2ḡ1λ1 kb (t) − z1

where W1∗ is an optimal weight vector of the NN, and Step i, 2 ≤ i ≤ n − 1: Based on the systems (1), the time
approximation error η1 is bounded over the compact set, i.e., derivative of variable zi = xi − αi−1 is
|η1 | ≤ η1∗ with constant η1∗ > 0. żi = ẋi − α̇i−1 = fi (x̄i , xi+1 ) − α̇i−1 . (26)
Design the virtual control α1 as
µ µ ¶ ¶ Substituting (6), (8) and (9) into (26) yields
1
α1 = − ρ1 + ρ̄1 (t) z1 + Ŵ1T S1 (Y1 )
ḡ1λ1 żi = −ki zi + giλi (xi+1 − αi∗ ) . (27)
1
− µ ¶ (20) By defining the error variable zi+1 = xi+1 −αi , we express
z1
2 ³ ´
kb2 (t)−z12
1
żi = −ki zi + giλi (zi+1 + αi − αi∗ ) . (28)
q
where ρ̄1 (t) = (k̇b1 (t)/kb1 (t))2 + β1 , β1 > 0 is a constant. Because αi−1 is the function of x̄i−1 , yd , kbi−1 , k̇bi−1 and
Ŵ1 , . . . , Ŵi−1 , α̇i−1 can be described as
Note that even if k̇b1 (t) is zero, β1 can ensure that the time
derivative of α1 is bounded. Ŵ1 is the estimation of W1∗ and i−1
X 1
X ∂αi−1 (l+1)
∂αi−1 ∂αi−1
ρ1 is a design constant, ρ1 > 0. α̇i−1 = ẋl + ẏd + (l)
kbi−1
∂xl ∂yd ∂k
Substituting (19) and (20) into (18), on account of the fact l=1 l=0 bi−1
that ρ̄1 (t) + k̇b1 (t)/kb1 (t) ≥ 0, (18) can be further rewritten i−1
à !
as X ∂αi−1 zl
+ Γl − 2 Sl (Yl ) − σl Ŵl .
k z2 z1 z2 ρ1 z12 ∂ Ŵl kbl (t) − zl2
l=1
V̇1 ≤ − ¡ 21 1 ¢ + − (29)
g1λ1 kb1 (t) − z12 kb21 (t) − z12 kb21 (t) − z12
For simplicity, the following definition is given
z1 W̃1T S1 (Y1 ) z12 z1 η1
+ 2 2 − ¡ ¢2 − 2 1
kb1 (t) − z1 2 2
2 kb1 (t) − z1 kb1 (t) − z12 X ∂αi−1 (l+1) ∂αi−1
Θi−1 = (l)
kbi−1 + ẏd
à ! ∂yd
l=0 ∂kbi−1
ġ1λ1 kb21 (t) ˙
− 2 log + W̃1T Γ-11 Ŵ1 . (21) Ã !
2g1λ1 kb21 (t) − z12 i−1
X ∂αi−1 zl
+ Γl − 2 Sl (Yl ) − σl Ŵl .
From Young’s inequality, it follows that ∂ Ŵl kbl (t) − zl2
l=1
(30)
z1 η1 z12 η1∗2
− ≤ ¡ ¢ + . (22) Consider the time-varying Lyapunov function candidate
kb21 (t) − z12 2 kb21 (t) − z12
2 2
à !
Further, (21) can be rewritten as 1 kb2i (t) 1
Vi = Vi−1 + log + W̃iT Γ−1
i W̃i
2giλi kb2i (t) − zi2 2
k1 z12 z z
V̇1 ≤ − ¡ ¢+ 2 1 2 2 (31)
g1λ1 kb21 (t) − z12 kb1 (t) − z1
where W̃i = Ŵi − Wi∗ , and Γi = ΓTi is a positive constant
ρ1 z12 η ∗2 z1 W̃ T S1 (Y1 )
− 2 2 + 1 + 21 gain matrix.
kb1 (t) − z1 2 kb1 (t) − z12 Substituting (28) into the time derivative of Vi leads to
GAO et al.: ADAPTIVE NN-BASED CONTROL FOR A CLASS OF NONLINEAR PURE-FEEDBACK SYSTEMS 927

à ! à !
zi k̇bi (t) ˙ zi
V̇i = V̇i−1 + ¡ ¢ żi − zi Ŵi = Γi − 2 Si (Yi ) − σi Ŵi (38)
giλi kb2i (t) − zi2 kbi (t) kbi (t) − zi2
à !
ġiλ kb2i (t) ˙ where σi is a positive constant, and σi -modification term
− 2 i log + W̃iT Γ−1i Ŵi σi Ŵi is used to improve the robustness when there is a NN
2giλi kb2i (t) − zi2
approximation error ηi .
zi Using Lemma 2, the following inequality is obtained:
= V̇i−1 + ¡ 2
¢
giλi kbi (t) − zi2 ki zi2 zz
à ! V̇i ≤ V̇i−1 − ¡ ¢ + 2 i i+1 2
k̇bi (t) giλi kb2i (t) − zi2 kbi (t) − zi
× −ki zi + giλi (zi+1 + αi − αi∗ )−zi
kbi (t) ρi zi2 ηi∗2 zi−1 zi
à ! − + − 2
kbi (t) − zi2
2 2 2
kbi−1 (t) − zi−1
ġiλi kb2i (t) ˙
− 2 log + W̃iT Γ−1
i Ŵi . (32)
2giλ kb2i (t) − zi2 gidi zi2
i + − σi W̃iT Ŵi . (39)
2ḡiλi kbi (t) − zi2
2 2
By using an NN WiT Si (Yi ) to approximate αi∗ (x̄i , φi ),
where Yi = [x̄Ti , ∂α ∂αi−1 T
αi∗ can be Step n: In the last step of backstepping, the actual controller
∂x1 , . . ., ∂xi−1 , Θi−1 , αi−1 ] ,
i−1

expressed as u will be obtained. The time derivative of zn = xn − αn−1


can be written as
αi∗ = Wi∗T Si (Yi ) + ηi (33)
żn = ẋn − α̇n−1 = fn (x̄n , u) − α̇n−1 . (40)
where approximation error |ηi | ≤ ηi∗ with constant ηi∗ > 0.
The virtual control αi is designed as Substituting (6), (8) and (9) into (40) yields
µ µ ¶ ¶
1 żn = −kn zn + gnλn (u − αn∗ ) . (41)
αi = − ρi + ρ̄i (t) zi + ŴiT Si (Yi )
ḡiλi
Because αn−1 is the function of x̄n−1 , yd , kbn−1 , k̇bn−1 and
k 2 (t) − zi2 zi Ŵ1 , . . . , Ŵn−1 , the derivative of αn−1 obtain
− zi−1 2 bi 2 − ¡ 2
¢ (34)
kbi−1 (t) − zi−1 2 kbi (t) − zi2 n−1
X ∂αn−1 ∂αi−1 X ∂αn−1 (l+1) 1
α̇n−1 = ẋl + ẏd + (l)
kbn−1
where design constant ρi > 0 and the definition of ρ̄i (t) is ∂xl ∂yd ∂k
l=1 l=0 bn−1
similar to ρ̄1 (t) in Step 1.
à !
Substituting (33) and (34) into (32), and due to the fact that n−1
X ∂αn−1 zl
ρ̄i (t) + k̇bi (t)/kbi (t) ≥ 0, then (32) can be described as + Γl − 2 Sl (Yl ) − σl Ŵl .
∂ Ŵl kbl (t) − zl2
l=1
k z2 zi zi+1 (42)
V̇i ≤ V̇i−1 − ¡ 2i i ¢+
giλi kbi (t) − zi2 kbi (t) − zi2
2
For simplicity, the following definition is given
ρi zi2 zi−1 zi zi W̃ T Si (Yi ) 1
X
− 2 2 − 2 2 + 2 i ∂αn−1 (l+1) ∂αi−1
kbi (t) − zi kbi−1 (t) − zi−1 kbi (t) − zi2 Θn−1 = (l)
kbn−1 + ẏd
∂yd
l=0 ∂kbn−1
zi2 zi ηi à !
− ¡ ¢2 − n−1
2 kb2i (t) − zi2 k 2 (t) − z 2
bi i
X ∂αn−1 zl
à ! + Γl − 2 Sl (Yl ) − σl Ŵl .
∂ Ŵl kbl (t) − zl2
ġiλi kb2i (t) ˙ l=1
− 2 log + W̃iT Γ−1
i Ŵi . (35) (43)
2giλ i
kbi (t) − zi2
2
Consider the following positive definite time-varying Lya-
Using Young’s inequality yields punov function
zi ηi zi2 η ∗2 Ã !
− ≤ ¡ ¢ + i . (36) 1 kb2n (t) 1
kb2i 2
(t) − zi 2 kb2i (t) − zi2
2 2 Vn = Vn−1 + log + W̃nT Γ−1
n W̃n
2gnλn kb2n (t) − zn2 2
Further, (35) becomes (44)
ki zi2 zi zi+1 ηi∗2 where W̃n = Ŵn − Wn∗ and Γn = ΓTn > 0 is the constant
V̇i ≤ V̇i−1 − ¡ ¢ + +
giλi kb2i (t) − zi2 kb2i (t) − zi2 2 gain matrix.
ρi zi2 zi−1 zi zi W̃iT Si (Yi ) Substituting (41) into the time derivative of Vn leads to
− − + Ã !
kbi (t) − zi2
2 kb2i−1 (t) − zi−1
2 kb2i (t) − zi2 zn k̇bn (t)
à ! V̇n = V̇n−1 + ¡ ¢ żn − zn
gnλn kb2n (t) − zn2 kbn (t)
ġiλ kb2i (t) ˙
− 2 i log + W̃iT Γ−1
i Ŵi . (37) Ã !
2giλi kb2i (t) − zi2 ġnλ kb2n (t) ˙
− 2 n log + W̃nT Γ−1
n Ŵn
2gnλn kb2n (t) − zn2
Design the adaptation law Ŵi as
928 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 5, NO. 5, SEPTEMBER 2018

Ã
zn Using Lemma 2, the following inequality is obtained:
= V̇n−1 + ¡ ¢ − kn zn + gnλn (u − αn∗ )
gnλn kb2n (t) − zn2 kn zn2 ρ z2
! Ã ! V̇n ≤ V̇n−1 − ¡ 2
¢− 2 n n 2
kb2n (t) gnλn 2
kbn (t) − zn kbn (t) − zn
k̇bn (t) ġnλn
− zn − 2 log
kbn (t) 2gnλn kb2n (t) − zn2 zn−1 zn ηn∗2 zn ηn
− + − 2
˙
2
kb2n−1 (t) − zn−1 2 kbn (t) − zn2
+ W̃nT Γ−1
n Ŵn . (45)
gndn zn2
By using an NN WnT Sn (Yn ) to approximate αn∗ (x̄n , φn ), + 2 2 − σn W̃nT Ŵn . (52)
2ḡnλn kbn (t) − zn2
where Yn = [x̄Tn , ∂α∂xn−1 , . . ., ∂α T
∂xn−1 , Θn−1 , αn−1 ] , leading to
n−1

the following equation:


1
With the first n − 1 steps, we have
n
à !
X kj gjdj zj2
αn∗ = Wn∗T Sn (Yn ) + ηn (46) V̇n ≤ − + ρj − 2
ḡjλj 2ḡjλj kb2j (t) − zj2
where |ηn | ≤ ηn∗ with constant ηn∗ > 0. j=1

The following actual controller u is constructed: n


X n
X ηj∗2
µ µ ¶ ¶ − σj W̃jT Ŵj + . (53)
1 2
u = − ρn + ρ̄n (t) zn + ŴnT Sn (Yn ) j=1 j=1
ḡnλn
The following inequalities hold:
k 2 (t) − zn2 zn ³ ´
− zn−1 2 bn 2 − ¡ 2 ¢ (47) −σj W̃jT Ŵj = −σj W̃jT W̃j + Wj∗
kbn−1 (t) − zn−1 2 kbn (t) − zn2
 ° °2 
° °2 ° ° ° ∗ °2
where design constant ρn > 0 the definition of ρ̄n (t) is similar ° °  °W̃ j ° ° °
Wj 
to ρ̄1 (t) in the Step 1. ≤ −σj °W̃j ° + σj  + 
2 2
Substituting (46) and (47) into (45), and based on the
situation that ρ̄n (t) + k̇bn (t)/kbn (t) ≥ 0, while (45) can be ° °2
° ° ° °2
further expressed as σj °W̃j ° σj °Wj∗ °
≤− + . (54)
kn zn2 ρn zn2 2 2
V̇n ≤ V̇n−1 − ¡ ¢ −
gnλn kb2n (t) − zn2 kb2n (t) − zn2 By selecting the suitable ρj , the inequality (ρ∗j = ρj + kj /
2
ḡjλj − (gjdj /2ḡjλ )) > 0 is established. Then, we obtain
zn2 zn−1 zn j
− ¡ ¢2 − Ã !
2 kb2n (t) − zn2 kb2n−1 2
(t) − zn−1 kj gjdj zj2 zj2 ρ∗j
− + ρj − 2 ≤ − .
ḡjλj 2ḡjλj kb2j (t) − zj2 kb2j (t) − zj2
zn W̃nT Sn (Yn ) zn ηn
+ − 2 (55)
kb2n (t) − zn2 kbn (t) − zn2
à ! Combining Assumption 3, Lemma 2 with the above inequal-
ġnλn kb2n (t) ˙ ity, we can get the following inequality
− 2 log + W̃nT Γ−1
n Ŵn . (48) Ã !
2gnλn kb2n (t) − zn2 Xn
kb2j (t) Xn

ηj∗2
V̇n ≤ − ρj log +
Based on the following inequality: kb2j (t) − zj2 2
j=1 j=1
zn ηn zn2 η ∗2 ° °2
− ≤ ¡ ¢ + n (49) ° °2
kb2n 2
(t) − zn 2 2 n σ ° °
j °W̃j ° σj °Wj∗ °
n
2 kb2n (t) − zn2 X X
− + . (56)
(48) can be computed as follows: j=1
2 j=1
2
kn zn2 ρ z2 Theorem 1: Under Assumptions 1−3, consider the systems
V̇n ≤ V̇n−1 − ¡ 2
¢− 2 n n 2
gnλn kbn (t) − zn2 kbn (t) − zn (1) with time-varying full state constraints. If the initial states
|xj (0)| < Kc0j , j = 1, 2, . . . , n are satisfied, then the following
zn−1 zn ηn∗2 zn W̃nT Sn (Yn ) properties can be guaranteed: the error variables zj , j = 1, 2,
− + +
kb2n−1 (t) − zn−1
2 2 kb2n (t) − zn2 . . . , n are bounded; the time-varying full state constraints are
à ! never violated; and all closed-loop signals are bounded.
ġnλ kb2n (t) ˙
− 2 n log 2 + W̃nT Γ−1n Ŵn . (50) Proof: According to (17), (31) and (44), we have
2gnλn kbn (t) − zn2 n
à ! n
X 1 kb2j (t) X 1 T −1
Design the adaptation law Ŵn as Vn = log 2 (t) − z 2 + W̃j Γj W̃j .
2gjλj k 2
à ! j=1 bj j j=1
˙ zn (57)
Ŵn = Γn − 2 Sn (Yn ) − σn Ŵn (51)
kbn (t) − zn2 By choosing the ρ∗j where ρ∗j ≥ τ /2g j0 , and ρj ≥ τ /2g j0 −
2
where σn > 0 is a design constant, and σn -modification term kj /ḡjλj + (gjdj /2ḡjλ j
) where τ > 0 is a constant. Then, there
σn Ŵn is used to improve the robustness when there is an NN is the inequality σj ≥ τ λmax {Γ−1 j } by selecting τ and σj . The
approximation error ηn . derivative of Vn along (56) and (57) satisfies
GAO et al.: ADAPTIVE NN-BASED CONTROL FOR A CLASS OF NONLINEAR PURE-FEEDBACK SYSTEMS 929

à ! ° °2
n σ ° ° IV. S IMULATION E XAMPLE
j °W̃j °
n
X kb2j (t) X
V̇n ≤ − ρ∗j log − +B In this section, two simulation studies are raised to demon-
j=1
kb2j(t) − zj2 j=1
2
strate the effectiveness of the proposed adaptive NN control
n
à ! n scheme in this paper.
X τ kb2j (t) X τ W̃jT Γ−1
j W̃j
≤− log 2 (t) − z 2 − + B Example 1: Consider the third order pure-feedback system
j=1
2g jλ
k b j j j=1
2 
j 2

ẋ1 = x1 + x2
" n à ! 

X 1 kb2j (t) ẋ2 = x2 − x1 x2 + (0.1 sin x2 + 1.2) x3
≤−τ log
1 ³ ´ (64)
2g k 2 (t) − z 2
 2 −x21
b j  ẋ3 = 0.1x 1 x2 − 0.5x 2 + sin x 3 + e + 1 u
j=1 jλ j j 

# 
n T −1
y = x1
X W̃j Γj W̃j
+ + B ≤ −τ Vn + B (58) where the states of system are constrained in |x1 | < kc1 = 1.4
j=1
2
+ 0.2 sin(0.5t), |x2 | < kc2 = 2.6 + 0.6 sin(t) and |x3 | < kc3
where = 5 + 0.8 sin(t), the desired trajectory yd = 0.8 sin(2t), and
° ° 2 the virtual controllers and the actual controller are designed
X n °W ∗ ° X n
ηj∗2
j as
B= σj + . ³ ³ ´ ´
2 2 
j=1 j=1 1 T

 α1 = − ρ 1 + ḡ1λ1 ρ̄1 z1 + Ŵ1 S1 (Y1 )


τt
Multiplying e by both sides of (58) yields 


 − ³ 2 z1 2 ´
τt 
 2 kb (t)−z1
d(Vn (t)e ) 
 1
≤ Beτ t . (59) 

 ³ ³ ´ ´
dt 
α2 = − ρ2 + 1 T

 ḡ2λ2 ρ̄2 z2 + Ŵ2 S2 (Y2 )
Integrating (59) over [0, t], which leads to
µ ¶ k2 (t)−z 2 (65)
B −τ t B −τ t B 

 −z1 kb22 (t)−z22 − ³ 2 z2 2 ´
0 ≤ Vn (t) ≤ Vn (0) − e + ≤ Vn (0)e + . 
 b1 1 2 kb (t)−z2
τ τ τ 

2

 ³ ³ ´ ´
(60) 
 u = − ρ3 + ḡ3λ 1
ρ̄3 z3 + Ŵ3T S3 (Y3 )



 3
Combining (57) with (60), the following inequality can be 
 k2 (t)−z 2

 −z2 kb23 (t)−z32 − ³ 2 z3 2 ´
obtained 2 kb (t)−z3
n
à ! b2 2
3

1 X kb2j (t) B
log ≤ Vn (t) ≤ Vn (0)e −τ t
+ while the adaptation laws are designed as
2g ∗ j=1 kb2j (t) − zj2 τ " #
˙ zi
(61) Ŵi = Γi − 2 Si (Yi ) − σi Ŵi , i = 1, 2, 3 (66)
kbi (t) − zi2

where g = max1≤j≤n {ḡjλj }. And taking the e-exponential
of both sides of (61), we have where
kb2j (t) ∗ ∗
Vn (0)e−τ t + 2gτ B
z1 = x1 − yd , z2 = x2 − α1 , z3 = x3 − α2
≤ e2g . (62) T
kb2j (t) − zj2 Y1 = [x1 , ẏd ,yd ]
· ¸T
Then we can get the tracking error which satisfies T ∂α1
q Y2 = x̄2 , , Θ1 , α1
∂x1
2g ∗ B
|zj | ≤ kbj (t) 1 − e−[2g Vn (0)e + τ ] .
∗ −τ t
(63) · ¸T
∂α2 ∂α2
Y3 = x̄T3 , , , Θ2 , α2 .
According to Lemma 2, |yd (t)| ≤ Y0 and the fact x1 = z1 + ∂x1 ∂x2
yd (t), we have an inequality |x1 | ≤ |z1 |+|yd (t)| < kb1 (t)+Y0 . The initial conditions of the system are defined as x1 (0) =
Then we have |x1 | ≤ kc1 , where kc1 (t) = kb1 (t)+Y0 . Further, 0, x2 (0) = 1.4, and x3 (0) = 0, while the design parameters
there is |y| = |x1 | ≤ kc1 . Thus, the output signal is bounded. are given as follows ρ1 = 40, ρ2 = 15, ρ3 = 18, ḡ1λ1 = 0.4,
It is obvious that the virtual controller which is defined in (20) ḡ2λ2 = 0.5, ḡ3λ3 = 0.6, Γ1 = 7, Γ2 = 8, Γ3 = 4, σ1 = 0.6,
is bounded |α1 | ≤ ᾱ1 . From the fact that x2 = z2 + α1 and σ2 = 0.5, σ3 = 0.4, β1 = 6, β2 = 7 and β3 = 9.
Lemma 2, it follows that |x2 | ≤ |z2 |+|α1 | ≤ kb2 + ᾱ1 . So, |x2 | Figs. 1−4 are the simulation results of system (64). Fig. 1
≤ kc2 , where kc2 (t) = kb2 (t)+ᾱ1 . Similarly, we can prove that clearly illustrates the excellent tracking performance of the
|xi+1 | ≤ kci+1 , i = 2, . . . , n − 1, where kci+1 (t) = kbi+1 (t) + system. At the same time, Figs. 1 and 2 are given to descript
ᾱi . Based on (60), wepcan get ( 12 )W̃jT Γ−1 j W̃j ≤ Vn (0)e
−τ t
+ the trajectory of state and error variables, respectively, and it is
B/τ and kW̃j k ≤ 2λmax (Γ)(Vn (0)e −τ t + B/τ ). Hence, easy to see that the time-varying constraints are not violated.
the states xj , j = 1, 2, . . . , n, the adaptive laws Ŵj , j = 1, 2, The trajectories of the adaptive laws and actual controller are
. . . , n and the controller u are bounded. In view of the above displayed in Figs. 3 and 4. From the above Figures, it can
analysis, we can determine that all signals in the closed-loop be concluded that all closed-loop signals remain bounded and
systems are bounded and the time-varying full state constraints the time-varying constraints are never violated in the pure-
are never violated. ¥ feedback system.
930 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 5, NO. 5, SEPTEMBER 2018

Example 2: Refer to [48], the control scheme proposed in


this paper is applied to deal with the stability of the wing rock
model with ailerons modelled by first-order actuator dynamics
under the condition of time-varying full state constraints.
Meanwhile, the dynamics equation of the wing rock model
is constructed as follows:

Fig. 4. The trajectory of u.




ẋ1 = x2


ẋ = bx + θ + θ x + θ x + θ |x | x
2 3 0 1 1 2 2 3 1 2
(67)

 + θ 4 |x2 | x2 + θ x3
5 1


 x3 u
ẋ3 = − ω + ω
Fig. 1. The trajectories of yd , y = x1 , x2 and x3 .
where x1 , x2 , x3 and u are the roll angle, the roll rate, the
aileron deflection angle and the control input, respectively.
The aerodynamic parameters of delta wing for 25◦ angle of
attack are designed as b = 1.5, θ0 = 0, θ1 = −0.01859521,
θ2 = 0.015162375, θ3 = −0.062445153, θ4 = 0.00954708,
θ5 = 0.02145291 and ω = 1/15.
Figs. 5−8 are the simulation results of the practical system
(67). Fig. 5 demonstrates the excellent tracking performance
of the system, and shows that time-varying state constraints
are never violated. The trajectories of the errors, the adaptive
laws and the actual controller are plotted in Figs. 6−8. From
the above Figures, it is easy to verify the effectiveness of
the control scheme proposed in the practical experimental
research.

Fig. 2. The trajectories of z1 , z2 and z3 .

Fig. 5. The trajectories of yd , y = x1 , x2 and x3 .

V. C ONCLUSION
° ° ° ° ° ° In this paper, we have proposed adaptive NN control scheme
° ° ° ° ° °
Fig. 3. The trajectories of °Ŵ1 °, °Ŵ2 ° and °Ŵ3 °.
for the nonlinear pure-feedback systems with time-varying full
GAO et al.: ADAPTIVE NN-BASED CONTROL FOR A CLASS OF NONLINEAR PURE-FEEDBACK SYSTEMS 931

R EFERENCES
[1] Q. L. Wei, D. R. Liu, Q. Lin, and R. Z. Song, “Discrete-time optimal
control via local policy iteration adaptive dynamic programming,” IEEE
Trans. Cybern., vol. 47, no. 10, pp. 3367−3379, Oct. 2017.

[2] C. C. Hua, X. P. Guan, and P. Shi, “Robust backstepping control for a


class of time delayed systems,” IEEE Trans. Autom. Control, vol. 50,
no. 6, pp. 894−899, Jun. 2005.

[3] Y. Q. Xia, M. Y. Fu, P. Shi, Z. J. Wu, and J. H. Zhang, “Adaptive


backstepping controller design for stochastic jump systems,” IEEE
Trans. Autom. Control, vol. 54, no. 12, pp. 2853−2859, Dec. 2009.

[4] M. Wang, X. Y. Wang, B. Chen, and S. C. Tong, “Robust adaptive fuzzy


tracking control for a class of strict-feedback nonlinear systems based
on backstepping technique,” J. Control Theor. Technol., vol. 5, no. 3,
pp. 317−322, Aug. 2007.
Fig. 6. The phase portrait of z1 , z2 and z3 .
[5] P. Li and G. H. Yang, “A novel adaptive control approach for nonlinear
strict-feedback systems using nonlinearly parameterised fuzzy approxi-
mators,” Int. J. Syst. Sci., vol. 42, no. 3, pp. 517−527, Mar. 2011.

[6] T. S. Li, S. C. Tong, and G. Feng, “A novel robust adaptive-fuzzy-


tracking control for a class of nonlinear multi-input/multi-output sys-
tems,” IEEE Trans. Fuzzy Syst., vol. 18, no. 1, pp. 150−160, Feb. 2010.

[7] J. P. Yu, P. Shi, W. J. Dong, and C. Lin, “Adaptive fuzzy control


of nonlinear systems with unknown dead zones based on command
filtering,” IEEE Trans. Fuzzy Syst., vol. 26, no. 1, pp. 46−55, Feb. 2018.

[8] Y. X. Li, G. H. Yang, and S. C. Tong, “Fuzzy adaptive distributed


event-triggered consensus control of uncertain nonlinear multi-agent sys-
tems,” IEEE Trans. Syst. Man Cybern.: Syst., doi: 10.1109/TSMC.2018.
2812216.

[9] M. Z. Hou, Z. Q. Deng, and G. R. Duan, “Adaptive control of uncertain


pure-feedback nonlinear systems,” Int. J. Syst. Sci., vol. 48, no. 10, pp.
2137−2145, Jul. 2017.

[10] Q. L. Wei, D. R. Liu, and H. Q. Lin, “Value iteration adaptive dynamic


° ° ° ° ° °
° ° ° ° ° ° programming for optimal control of discrete-time nonlinear systems,”
Fig. 7. The trajectories of °Ŵ1 °, °Ŵ2 ° and °Ŵ3 °.
IEEE Trans. Cybern., vol. 46, no. 3, pp. 840−853, Mar. 2016.

[11] Q. L. Wei, F. L. Lewis, D. R. Liu, R. Z. Song, and H. Q. Lin, “Discrete-


time local value iteration adaptive dynamic programming: convergence
analysis,” IEEE Trans. Syst. Man Cybern.: Syst., vol. 18, no. 1, pp. 150−
160, Feb. 2018.

[12] S. S. Ge and C. Wang, “Adaptive NN control of uncertain nonlinear


pure-feedback systems,” Automatica, vol. 38, no. 4, pp. 671−682, Apr.
2002.

[13] B. B. Ren, S. S. Ge, K. P. Tee, and T. H. Lee, “Adaptive neural control for
output feedback nonlinear systems using a Barrier Lyapunov function,”
IEEE Trans. Neur. Netw., vol. 21, no. 8, pp. 1339−1345, Aug. 2010.

[14] F. Wang, B. Chen, C. Lin, J. Zhang, and X. Z. Meng, “Adaptive neural


network finite-time output feedback control of quantized nonlinear
systems,” IEEE Trans. Cybern., vol. 48, no. 6, pp. 1839−1848, Jun.
2018.

[15] W. Si and W. Zeng, “Adaptive neural output-feedback control for


nonstrict-feedback stochastic nonlinear time-delay systems with hys-
teresis,” IEEE/CAA J. of Autom. Sinica, 2017, doi: 10.1109/JAS.2017.
7510451.
Fig. 8. The trajectory of u. [16] H. Wang, K. Liu, X. Liu, B. Chen, and C. Lin, “Neural-based adaptive
output-feedback control for a class of nonstrict-feedback stochastic
state constraints. The use of BLF prevents time-varying state nonlinear systems,” IEEE Trans. Cybern., vol. 45, no. 9, pp. 1977−1987,
Sep. 2015.
constraints from being exceeded. The approximation property
of NNs is applied to approximate the unknown nonlinear [17] C. Peng, Y. Bai, X. Gong, Q. J. Gao, C. J. Zhao, and Y. T. Tian,
“Modeling and robust backstepping sliding mode control with adaptive
function generated by the controller design process. Though RBFNN for a novel coaxial eight-rotor UAV,” IEEE/CAA J. of Autom.
Lyapunov stability analysis and two simulation examples, the Sinica, vol. 2, no. 1, pp. 56−64, Jan. 2015.
proposed control scheme ensures that all signals of the closed-
[18] X. M. Sun and S. S. Ge, “Adaptive neural region tracking control of
loop systems are uniformly ultimately bounded and asymptotic multi-fully actuated ocean surface vessels,” IEEE/CAA J. of Autom.
tracking is implemented without violation of constraints. Sinica, vol. 1, no. 1, pp. 77−83, Jan. 2014.
932 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 5, NO. 5, SEPTEMBER 2018

[19] Y. Yang and D. Yue, “Distributed tracking control of a class of output nonlinearity,” IEEE Trans. Neur. Netw. Learn. Syst., vol. 26, no.
multi-agent systems in non-affine pure-feedback form under a directed 8, pp. 1789−1802, Aug. 2015.
topology,” IEEE/CAA J. of Autom. Sinica, vol. 5, no. 1, pp. 169−180,
Jan. 2018. [38] K. P. Tee, S. S. Ge, and E. H. Tay, “Barrier Lyapunov functions for the
control of output-constrained nonlinear systems,” Automatica, vol. 45,
[20] R. W. Zuo, X. M. Dong, Y. Chen, Z. C. Liu, and C. Shi, “Adaptive neural no. 4, pp. 918−927, Apr. 2009.
control for a class of non-affine pure-feedback nonlinear systems,” Int.
J. Control, doi: 10.1080/00207179.2017.1393106. [39] H. Y. Li, L. J. Wang, H. P. Du, and A. Boulkroune, “Adaptive fuzzy
backstepping tracking control for strict-feedback systems with input
[21] S. S. Ge and C. Wang, “Direct adaptive NN control of a class of delay,” IEEE Trans. Fuzzy Syst., vol. 25, no. 3, pp. 642−652, Jun. 2017.
nonlinear systems,” IEEE Trans. Neur. Netw., vol. 13, no. 1, pp. 214−
221, Jan. 2002. [40] K. P. Tee and S. S. Ge, “Control of nonlinear systems with full state
constraint using a Barrier Lyapunov function,” in Proc. 48h IEEE Conf.
[22] L. Liu, Z. S. Wang, and H. G. Zhang, “Adaptive fault-tolerant tracking Decision and Control (CDC) held jointly with 28th Chinese Control
control for MIMO discrete-time systems via reinforcement learning Conf., Shanghai, China, 2009, pp. 8618−8623.
algorithm with less learning parameters,” IEEE Trans. Autom. Sci. Eng.,
vol. 14, no. 1, pp. 299−313, Jan. 2017. [41] Y. J. Liu and S. C. Tong, “Barrier Lyapunov functions for Nussbaum gain
adaptive control of full state constrained nonlinear systems,” Automatica,
[23] H. Q. Wang, P. Shi, H. Y. Li, and Q. Zhou, “Adaptive neural tracking vol. 76, pp. 143−152, Feb. 2017.
control for a class of nonlinear systems with dynamic uncertainties,”
IEEE Trans. Cybern., vol. 47, no. 10, pp. 3075−3087, Oct. 2017. [42] K. P. Tee and S. S. Ge, “Control of nonlinear systems with partial state
constraints using a Barrier Lyapunov function,” Int. J. Control, vol. 84,
[24] Y. J. Liu, S. Li, S. C. Tong, and C. L. P. Chen, “Neural approximation- no. 12, pp. 2008−2023, Dec. 2011.
based adaptive control for a class of nonlinear nonstrict feedback
discrete-time systems,” IEEE Trans. Neur. Netw. Learn. Syst., vol. 28, [43] Y. J. Liu, S. M. Lu, S. C. Tong, X. K. Chen, C. L. P. Chen, and D.
no. 7, pp. 1531−1541, Jul. 2017. J. Li, “Adaptive control-based Barrier Lyapunov functions for a class
of stochastic nonlinear systems with full state constraints,” Automatica,
[25] Z. Liu, G. Y. Lai, Y. Zhang, X. Chen, and C. L. P. Chen, “Adaptive vol. 87, pp. 83−93, Jun. 2018.
neural control for a class of nonlinear time-varying delay systems with
unknown hysteresis,” IEEE Trans. Neur. Netw. Learn. Syst., vol. 25, no. [44] Y. J. Liu, M. Z. Gong, S. C. Tong, C. L. P. Chen, and D. J. Li, “Adaptive
12, pp. 2129−2140, Dec. 2014. fuzzy output feedback control for a class of nonlinear systems with full
state constraints,” IEEE Trans. Fuzzy Syst., doi: 10.1109/TFUZZ.2018.
[26] M. Chen and S. S. Ge, “Direct adaptive neural control for a class of 2798577.
uncertain nonaffine nonlinear systems based on disturbance observer,”
IEEE Trans. Cybern., vol. 43, no. 4, pp. 1213−1225, Aug. 2013. [45] D. J. Li, S. M. Lu, Y. J. Liu, and D. P. Li, “Adaptive fuzzy tracking
control-based Barrier functions of uncertain nonlinear MIMO systems
[27] M. Chen, P. Shi, and C. C. Lim, “Adaptive neural fault-tolerant control with full state constraints and applications to chemical process,” IEEE
of a 3-DOF model helicopter system,” IEEE Trans. Syst. Man Cybern.: Trans. Fuzzy Syst., doi: 10.1109/TFUZZ.2017.2765627.
Syst., vol. 46, no. 2, pp. 260−270, Feb. 2016.
[46] L. Bai, H. Y. Li, H. J. Liang, Q. Zhou, and L. J. Wang, “Adaptive
[28] T. S. Li, Z. F. Li, D. Wang, and C. L. P. Chen, “Output-feedback adaptive fuzzy control for nonstrict-feedback stochastic nonlinear systems with
neural control for stochastic nonlinear time-varying delay systems with full-state constraints and unknown dead zone,” in Proc. 4th Int. Conf.
unknown control directions,” IEEE Trans. Neur. Netw. Learn. Syst., vol. Information, Cybernetics and Computational Social Systems (ICCSS),
26, no. 6, pp. 1188−1201, Jun. 2015. Dalian, China, 2017, pp. 26−31.
[29] J. P. Yu, B. Chen, H. S. Yu, C. Lin, and L. Zhao, “Neural networks-
[47] D. J. Li and D. P. Li, “Adaptive controller design-based neural networks
based command filtering control of nonlinear systems with uncertain
for output constraint continuous stirred tank reactor,” Neurocomputing,
disturbance,” Inf. Sci., vol. 426, pp. 50−60, Feb. 2018.
vol. 153, pp. 159−163, Apr. 2015.
[30] Z. Wang, Y. Xu, R. Q. Lu, and H. Peng, “Finite-time state estimation
[48] B. S. Kim and S. J. Yoo, “Approximation-based adaptive control of
for coupled markovian neural networks with sensor nonlinearities,” IEEE
uncertain non-linear pure-feedback systems with full state constraints,”
Trans. Neur. Netw. Learn. Syst., vol. 28, no. 3, pp. 630−638, Mar. 2017.
IET Control Theor. Appl., vol. 8, no. 17, pp. 2070−2081, Nov. 2014.
[31] Z. Wang, R. Q. Lu, F. R. Gao, and D. R. Liu, “An indirect data-driven
[49] W. He, Y. H. Chen, and Z. Yin, “Adaptive neural network control of an
method for trajectory tracking control of a class of nonlinear discrete-
uncertain robot with full-state constraints,” IEEE Trans. Cybern., vol.
time systems,” IEEE Trans. Ind. Electron., vol. 64, no. 5, pp. 4121−
46, no. 3, pp. 620−629, Mar. 2016.
4129, May 2017.

[32] Q. L. Wei, F. L. Lewis, Q. Y. Sun, P. F. Yan, and R. Z. Song, “Discrete- [50] D. P. Li, Y. J. Liu, S. C. Tong, C. L. P. Chen, and D. J. Li, “Neural
time deterministic Q-learning: a novel convergence analysis,” IEEE networks-based adaptive control for nonlinear state constrained systems
Trans. Cybern., vol. 47, no. 5, pp. 1224−1237, May 2017. with input delay,” IEEE Trans. Cybern., doi: 10.1109/TCYB.2018.
2799683.
[33] L. Liu, Y. J. Liu, and S. C. Tong, “Neural networks-based adaptive finite-
time fault-tolerant control for a class of strict-feedback switched nonlin- [51] Y. J. Liu, S. C. Tong, C. L. P. Chen, and D. J. Li, “Adaptive NN control
ear systems,” IEEE Trans. Cybern., doi: 10.1109/TCYB.2018.2828308. using integral Barrier Lyapunov functionals for uncertain nonlinear
block-triangular constraint systems,” IEEE Trans. Cybern., vol. 47, no.
[34] A. Bemporad, “Reference governor for constrained nonlinear systems,” 11, pp. 3747−3757, Nov. 2017.
IEEE Trans. Autom. Control, vol. 43, no. 3, pp. 415−419, Mar. 1998.
[52] Q. Zhou, L. J. Wang, C. W. Wu, H. Y. Li, and H. P. Du, “Adaptive
[35] R. Q. Lu, Y. Xu, and R. D. Zhang, “A new design of model predictive fuzzy control for nonstrict-feedback systems with input saturation and
tracking control for networked control system under random packet loss output constraint,” IEEE Trans. Syst. Man Cybern.: Syst., vol. 47, no.
and uncertainties,” IEEE Trans. Ind. Electron., vol. 63, no. 11, pp. 6999− 1, pp. 1−12, Jan. 2017.
7007, Nov. 2016.
[53] H. Y. Li, L. Bai, L. J. Wang, Q. Zhou, and H. Q. Wang, “Adaptive neural
[36] M. Sampei, H. Kiyota, and M. Ishikawa, “Control strategies for mechan- control of uncertain nonstrict-feedback stochastic nonlinear systems with
ical systems with various constraints-control of non-holonomic systems,” output constraint and unknown dead zone,” IEEE Trans. Syst. Man
in Proc. 1999 IEEE Int. Conf. Systems, Man, and Cybernetics, Tokyo, Cybern.: Syst., vol. 47, no. 8, pp. 2048−2059, Aug. 2017.
Japan, pp. 158−165.
[54] Y. J. Liu, S. M. Lu, and S. C. Tong, “Neural network controller design
[37] Z. Liu, G. Y. Lai, Y. Zhang, and C. L. P. Chen, “Adaptive neural output for an uncertain robot with time-varying output constraint,” IEEE Trans.
feedback control of output-constrained nonlinear systems with unknown Syst. Man Cybern.: Syst., vol. 47, no. 8, pp. 2060−2068, Aug. 2017.
GAO et al.: ADAPTIVE NN-BASED CONTROL FOR A CLASS OF NONLINEAR PURE-FEEDBACK SYSTEMS 933

[55] K. P. Tee, B. B. Ren, and S. S. Ge, “Control of nonlinear systems with Yan-Jun Liu (M’2015−SM’2017) received the
time-varying output constraints,” Automatica, vol. 47, no. 11, pp. 2511− B.S. degree in applied mathematics and the M.S.
2516, Nov. 2011. degree in control theory and control engineering
from Shenyang University of Technology, Shenyang,
[56] B. S. Kim and S. J. Yoo, “Approximation-based adaptive tracking China, in 2001 and 2004, respectively. He received
control of nonlinear pure-feedback systems with time-varying output the Ph.D. degree in control theory and control en-
constraints,” Int. J. Control Autom. Syst., vol. 13, no. 2, pp. 257−265, gineering from Dalian University of Technology,
Apr. 2015. Dalian, China, in 2007. He is currently a Professor
with the College of Science, Liaoning University of
[57] D. P. Li, D. J. Li, Y. J. Liu, S. C. Tong, and C. L. P. Chen,
Technology. He is now an Associate Editor of IEEE
“Approximation-based adaptive neural tracking control of nonlinear
Transactions on Systems, Man, and Cybernetics:
MIMO unknown time-varying delay systems with full state constraints,”
Systems. His research interests include adaptive fuzzy control, nonlinear
IEEE Trans. Cybern., vol. 47, no. 10, pp. 3100−3109, Oct. 2017.
control, neural network control, reinforcement learning, and optimal control.
[58] D. P. Li and D. J. Li, “Adaptive neural tracking control for an uncertain
state constrained robotic manipulator with unknown time-varying de-
lays,” IEEE Trans. Syst. Man Cybern.: Syst., doi: 10.1109/TSMC.2017.
2703921.
[59] Y. J. Liu, S. M. Lu, D. J. Li, and S. C. Tong, “Adaptive controller
design-based ABLF for a class of nonlinear time-varying state constraint
systems,”IEEE Trans. Syst. Man Cybern.: Syst., vol. 47, no. 7, pp. 1546− Lei Liu received the B.S. degree in information and
1553, Jul. 2017. computing science and the M.S. degree in applied
mathematics from Liaoning University of Technol-
[60] S. M. Lu, D. P. Li, and Y. J. Liu, “Adaptive neural network control for
ogy, Jinzhou, China, in 2010 and 2013, respectively.
uncertain time-varying state constrained robotics systems,” IEEE Trans.
He received the Ph.D. degree in 2017 from North-
Syst. Man Cybern.: Syst., doi: 10.1109/TSMC.2017.2755377.
eastern University, Shenyang, China. Currently, he
[61] L. Ma and D. P. Li, “Adaptive neural networks control using Barrier is a Lecturer at Liaoning University of Technology,
Lyapunov functions for DC motor system with time-varying state Jinzhou, China. His current research interests in-
constraints,” Complexity, vol. 2018, pp. Article No. 5082401, Jan. 2018. clude fault-tolerant control, fault detection and diag-
nosis, optimal control for nonlinear systems, neural
[62] C. X. Wang, Y. Q. Wu, and J. B. Yu, “Barrier Lyapunov functions-based network control and their industrial applications.
adaptive control for nonlinear pure-feedback systems with time-varying
full state constraints,” Int. J. Control Autom. Syst., vol. 15, no. 6, pp.
2714−2722, Dec. 2017.

Tingting Gao received the B.S. degree in informa- Dapeng Li received the B.S. degree in applied
tion and computing science from Liaoning Univer- information science and engineering from Shenyang
sity of Technology, Jinzhou, China, in 2016, where Ligong University, Shenyang, China, in 2012, and
she is currently pursuing the M.S. degree in applied the M.S. degree at the School of Electrical Engineer-
mathematics. Her research interests include non- ing, Liaoning University of Technology, Jinzhou,
linear control, adaptive control, stochastic control, China, in 2015. His research interests include nonlin-
neural network control, and constraint control. ear control, adaptive fuzzy control, time-delay con-
trol, neural network control, and constraint control.

You might also like