You are on page 1of 13

278 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 6, NO.

1, JANUARY 2019

Neural Network Based Adaptive Tracking Control


for a Class of Pure Feedback Nonlinear Systems
With Input Saturation
Nassira Zerari, Mohamed Chemachema, and Najib Essounbouli

Abstract—In this paper, an adaptive neural networks (NNs) networks (NNs) or fuzzy control schemes were proposed to
tracking controller is proposed for a class of single-input/single- address the strict-feedback stochastic nonlinear systems with
output (SISO) non-affine pure-feedback non-linear systems with time-delays. In [6], adaptive neural output feedback controller
input saturation. In the proposed approach, the original input
saturated nonlinear system is augmented by a low pass filter. was presented for strict-feedback nonlinear systems with un-
Then, new system states are introduced to implement states modeled dynamics. However, the back-stepping technique was
transformation of the augmented model. The resulting new extended to deal with the more difficult control problem of
model in affine Brunovsky form permits direct and simpler larger class of uncertain pure-feedback nonlinear systems in
controller design by avoiding back-stepping technique and its which no affine appearance of the state variables was used as
complexity growing as done in existing methods in the literature.
In controller design of the proposed approach, a state observer, virtual control inputs. In [7]−[11], authors have investigated
based on the strictly positive real (SPR) theory, is introduced and the adaptive back-stepping control problem for a class of pure-
designed to estimate the new system states, and only two neural feedback nonlinear systems based on NN and fuzzy inference
networks are used to approximate the uncertain nonlinearities systems. Using the implicit function theorem, the mean value
and compensate for the saturation nonlinearity of actuator. The theorem and back-stepping technique, adaptive NN control
proposed approach can not only provide a simple and effective
way for construction of the controller in adaptive neural networks schemes were proposed in [12]−[14], for the same class of
control of non-affine systems with input saturation, but also uncertain non-linear pure-feedback systems. In [15], authors
guarantee the tracking performance and the boundedness of all proposed an adaptive fuzzy neural networks control method
the signals in the closed-loop system. The stability of the control for SISO stochastic nonlinear systems in pure-feedback form.
system is investigated by using the Lyapunov theory. Simulation A completely non-affine pure-feedback system is dealt with
examples are presented to show the effectiveness of the proposed
controller. in [16] using the input-to-state stability analysis and the small
gain theorem to develop an improved NN adaptive controller.
Index Terms—Adaptive control, input saturation, neural net- A common problem in all cited papers is the drawback in-
works systems (NNs), nonlinear pure-feedback.
herent to the application of the back-stepping design technique
known as the explosion of complexity. The latter problem is
I. I NTRODUCTION caused by the repeated differentiations of certain nonlinear

T HE control design of nonlinear systems with parameter


uncertainties using universal function approximators is
an active research area that has received increasing attention
functions, such as virtual control inputs [17]. To deal with
the complexity growing problem, a dynamic surface control
(DSC) technique was introduced in [18] for a class of strict-
along the past decades. Adaptive back-stepping becomes one feedback nonlinear systems. In [19], by using a first-order filter
of the most popular design methods in nonlinear control de- at each recursive step, neural network based adaptive DSC
sign for synthesizing controllers of lower-triangular nonlinear was further developed for a class of pure-feedback nonlinear
systems. Many adaptive back-stepping design schemes have systems with unknown time-delay functions and uncertainties.
been reported for strict-feedback nonlinear systems with un- By combining DSC technique with the back-stepping design,
known nonlinear functions [1]−[6]. In [4], [5], adaptive neural the pure-feedback nonlinear systems with unknown dead zone
and uncertainties were considered in [20] with the restriction
Manuscript received March 2, 2018; revised April 24, 2018; accepted June that the (n − 1)th and nth state equations were assumed
7, 2018. Recommended by Associate Editor Yanjun Liu. (Corresponding
author: Nassira Zerari.) to be affine such that the repeated differentiations of the
Citation: N. Zerari, M. Chemachema, and N. Essounbouli, “Neural network virtual control inputs can be eliminated. However, from a
based adaptive tracking control for a class of pure feedback nonlinear systems point of view of practical applications, the methods developed
with input saturation,” IEEE/CAA J. Autom. Sinica, vol. 6, no. 1, pp. 278−290,
Jan. 2019. in [20] are computationally expensive because the NN and
N. Zerari and M. Chemachema are with the Department of Electronics, fuzzy approximators were used at every design step to online
Faculty of Technology, University of Constantine1, Compus A. Hamani, Route approximate the unknown dynamics. Furthermore, the adaptive
Aine El Bey, 2500 Constantine, Algeria (e-mail: zer.napg2010@gmail.com;
m chemachema@yahoo.fr). laws involved in DSC require a large number of parameters
N. Essounbouli is with CReSTIC Laboratory, University of Reims Cham- to updated online, which make the control law and stability
pagne Ardennes, France (e-mail: najib.essounbouli@univ-reims.fr). analysis very complicated. Only few results were reported on
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. the adaptive neural control for general nonlinear systems in
Digital Object Identifier 10.1109/JAS.2018.7511255 pure-feedback or strict-feedback form without using back-
ZERARI et al.: NN BASED ADAPTIVE TRACKING CONTROL FOR A CLASS OF PURE FEEDBACK NONLINEAR SYSTEMS WITH INPUT SATURATION 279

steeping and DSC methodology [21]−[23]. Despite these the tracking errors converge to a small neighbourhood of the
efforts, the existing results [21]−[23] are focusing on uncertain origin. Compared to the literature, the main contributions of
nonlinear systems without considering the input nonlinearities. this paper are summarized as follows.
In real world, input nonlinearities exist widely in physical 1) This paper solves the adaptive tracking control problem
systems such as mechanical, hydraulic, magnetic, and other of nonlinear systems with input saturation in lower trian-
types of system components, so dealing with these nonlinear- gular structure like pure-feedback and strict-feedback forms
ities in controller design is an important research field. Dead- along with avoiding the well-known explosion of complexity
zone in [24]−[26], quantized inputs in [27], [28], backlash problem of backstepping. In [35]−[42], [46], authors treated
and hysteresis in [29]−[31] and saturation nonlinearities in the same control problem, however, their proposed controllers
[32]−[37] are common non-smooth nonlinear characteristics suffer from the explosion of complexity.
dealt with in the literature. Input saturation is one of the most 2) Compared to the works in [37]−[46] dealing with input
important non-smooth nonlinearities present in many practical saturated strict feedback systems, the considered nonlinear
systems, which can hard limit its performance. Therefore, the systems in this paper are in input saturated pure-feedback
effect of input saturation should be taken into consideration form. Therefore, the control design and stability analysis of
in the design and analysis of control systems. If the input the closed-loop system in this paper are more difficult and
saturation is ignored in the control design, the closed-loop cover a larger class of nonlinear systems.
control performance will be severely degraded or even the 3) The number of the online adaptive learning parameters
system may become instable. Many researchers have devoted is highly reduced since the proposed methods only needs two
a lot of effort to cope with the adaptive control problem for NNs whatever the order of the nonlinear system, while the
uncertain nonlinear systems with input saturation [38]−[46]. methods in [29]−[37], [48]−[50] require a number of NN
In [38], output tracking control schemes based on fuzzy back- units as many as the order of the system.
stepping dynamic surface control was presented for strict- This paper is organized as follows. In Section II, the
feedback nonlinear systems with input saturation. In [47], problem formulation is presented. In Section III, adaptive
back-steeping-based fuzzy adaptive control is developed for neural tracking control and observer design are developed by
unknown nonlinear non-strict-feedback systems where an aux- using NNs and the closed-loop system stability is analysed.
iliary system has been introduced to compensate for the effects Simulation results are presented to show effectiveness of the
of input saturation. This result was extended to deal with pure- approach in Section IV, and Section V contains the conclu-
feedback nonlinear systems with input saturation [48]−[51]. sions.
Specifically, in [49], [50], the class of saturated stochastic
nonlinear pure-feedback systems was considered. However, in II. P ROBLEM F ORMULATION
all previous works [38]−[51], the proposed controllers require
N neural networks to online learn the uncertain dynamics for A. Problem Formulation
N th order nonlinear systems. To our best knowledge, there Consider the following class of pure-feedback nonlinear
are no results on neural networks adaptive tracking control for systems:
input saturated uncertain nonlinear pure-feedback systems.
Motivated by the above observation, in this paper an adap- ẋi = fi (x̄i , xi+1 ), i = 1, . . . , n − 1
tive neural network tracking control approach is proposed for ẋn = fn (x̄n , u(v)), n≥2
the class of completely non-affine pure-feedback nonlinear y = x1 (1)
systems in the presence of input saturation. In the proposed
approach, the back-stepping technique is avoided and thus the where x̄i = [x1 , . . . , xi ]T ∈ R, i = 1, . . . , n are measurable
well-known problem of explosion of complexity is completely state vector, y ∈ R is the output of the system, and fi (·, ·) are
eliminated. Using a function transformation, the original in- unknown smooth functions. Moreover, v denotes the actual
put saturated non-affine pure-feedback system is transformed control input and u(v) is the input saturation type of the
into the strict-feedback form. Then, the obtained model is nonlinearity.
transformed again into the affine triangular Brunovsky form According to [37], input saturation u(v) can be described
by the introduction of new states definitions along with a by
low pass filter for the actual saturated control input. In the (
sign(v(t))um , |v(t)| > um
controller design of the proposed approach, a state observer, u(v) = sat(v) = (2)
based on strictly positive real (SPR) theory, is introduced v(t), |v(t)| ≤ um
and designed to estimate the new system states, and Neural
where um is a known bound on u(t). The relationship between
Networks are used to approximate the uncertain nonlinearities
the applied control u(t) and the control input v(t) has a sharp
and compensate for the saturation nonlinearity of actuator. To
corner when |v(t)| = um . This saturation description cannot
deal with the approximation errors, a robustifying control term
be applied directly. According to [37], the saturation can be
is employed in addition to the σ-modification algorithm in the
approximated by the following smooth function and is shown
adaptation laws [52]. The stability of closed-loop system is
in Fig. 1.
investigated by using the Lyapunov theory and it is shown
that the proposed adaptive control approach can guarantee that v
g(v) = um tanh( )
all the signals in the closed-loop system are bounded, and um
280 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 6, NO. 1, JANUARY 2019

exp( uvm ) − exp −( uvm ) R+ → R will be approximated by neural networks (NNs) with
= um . (3)
exp( uvm ) + exp −( uvm ) the output given by
fˆ(Z/Ŵ ) = Ŵ T S(Z) (6)
where Z ∈ ΩZ ∈ Rm is the input to NN, Ŵ ∈ RN is
called the weight vector, and N is the number of network
T
nodes, S(Z) = [s1 (Z), . . . , sN (Z)] is the so-called vector of
activation function chosen as the hyperbolic tangent function
in our study
1 − e−Z
si (Z) = , i = 1, . . . , N. (7)
1 + e−Z
According to the approximation property of the NN, any
continuous function over a compact set Ω, can be approxi-
Fig. 1. Saturation um tanh(v/um ) (dash-dotted) and sat(v(t)) (solid line). mated as follows:

Then, sat(v(t)) in (2) can be expressed as f (Z) = W T S(Z) + ε(Z) (8)


where W are the optimal weights and ε(Z) is the smallest
u(v) = g(v) + d(v)
approximation error, while N is sufficiently large so as we
v
= um tanh( ) (4) have that for any ε̄ > 0, the approximation error is bounded
um by ε̄ > 0, i.e., |ε(Z)| ≤ ε̄. From the universal approximation
where d(v) =sat(v(t)) − g(v) is bounded as proved in [37]. results for neural networks stated in [55], it is known that the
The control objective is to design adaptive control v(t) for constant ε̄ can be made arbitrarily small by increasing the NN
a class of pure-feedback nonlinear systems (1) with input nodes number N .
saturation such that all the signals of the closed-loop system
are bounded and the system output y(t) of system (1) follows III. A DAPTIVE N EURAL N ETWORK C ONTROL AND
any given bounded desired output signal yd (t). O BSERVER D ESIGN
For designing the adaptive control for pure-feedback sys- A. Controller Design
tems, system (1) is rewritten as In this section, an adaptive neural network control scheme
will be presented. In order to improve tracking performance,
ẋi = Fi (x̄i , xi+1 ) + xi+1 , i = 1, . . . , n − 1
we develop a new scheme to handle the difficulties introduced
ẋn = Fn (x̄n , u(v)) + u(v), n≥2 by the cascading for input saturated strict-feedback system. By
y = x1 (5) incorporating the strict-feedback system (5) with a low-pass
filter, it is transformed into an affine strict-feedback system
where Fi (x̄i , xi+1 ) = fi (x̄i , xi+1 )−xi+1 and Fn (x̄n , u(v)) =
in normal form but with states unavailable for feedback, the
fn (x̄n , u(v))−u(v). System (5) is used to design the observer
system (5) becomes
and the controller in the next section.
For the development of control laws, the following assump- ẋi = Fi (x̄i , xi+1 ) + xi+1 , i = 1, . . . , n − 1
tions are made. ẋn = Fn (x̄n , u(v)) + g(v) + d(v), n≥2 (9)
Assumption 1 [53]: fi (x̄i , xi+1 ), i = 1, . . . , n, are the v̇ = −cv + ω (10)
unknown but smooth functions, i.e., they are differentiable,
and their partial derivatives ∂Fi∂x (x̄i ,xi+1 )
and ∂Fn (x∂v
n ,u(v))
are where c ∈ R+ is a positive constant and ω is an auxiliary
i+1
continuous. control signal to be designed to generate a stable control signal
Assumption 2 [21]: The desired trajectory yd (t) is contin- v, in the augmented system given by (10). For the augmented
uously differentiable up to (n + 1)th order and the vector Yd system, it is considered that the control input v is one of the
(n) (n+1)
= [yd , . . . , yd ] and yd are available where Yd ∈ Ωd ⊂ system states, and ω is a new control input.
R n+1
with Ωd being a compact set. Using the transformations reported in [22], we show that
Remark 1: When there is no input nonlinearity, the system the system (5) can be transformed into the normal form, by
given by (1) can be viewed as a simplified version of pure- selecting new state variables as
feedback nonlinear system in [21]. However, the focus of this (i−1)
si = x1 , i = 1, 2, . . . , n + 1.
paper is on control design for the same class of nonlinear
systems with input saturation. Therefore, the controller design Step 1: Define new state s1 = y = x1 = ϕ1 (x1 ).
Step 2: Let s2 = ṡ1 = F1 (x1 , x2 ) + x2 = ϕ2 (x̄2 ) and the
and the stability of the closed-loop system are more difficult
time derivative of s2 can be expressed as
than those in [21].
ṡ2 = ϕ̇2 (x̄2 )
B. Neural Networks Approximation ∂ϕ2 (x̄2 ) ∂ϕ2 (x̄2 )
= ẋ1 + ẋ2
Neural networks have been frequently used as function ∂x1 ∂x2
approximators [54], in this study, a continuous function f (Z) : = ϕ3 (x̄3 ). (11)
ZERARI et al.: NN BASED ADAPTIVE TRACKING CONTROL FOR A CLASS OF PURE FEEDBACK NONLINEAR SYSTEMS WITH INPUT SATURATION 281

Step i (i = 3, . . . , n): By a similar way, let si = ṡi−1 = e1 = y − yd = s1 − yd . (19)


ϕi (x̄i ) and the time derivative of si is obtained as
Then, the error dynamics can be given by
ṡi = ϕ̇i (x̄i ) ėn+1 = F (x̄n , u(v)) + g(x̄n , u(v))ω
i
X ∂ϕi (x̄i ) + β(Zβ )(g(v) + d(v)) (20)
= ẋj
∂xj
j=1 which can be written in matrix form as
h
= ϕi+1 (x̄i+1 ). (12) ė = Ae + B F (x̄n , u(v)) + g(x̄n , u(v))ω
Step n + 1: This is the final step. i
(n+1)
+ β(Zβ )(g(v) + d(v)) − yd
Let sn+1 = ṡn = ϕn+1 (x̄n , u(v)), and taking its time
derivative yields y = CT e (21)
ṡn+1 = ϕ̇n+1 (x̄n , u(v)) with
n
 
X ∂ϕn+1 (x̄n , u(v)) ∂ϕn+1 (x̄n , u(v)) 0 1 0 0
= ẋj + v̇  ···

∂xj ∂v  0 0 1  0
j=1  .. ..  ..
n−1
A= . .  ∈ R(n+1)×(n+1)
.
X ∂ϕn+1 (x̄n , u(v)) ∂ϕn+1 (x̄n , u(v))  
= ẋj + v̇  0 0 0 1 
∂xj ∂v ···
j=1 0 0 0 0
∂ϕn+1 (x̄n , u(v))    
+ (Fn (x̄n , u(v)) + g(v) + d(v)) 0 0
∂xn  0   0 
n−1
X ∂ϕn+1 (x¯n , u(v))    
∂ϕn+1 (x̄n , u(v))    
= ẋj + v̇ B =  ... ∈R (n+1)×1
, C =  ...  ∈ R(n+1)×1 .
∂xj ∂v    
j=1  0   0 
∂ϕn+1 (x̄n , u(v)) 1 1
+ Fn (x̄n , u(v))
∂xn Note that the pair (A, B) is controllable and the pair (C T ,
∂ϕn+1 (x̄n , u(v)) A) is observable.
+ (g(v) + d(v)). (13)
∂xn Remark 2: From the above transformation, we see that the
By substituting (10) into (13), an affine term with respect system (5) is similar to the affine systems in [42], [46]. In
to w appears as [42], adaptive DSC approaches were developed. In [46], the
traditional adaptive back-stepping design method is applied.
ṡn+1 = F (x̄n , u(v)) + g(x̄n , u(v))ω + β(Zβ )(g(v) + d(v)) Though the above design procedure contains N steps, the
(14) virtual control laws at intermediate steps are not necessary to
where be implemented. However, in our study the designed controller
n−1
X only contains the actual control law with its adaptive laws
∂ϕn+1 (x̄n , u(v)) ∂ϕn+1 (x̄n , u(v)) which are given directly. Thus, the structure of the designed
F (x̄n , u(v)) = ẋj −c v
j=1
∂xj ∂v controller is much simpler than that of existing design ap-
∂ϕn+1 (x̄n , u(v)) proaches, and the controller realization is rather easier.
+ Fn (x̄n , u(v)) (15) Remark 3: The new states si , i = 2, . . . , n + 1 of the
∂xn
∂ϕn+1 (x̄n , u(v)) transformed system (18) are not available for the controller
g(x̄n , u(v)) = (16) design, thus an observer is introduced to estimate the tracking
∂v
∂ϕn+1 (x̄n , u(v)) error vector. From the estimate of the tracking error vector,
β(Zβ ) = (17) we can directly determine the estimate of the new state vector
∂xn
s = [s1 , . . . , s(n+1) ]T .
with Zβ = [x̄n , u(v)] .
As a result, the system (5) can be described by the following B. Observer Design
normal form with respect to the newly defined state variables
Consider now the following observer for estimating the
ṡi = si+1 , 1 ≤ i ≤ n tracking error vector e:
ê˙ = (A − Bk T )ê − Byd
(n+1)
ṡn+1 = F (x̄n , u(v)) + g(x̄n , u(v))ω + β(Zβ )(g(v) + d(v)) + Lẽ1
y = x1 (18) ê1 = C T ê. (22)
T T
where s = [s1 , . . . sn+1 ] are the new state variables, and y ∈ Let k = [k0 , k1 , . . . , kn ] be a feedback gain vector, and L
R is the measured output and F (x̄n , u(v)) and g(x̄n , u(v)) = [L1 , . . . , L(n+1) ] be an observer gain vector and they are
are unknown nonlinear functions. Moreover, v denotes the chosen such that the characteristic polynomials of A − Bk T
actual control input and u(v) is the input saturation type of and A−LC T are Hurwitz. Note that ẽ = e−ê, ê = ŝ−Yd with
the nonlinearity. Let us define the tracking error as e = [e1 , ŝ being estimate of the state vectors and ê being the estimate
(n)
e2 , . . . , en+1 ] = [e1 , ė1 , . . . , e1 ], where of the tracking error vector e. Consequently, if the observer
282 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 6, NO. 1, JANUARY 2019

error ẽ converges to zero then the estimated tracking error ê to generate the adaptive laws to estimate the parameter vectors
will also converge to zero. and unknown constants.
To facilitate the controller design, and according to the Due to the canonical form of matrices A, B and C, the
universal approximation property of neural networks [55], the observation error dynamics (30) can be expressed in frequency
following unknown nonlinear functions can be approximated domain using mixed notation (i.e., time-frequency).
as follows:
h
g −1 (x̄n , u(v))(F (x̄n , u(v)) + β (Zβ ) (g(v) + d(v))) ẽ = H(p) g(x̄n , u(v))(W̃1T S(Z) + ε1 (Z) + vr
i
= W1T S(Z) + ε1 (Z) (23)
+ k T ê(W̃2T S(Z) + ε2 (Z)) (31)
g −1 (x̄n , u(v)) = W2T S(Z) + ε2 (Z) (24)
where S(·) is NN vectors fixed a priori by the designer, ε1 (Z) where the p is the Laplace variable and H(p) = C T (P I − (A
and ε2 (Z) are the NN approximation errors, and the input − LC T )B is the stable transfer function of (31). Note that
vector Z = [x̄n , v, u(v)]T . W1 and W2 are ideal parameter this notation is very common in the adaptive control literature
vectors, which are introduced only for analysis purpose, and such as in [57]−[60].
their values are not needed when implementing the controller. Now, since H(p) in (31) is not SPR in general, we introduce
According to the discussion in [56], W is defined as follows: a low pass filter L(p) such that H(p)L−1 (p) = H(p) is SPR.
n ¯ ¯o
¯ ¯ h
Wi = arg min supz∈Ωz ¯fˆ (Z/W ) − f (Z)¯ . (25)
W ẽ1 = H(p)L−1 (p) L(p)(W̃1T S(Z) + ε1 (Z) + vr
Since the ideal parameter vectors W are unknown, it should i
+ k T ê(W̃2T S(Z) + ε2 (Z))
be estimated by a suitable adaptation laws. The estimates are
h
represented by Ŵ . Let us define the parameter error vectors
= H(p)L−1 (p) W̃1T S(Z) + k T êW̃2T S(Z)
as i
W̃i = Wi − Ŵi , i = 1, 2. (26) + U + vr (32)

According to (23), the dynamic error (21) can be repre- where


sented as
h i h
(n+1)
ė = Ae + B g(x̄n , u(v))(W1T S(Z) + ε1 (Z) + ω) − yd U = L(p) g(x̄n , u(v))(W̃1T S(Z) + ε1 (Z) + vr
i
y = C T e. (27) + k T ê(W̃2T S(Z) + ε2 (Z))
Let us choose the following controller: − W̃1T S(Z) − k T êW̃2T S(Z) + vr . (33)
ω = k T êŴ2T S(Z) − Ŵ1T S(Z) + vr (28)
Now, to facilitate the controller design and the stability
where vr is a compensative control term for the approximation analysis, we make the following mild assumption:
errors to be specified later.
Assumption 3: Each element τ̄ of the unknown parameter
Substituting (28) into (27) results in the following error defined in (33) is bounded by a constant, i.e., there exists
system dynamics unknown positive constant τ̄ such that
h
ė = Ae − Bk T ê + B g(x̄n , u(v))(W̃1T S(Z) + ε1 (Z))
|U| ≤ τ̄ . (34)
+ k T êg(x̄n , u(v))(W̃2T S(Z) + ε2 (Z))
i
(n+1)
+ g(x̄n , u(v))vr − yd . (29) Thus, the state space realization of (30) leads to
Let us define the observation error vector as
ẽ˙ = (A − LC T )e
T
ẽ = [ẽ1 , . . . , ẽn ] = e − ê. h i
+ Bc (W̃1T S(Z) + k T ê(W̃2T S(Z) + U + vr
Subtracting (22) from (29), we get the dynamics of the
observation error as ẽ1 = C T ẽ (35)
h
ẽ˙ = (A − LC T )e + Bg(x̄n , u(v)) (W̃1T S(Z) + ε1 (Z)) with (A − LC T ) being stable. Thus, for any given positive
i
definite symmetric matrix, there exists a unique positive defi-
+ k T ê(W̃2T S(Z) + ε2 (Z)) + vr
nite symmetric solution P to the following Lyapunov algebraic
ẽ1 = C T ẽ. (30) equation.
Since only the output observation error ẽ1 in (30) is
measurable, we use the strictly positive real (SPR) Lyapunov (A − LC T )T P + P (A − LC T ) = −Q (36)
approach to analyse the stability of the closed-loop system and P Bc = C. (37)
ZERARI et al.: NN BASED ADAPTIVE TRACKING CONTROL FOR A CLASS OF PURE FEEDBACK NONLINEAR SYSTEMS WITH INPUT SATURATION 283

C. Stability Analysis By completion of squares, we have


According to the above discussions, the following theorem 1 1
σ1 W̃1T Ŵ1 ≤ kW k21 − kW̃ k21 (47)
is given to explain the control performance of the closed-loop σ1 σ1
system. 1 1
Theorem 1: Consider the nonlinear system (1) that satisfies σ2 W̃2T Ŵ2 ≤ kW k22 − kW̃ k22 (48)
σ2 σ2
Assumptions 1−3. Suppose that the tracking error observer is 1 2 1
designed as given in (22), and the control law is chosen as −σ3 τ̃ τ̂ ≤ τ̃ − τ̄ 2 . (49)
σ3 σ2
given in (28) with the following compensative control term:
Then, V̇ can be rewitten as follows:
ẽ1
vr = −τ̃ tanh( ) (38) 1 1 1
² V̇ ≤ −ẽT Qẽ − kW̃ k21 − kW̃ k22 − τ̄ 2
σ1 σ2 σ2
with the corresponding adaptive control laws 1 1 1 2
2 2
+ kW k1 + kW k2 + τ̄ + κ²τ̄
˙ σ1 σ2 σ2
Ŵ 1 = Γ1 (ẽ1 S(Z) − σ1 Ŵ1 ) (39)
˙ T ≤ −αV + θ (50)
Ŵ 2 = Γ2 (k êẽ1 S(Z) − σ2 Ŵ2 ) (40)
σ1 2 σ2 2 σ3 2
ẽ1 where θ = 2 kW1 k + 2 kW2 k + 2 τ + κ²τ .
τ̂˙ = γ1 (ẽ1 tanh( ) − σ2 τ̂ ) (41) Let
² ( )
where Γ1 , Γ2 , γ1 , σ1 , σ2 and σ3 are positive constants. Then, λmin (Q) σ1 σ2 σ3
α = min , ¡ ¢, ,
it can be guaranteed that all the signals of the closed-loop λmax (P ) λmax Γ1 −1 λmax (Γ2 −1 ) γ1
system are bounded, and the tracking errors converge to a ¡ ¢
small neighborhood of origin. where λmax Γ1 −1 and λmax (Γ2 −1 ) denote the maximum
Remark 4: The adaptive laws (39)−(41) are given in a eigenvalues of Γ1 −1 and Γ2 −1 , respectively.
σ-modification form to help in establishing the differential Multiplying both sides by eαt , (50) can be expressed as
inequality in terms of the Lyapunov function V for stability d
(V (t) eαt ) ≤ θeαt . (51)
proof. dt
Proof: Define the following Lyapunov function candidate: Integrating (51) over [0, t]
· ¸
1 1 1 θ θ −αt
V = ẽT P ẽ + W̃1T Γ1 −1 W̃1 + W̃2T Γ2 −1 W̃2 + γ1 τe2 0 ≤ V (t) ≤ + V (0) − e . (52)
2 2 2 α α
(42)
From (52), therefore, ẽ, W̃1 , W̃2 , τ̃ are bounded.
where τ̃ = τ̄ −τ̂ and W̃i = Ŵi −Wi , i = 1, 2 are the parameter Using (52), V (0) is defined as follows:
estimation errors. The time derivative of (42) is given by 1
h V (0) = ẽT (0) P ẽ (0) + W̃1T (0) Γ1 −1 W̃1 (0)
2
V̇ = − ẽT Qẽ + ẽT P Bc (W̃1T S(Z) + k T ê(W̃2T S(Z) 1 T 1 2
−1
i + W̃2 (0)Γ2 W̃2 (0) + τe (0). (53)
˙ ˙ 1 2 2γ1
+ U + vr − W̃1T Γ1 −1 Ŵ1 − W̃2T Γ2 −1 Ŵ2 − τe τḃ.
γ1 From (42) and (53), one has
(43) µ µ µ ¶ ¶¶ 12
1 θ θ
Considering (37), (43) can be written as kẽk ≤ + V (0) − e−αt . (54)
λmin (P ) α α
h i
˙ Then the solution ẽ exponentially converges to a bounded
V̇ = − ẽT Qẽ + [ẽ1 U + ẽ1 vr ] + W̃1T ẽ1 S(Z) − Γ1 −1 Ŵ1
h i region
˙ 1 ( ¯
+ W̃2T ẽ1 k T êS(Z) − Γ2 −1 Ŵ2 − τe τḃ. (44) ¯ µ ¶1 )
γ1 ¯ 1 θ 2
Ωẽ = ẽ ¯ kẽk ≤ . (55)
By using Assumption 3, substituting (39) into (44) results ¯ λmin (P ) α
in Similarly, the parameter error vector W̃1 and W̃2 are also
· ¸
T ẽ1 1 bounded and converge to ΩW̃1 and ΩW̃2 respectively, which
V̇ ≤ −ẽ Qẽ + κ² τ̄ + ẽ1 τ̃ tanh( ) + τe τḃ
² γ1 are defined as
h i  ¯ ! 12 
T −1 ˙  ¯ ° ° Ã 
+ W̃1 ẽ1 S(Z) − Γ1 Ŵ1 ¯ ° ° 2 θ
h i ΩW̃1 = W̃1 ¯¯ °W̃1 ° ≤ ¡ −1 ¢ (56)
˙  ¯ λmin Γ1 α 
+ W̃2T ẽ1 k T êS(Z) − Γ2 −1 Ŵ2 . (45)
 ¯ ! 12 
 ¯ ° ° Ã 
Noting that for any ² > 0, the following inequality, as in ¯ ° ° 2 θ
ΩW̃2 = W̃2 ¯¯ °W̃2 ° ≤ ¡ −1 ¢ . (57)
[61], holds −x tanh( x² ) + |x| ≤ κ², with κ = 0.2785, and  ¯ λmin Γ2 α 
according to the parameter adaptation laws (39)−(41), one
has From the boundedness of ẽ, W̃1 , W̃2 and τ̃ one can directly
conclude about the boundedness of e, ê, s̃, Ŵ1 , Ŵ2 and τ̂ since
V̇ ≤ −ẽT Qẽ + σ1 W̃1T Ŵ1 + σ2 W̃2T Ŵ2 − σ3 τ̃ τ̂ + κ²τ̄ . (46) ẽ = e − ê = s̃, Ŵ1 = W1 − W̃1 , Ŵ2 = W2 − W̃2 , τ̂ = τ̄ − τ̄
284 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 6, NO. 1, JANUARY 2019

T T
and the ideal parameter W1 , W2 and τ̄ must be bounded. ê (0) = [ê1 (0) , ê2 (0) , ê3 (0) , ê4 (0)] = [1, 0, 0, 0] , and
According to the boundedness of ẽ and the definition of ê in the initial values of the estimate parameters W1 (0) and W2 (0)
(22), it can be obtained that ê is bounded and thus e is bounded are randomly selected.
too. Based on e = s−Yd we have that s is also bounded and it Therefore, it is conclude that all the signals in the resulting
thus follows that ŝ is bounded since s̃ is bounded. Additionally, closed-loop system are bounded. The design parameters used
From (10), the control input v can be viewed as output of a in this simulation are chosen as follows: Γ1 = Γ2 = 0.001,
stable linear time invariant system driven by signal ω, which is σ1 = σ2 = 0.001, γ1 = 10, σ3 = 10, ² = 0.1, a = 20.
bounded. The state x1 is bounded since x1 = s1 . Furthermore, Fig. 2 illustrates the simulation results of Example 1.
∂b2
since s2 is bounded and ∂x 2
> 0, we conclude that x2 is Fig. 2 (a) shows the trajectories of the system output along with
bounded based on the definition of s2 = b2 (x2 ). Similarly the reference signal, it can be seen that fairly good tracking
and recursively, the states xi , i = 3, . . . , n are all bounded.¥ performance is obtained. Fig. 2 (b) shows the trajectories of
states xi . To verify the validity of the designed observer, the
IV. S IMULATION R ESULTS observer errors ẽ are described in Fig. 2 (c). It can be seen
In order to show the effectiveness of the proposed method, that the designed observer is very good for estimating the
two examples are given in this section. The first one is a unmeasured states. The norm of NN approximation parameter
nonlinear pure-feedback system in the from of (1), while, the is presented in Fig. 2 (d). The boundedness of the intermediate
second one is directly applied to a nonlinear strict-feedback control law ω(t) is illustrated in Fig. 2 (e). The control signals
system in the form of (5). u(t) and sat(u(t)) can be observed in Fig. 2 (f). Fig. 3 and
Fig. 4 show the tracking-errors and their estimates and the tra-
A. Example 1 jectories of new states along with their estimates, respectively.
The following third-order non-affine nonlinear system given To further show the effectiveness of the proposed controller
by [21]: in this paper, the comparison is conducted with traditional
back-stepping control without considering the input saturation
ẋ1 = f1 (x1 , x2 ) [16]. The actual control input is given as [16]
ẋ2 = f2 (x̄2 , x3 )
u = −z2 − k3 z3 − Ŵ3 S (Z3 ) (59)
ẋ3 = f3 (x̄3 , u(v))
y = x1 (58) where z1 = x1 − yd , z2 = x2 − α1 , z3 = x3 − α2 .
The virtual controllers are chosen as [16]
where
α1 = −k1 z1 − Ŵ1T S(Z1 )
f1 (x1 , x2 ) = 0.2x1 + x2
α2 = −z1 − k2 z2 − Ŵ2T S(Z2 )
1 − exp(x1 x2 )
f2 (x̄2 , x3 ) = + x3 + 0.05 sin(x3 ) T
1 + exp(x1 x2 ) where the inputs of NNs are chosen as Z1 = [x1 , ẏd ] ,
¡ ¢ Z2 = [x2 , ∂α
T ∂α2 ∂α2 T
∂x1 , ∅1 ] , ³Z3 =´ [x3 , ∂x1 , ∂x2 , ∅2 ] with ∅i =
1
f3 (x̄3 , u(v)) = 0.2 exp −x42 x63 ³ ´
¡ ¢ Pi ∂αi k+1 ∂αi ˙
+ 0.9 + 0.05 exp(−x21 ) sat(u) + 0.1 sin(sat(u)) k=0 ∂ydk yd + ∂ Ŵ Ŵ k , i = 1, 2 and the update
k
laws are given by
are unknown nonlinear functions. The input saturation ³ ´
sat(u(v)) is determined by (2) with the parameter um = 2.5. ˙
Ŵ i = Γi z1 S (Zi ) − σi Ŵi , i = 1, 2, 3. (60)
The simulation objective is to apply the proposed adaptive NN
controller such that the system output y1 tracks the desired The obtained results when applying the controller defined
trajectory yd = sin(t). To implement the proposed controller, in [16] without input saturation are shown in Fig. 5. It is very
the observer-gain vectors are selected respectively as follows: clear that our approach outperforms the approach proposed in
k = [15.596 × 102 , 149.78 × 102 , 463.1, 50.1], L = [102, [16] since the obtained response in Fig. 5 (a) shows equivalent
62.03 × 102 , 17.25 × 103 , 29.153 × 103 ]T . Selecting the tracking performance to our approach but with the cost of a
positive definite matrix Q = diag{10, 10, 10, 10} and solving large control signal as shown in Fig. 5 (b). This large control
the Lyapunov algebraic equation (36) to obtain the following signal may be nonrealistic. However, when the input saturation
positive definite symmetric matrix: is applied in Fig. 5 (c). Fig. 5 (d) shows the inability of the
  controller in [16] to derive the system to follow its reference.
25.06 15.06 −25.03 −15.03
 15.06 55.16 5.06 −75.08 
P =  −25.03
.
5.06 55.11 −35.03  B. Example 2
−15.03 −75.08 −35.03 124.86 We consider a second-order non-affine nonlinear system
Two NN systems in the form of (6) are used to generate given by [46]
the unknown nonlinear systems with Z = [x1 , x2 , x3 , v(t), x2 − x1
u(v)]T and inputs N = 5 network nodes and S(·) hyperbolic ẋ1 =
1 + x41
tangent function as activation function. The initial conditions 2

are chosen as ẋ2 = sat(u(t)) − x2 e−x1


T T
x (0) = [x1 (0) , x2 (0) , x3 (0) ] = [1, 0, 0] y = x1 (61)
ZERARI et al.: NN BASED ADAPTIVE TRACKING CONTROL FOR A CLASS OF PURE FEEDBACK NONLINEAR SYSTEMS WITH INPUT SATURATION 285

Fig. 2. Simulation results of Example 1. (a) Tracking performance: output y and the desired trajectory yd . (b) Trajectory of states x2 and x3 . (c) Observer
estimation error ẽ. (d) Norms of adjustable weights of NNs systems. (e) Control input intermediate ω(t). (f) Trajectory of u(t) and sat (u(t)).

Fig. 3. Tracking error e and its estimates ê for Example 1.

where the input saturation u(v(t)) is determined by (2) with and solving the Lyapunov algebraic equation (36) to obtain
the parameter um = 2. The simulation objective is to apply the the following positive definite symmetric matrix:
proposed adaptive NN controller such that the system output  
5.1686 × 103 −50.00 −43.40
y1 tracks the desired trajectory yd = sin(t). To implement
P = −50.00 43.40 −50.00  .
the proposed controller, the observer-gain vectors are selected
−43.40 −50.00 60.90
respectively as follows:
T
k = [15.813 × 102 , 407.8 350], L = [100, 6000, 5000] . Two NN systems in the form of (6) are used to generate the
Specify a positive definite matrix Q = diag{100, 100, 100} unknown nonlinear systems with Z = [x1 , x2 , v(t), u(v)]T
286 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 6, NO. 1, JANUARY 2019

Fig. 4. New states variables s along with their estimates ŝ for Example 1.

Fig. 5. Sumilation results of Example 1. (a) Performance comparison of different controllers without input saturation. (b) Control input u(t) comparison of
different controllers (without input saturation). (c) Control input u(t) of back-stepping method with input saturation. (d) Tracking performance of back-stepping
controller with input saturation.

and inputs N = 4 network nodes and S(·) hyperbolic tangent control law ω(t) is illustrated in Fig. 6 (e). The control signals
function as activation function. The initial conditions are u(t) and sat(u(t)) can be observed in Fig. 6 (f). Fig. 7 and
T
chosen as x(0) = [x1 (0), x2 (0)]T = [0.01, 0.2] , ê(0) = Fig. 8 show the tracking-errors and their estimates and the tra-
T T
[ê1 (0), ê2 (0), ê3 (0)] = [0, 0, 0] and the initial values of the jectories of new states along with their estimates, respectively.
estimate parameters W1 (1) and W2 (0) are randomly selected. To further show the effectiveness of the proposed controller
The design parameters used in this simulation are chosen in this paper, the comparison is conducted with traditional
as follows: Γ1 = 100, Γ2 = 0.001, σ1 = 0.01, σ2 = 100, γ1 back-stepping control with considering the input saturation
= 0.01, σ3 = 0.01, ² = 0.01, a = 20. [46]. The actual control input u(t), the intermediate control
Fig. 6 illustrates the simulation results of Example 2. function α1 , and adaptive laws Ŵ1 and Ŵ1 are given by [46]
µ ¶2
Fig. 6 (a) shows the trajectories of the system output along with 1 ∂α1
the reference signal, it can be seen that fairly good tracking u = −h − z1 − k2 z2 − z2 − 2 z2 − H2 (62)
2 ∂y
performance is obtained. Fig. 6 (b) show the trajectories of 7z1
states xi . To verify the validity of the designed observer, the α1 = −k1 z1 − Ŵ1T S1 (x1 ) − ẏd − 2 −z ) (63)
4 (kb1 1
observer errors ẽ are described in Fig. 6 (c). It can be seen z1
˙
that the designed observer is very good for estimating the Ŵ 1 = Γ1 S1 (x1 ) 2 − σ Ŵ1 (64)
kb1 − z1
unmeasured states. The norm of NN approximation parameter
˙
is presented in Fig. 6 (d). The boundedness of the intermediate Ŵ 2 = Γ2 z2 S2 (x̂2 ) − σ Ŵ2 (65)
ZERARI et al.: NN BASED ADAPTIVE TRACKING CONTROL FOR A CLASS OF PURE FEEDBACK NONLINEAR SYSTEMS WITH INPUT SATURATION 287

Fig. 6. Simulation results of Example 2. (a) Tracking performance: output y and the desired trajectory yd . (b) Trajectory of states x2 . (c) Observer estimation
error ẽ. (d) Norms of adjustable weights of NNs systems. (e) Control input intermediate ω(t). (f) Trajectory of u(t) and sat(u(t)).

Fig. 7. Tracking error e and its estimates ê for Example 2.


288 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 6, NO. 1, JANUARY 2019

Fig. 8. New states variables s along with their estimate ŝ for Example 2.

Fig. 9. Simulation results of Example 2. (a) Performance comparison of different controllers without input saturation. (b) Control input u(t) comparison of
different controllers.

∂α1 ˙ ∂α1 ∂α1


where H2 = k2 e1 + Ŵ2T S2 (x̂2 ) − ∂ Ŵ1
Ŵ 1 − ∂yd ẏd − ∂ ẏd ÿd V. C ONCLUSION
∂α1 T
− ∂y [x̂2 − Ŵ1 S1 (x1 )], and z1 = y − yd , z2 = x̂2 − α1 , e1 In this work, an NN adaptive control approach is proposed
= x2 − x̂2 . The state observer used in [46] is defined by for a class of input saturated pure-feedback nonlinear systems.
By introducing a low-pass filter followed by new states def-
 initions, the original input saturated non-affine pure-feedback
x̂˙ 1 = x̂2 + Ŵ1T S1 (x1 ) + α1 (y − x̂1 )
system is reformulated into the affine Bronovski canonical
˙ form. Compared with other approaches, the proposed tech-
x̂2 = g(v) + Ŵ2T S2 (x̂2 ) + α2 (y − x̂1 ).
nique avoids the back-stepping usually used in the literature
and then eliminates the explosion of complexity problem. The
It should be noted that two NNs are used in the back- proposed control algorithm is considerably simpler and can
stepping controller (62)−(64) for the second-order system handle a wider class of input saturated nonlinear systems.
(61). From the above equation defining the controller and Furthermore, only two NNs, whatever the order of the sat-
updating laws in [46], we can see clearly the explosion of urated nonlinear systems, are employed in controller design
complexity involved by the use of many design parameters in contrast to the literature where many NNs units as many
along with two NNs and one virtual control input with its as the order of the system are required. The Lyapunov theory
complicated derivatives. is used to prove the stability of the closed-loop system and
The obtained results when applying the controller defined simulation results shows that the proposed method is able to
in [46] with input saturation are shown in Fig. 9. It is very control systems with effective performance in the presence of
clear that our approach outperforms the approach proposed input saturation.
in [46] since the obtained response in Fig. 9 (a) shows slow
tracking performance compared to our approach. Furthermore, R EFERENCES
a transient larger control signal, as shown in Fig. 9 (b), is [1] H. Zargarzadeh, T. Dierks, and S. Jagannathan, “Optimal control of
applied on the system. This large control may be nonrealistic. nonlinear continuous-time systems in strict-feedback form,” IEEE Trans.
ZERARI et al.: NN BASED ADAPTIVE TRACKING CONTROL FOR A CLASS OF PURE FEEDBACK NONLINEAR SYSTEMS WITH INPUT SATURATION 289

Neural Netw. Learn. Syst., vol. 26, no. 10, pp. 2535−2549, Oct. 2015. systems with unknown dead zone in pure feedback form,” Automatica,
vol. 44, no. 7, pp. 1895−1903, Jul. 2008.
[2] C. P. Bechlioulis and G. A. Rovithakis, “Adaptive control with guaran-
teed transient and steady state tracking error bounds for strict feedback [21] W. C. Meng, Q. M. Yang, S. Jagannathan, and Y. X. Sun, “Adaptive neu-
systems,” Automatica, vol. 45, no. 2, pp. 532−538, Feb. 2009. ral control of high-order uncertain nonaffine systems: a transformation
to affine systems approach,” Automatica, vol. 50, no. 5, pp. 1473−1480,
[3] Y. P. Pan, Y. Q. Liu, and H. Y. Yu, “Simplified adaptive neural control of May 2014.
strict-feedback nonlinear systems,” Neurocomputing, vol. 159, pp. 251−
256, Jul. 2015. [22] J. H. Park, S. H. Kim, and C. J. Moon, “Adaptive neural control for
strict-feedback nonlinear systems without backstepping,” IEEE Trans.
[4] H. Q. Wang, B. Chen, and C. Lin, “Adaptive neural control for strict- Neural Netw., vol. 20, no. 7, pp. 1204−1209, Jul. 2009.
feedback stochastic nonlinear systems with time-delay,” Neurocomput-
ing, vol. 77, no. 1, pp. 267−274, 2012. [23] J. Na, X. M. Ren, and D. D. Zheng, “Adaptive control for nonlinear
pure-feedback systems with high-order sliding mode observer,” IEEE
[5] M. Wang, B. Chen, X. P. Liu, and P. Shi, “Adaptive fuzzy tracking Trans. Neural Netw. Learn. Syst., vol. 24, no. 3, pp. 370−382, Mar.
control for a class of perturbed strict-feedback nonlinear time-delay 2013.
systems,” Fuzzy Sets Syst., vol. 159, no. 8, pp. 949−967, Apr. 2008.
[24] S. C. Tong and Y. M. Li, “Adaptive fuzzy output feedback tracking
[6] H. Q. Wang, X. P. Liu, S. Li, and D. Wang, “Adaptive neural output- backstepping control of strict-feedback nonlinear systems with unknown
feedback control for a class of nonlower triangular nonlinear systems dead zones,” IEEE Trans. Fuzzy Syst., vol. 20, no. 1, pp. 168−180, Feb.
with unmodeled dynamics,” IEEE Trans. Neural Netw. Learn. Syst., vol. 2012.
29, no. 8, pp. 3658−3668, Aug. 2018.
[25] Z. Liu, F. Wang, Y. Zhang, X. Chen, and C. L. P. Chen, “Adaptive
[7] D. Wang and J. Huang, “Adaptive neural network control for a class of tracking control for a class of nonlinear systems with a fuzzy dead-zone
uncertain nonlinear systems in pure-feedback form,” Automatica., vol. input,” IEEE Trans. Fuzzy Syst., vol. 23, no. 1, pp. 193−204, Feb. 2015.
38, no. 8, pp. 1365−1372, Aug. 2002.
[26] B. Chen, X. P. Liu, K. F. Liu, and C. Lin, “Fuzzy approximation-based
[8] W. Rui, F. S. Yu, and J. Y. Wang, “Neural network-based adaptive adaptive control of nonlinear delayed systems with unknown dead zone,”
tracking control for a class of uncertain stochastic nonlinear pure- IEEE Trans. Fuzzy Syst., vol. 22, no. 2, pp. 237−248, Apr. 2014.
feedback systems,” in Proc. 25th Chinese Control and Decision Conf.,
Guiyang, China, 2013, pp. 495−500. [27] F. Wang, B. Chen, C. Lin, J. Zhang, and X. Z. Meng, “Adaptive neural
network finite-time output feedback control of quantized nonlinear
[9] Y. M. Li, S. C. Tong, and T. S. Li, “Adaptive fuzzy backstepping
systems,” IEEE Trans. Cybern., vol. 48, no. 6, pp. 1839−1848, Jun.
control design for a class of pure-feedback switched nonlinear systems,”
2018.
Nonlinear Anal. Hybrid Syst., vol. 16, pp. 72−80, May 2015.
[28] J. Zhou, C. Y. Wen, and G. H. Yang, “Adaptive backstepping stabilization
[10] Y. Gao, S. C. Tong, and Y. M. Li, “Adaptive fuzzy backstepping output
of nonlinear uncertain systems with quantized input signal,” IEEE Trans.
feedback control for a class of uncertain stochastic nonlinear system
Autom. Control, vol. 59, no. 2, pp. 460−464, Feb. 2014.
in pure-feedback form,” Neurocomputing, vol. 122, pp. 126−133, Dec.
2013. [29] Q. Ma, G. Z. Cui, and T. C. Jiao, “Neural-network-based adaptive
tracking control for a class of pure-feedback stochastic nonlinear systems
[11] S. S. Ge, C. G. Yang, and T. H. Lee, “Adaptive predictive control using
with backlash-like hysteresis,” Neurocomputing, vol. 144, pp. 501−508,
neural network for a class of pure-feedback systems in discrete time,”
Nov. 2014.
IEEE Trans. Neural Netw., vol. 19, no. 9, pp. 1599−1614, Sep. 2008.
[30] J. Zhou, C. J. Zhang, and C. Y. Wen, “Robust adaptive output control of
[12] Z. F. Chen, S. S. Ge, Y. Zhang, and Y. N. Li, “Adaptive neural control of
uncertain nonlinear plants with unknown backlash nonlinearity,” IEEE
MIMO nonlinear systems with a block-triangular pure-feedback control
Trans. Autom. Control, vol. 52, no. 3, pp. 503−509, Mar. 2007.
structure,” IEEE Trans. Neural Netw. Learn. Syst., vol. 25, no. 11, pp.
2017−2029, Nov. 2014. [31] B. B. Ren, S. S. Ge, C. Y. Su, and T. H. Lee, “Adaptive neural control
[13] T. P. Zhang, X. C. Shi, Q. Zhu, and Y. Q. Yang, “Adaptive neural tracking for a class of uncertain nonlinear systems in pure-feedback form with
control of pure-feedback nonlinear systems with unknown gain signs and hysteresis input,” IEEE Trans. Syst., Man, Cybern. B Cybern., vol. 39,
unmodeled dynamics,” Neurocomputing, vol. 121, pp. 290−297, Dec. no. 2, pp. 431−443, Apr. 2009.
2013. [32] Y. F. Gao, X. M. Sun, C. Y. Wen, and W. Wang, “Observer-based
[14] B. S. Kim and S. J. Yoo, “Adaptive control of nonlinear pure-feedback adaptive NN control for a class of uncertain nonlinear systems with
systems with output constraints: integral barrier Lyapunov functional nonsymmetric input saturation,” IEEE Trans. Neural Netw. Learn. Syst.,
approach,” Int. J. Control Autom. Syst., vol. 13, no. 1, pp. 249−256, vol. 28, no. 7, pp. 1520−1530, Jul. 2017.
Feb. 2015. [33] X. D. Zhao, H. J. Yang, W. G. Xia, and X. Y. Wang, “Adaptive fuzzy
[15] C. L. P. Chen, Y. J. Liu, and G. X. Wen, “Fuzzy neural network-based hierarchical sliding-mode control for a class of MIMO nonlinear time-
adaptive control for a class of uncertain nonlinear stochastic systems,” delay systems with input saturation,” IEEE Trans. Fuzzy Syst., vol. 25,
IEEE Trans. Cybern., vol. 44, no. 5, pp. 583−593, May 2014. no. 5, pp. 1062−1077, Oct. 2017.

[16] C. Wang, D. J. Hill, S. S. Ge, and G. R. Chen, “An ISS-modular approach [34] T. S. Li, R. H. Li, and J. F. Li, “Decentralized adaptive neural control of
for adaptive neural control of pure-feedback systems,” Automatica, vol. nonlinear interconnected large-scale systems with unknown time delays
42, no. 5, pp. 723−731, May 2006. and input saturation,” Neurocomputing, vol. 74, no. 14−15, pp. 2277−
2283, Jul. 2011.
[17] J. Zhou and C. Y. Wen, Adaptive Backstepping Control of Uncertain
Systems. Berlin, Heidelberg, Germany: Springer, 2008, pp. 723−731. [35] M. Chen, G. Tao, and B. Jiang, “Dynamic surface control using
neural networks for a class of uncertain nonlinear systems with input
[18] D. Swaroop, J. K. Hedrick, P. P. Yip, and J. C. Gerdes, “Dynamic surface saturation,” IEEE Trans. Neural Netw. Learn. Syst., vol. 26, no. 9, pp.
control for a class of nonlinear systems,” IEEE Trans. Autom. Control, 2086−2097, Sep. 2015.
vol. 45, no. 10, pp. 1893−1899, Oct. 2000.
[36] H. Q. Wang, B. Chen, X. P. Liu, K. F. Liu, and C. Lin, “Adaptive
[19] M. Wang, X. P. Liu, and P. Shi, “Adaptive neural control of pure- neural tracking control for stochastic nonlinear strict-feedback systems
feedback nonlinear time-delay systems via dynamic surface technique,” with unknown input saturation,” Inf. Sci., vol. 269, pp. 300−315, Jun.
IEEE Trans. Syst., Man, Cybern. B Cybern., vol. 41, no. 6, pp. 1681− 2014.
1692, Dec. 2011.
[37] C. Y. Wen, J. Zhou, Z. T. Liu, and H. Y. Su, “Robust adaptive control
[20] T. P. Zhang and S. S. Ge, “Adaptive dynamic surface control of nonlinear of uncertain nonlinear systems in the presence of input saturation and
290 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 6, NO. 1, JANUARY 2019

external disturbance,” IEEE Trans. Autom. Control, vol. 56, no. 7, pp. [53] T. P. Zhang, Q. Zhu, and Y. Q. Yang, “Adaptive neural control of
1672−1678, Jul. 2011. non-affine pure-feedback non-linear systems with input nonlinearity and
perturbed uncertainties,” Int. J. Syst. Sci., vol. 43, no. 4, pp. 691−706,
[38] Y. M. Li, S. C. Tong, and T. S. Li, “Composite adaptive fuzzy output Apr. 2012.
feedback control design for uncertain nonlinear strict-feedback systems
with input saturation,” IEEE Trans. Cybern., vol. 45, no. 10, pp. 2299− [54] J. Park and I. W. Sandberg, “Universal approximation using radial-basis-
2308, Oct. 2015. function networks,” Neural Comput., vol. 3, no. 2, pp. 246−257, Jun.
1991.
[39] Y. M. Li, S. C. Tong, and T. S. Li, “Direct adaptive fuzzy backstep-
ping control of uncertain nonlinear systems in the presence of input [55] M. M. Gupta and D. H. Rao, Neuro-Control Systems, Theory and
saturation,” Neural Comput. Appl., vol. 23, no. 5, pp. 1207−1216, Oct. Applications. Piscataway, NJ, USA: IEEE Press, 1994.
2013. [56] S. S. Ge, C. C. Hang, T. H. Lee, and T. Zhang, Stable Adaptive Neural
Network Control. Boston, MA, USA: Springer, 2002.
[40] Y. J. Liu and S. C. Tong, “Adaptive fuzzy control for a class of unknown
nonlinear dynamical systems,” Fuzzy Sets Syst., vol. 263, pp. 49−70, [57] M. French and E. Rogers, “Approximate models for adaptive feedback
Mar. 2015. linearization,” Int. J. Control, vol. 68, no. 6, pp. 1305−1322, Dec. 1997.
[41] H. Q. Wang, X. P. Liu, and K. F. Liu, “Adaptive neural data-based [58] N. Kim and A. J. Calise, “Several extensions in methods for adaptive
compensation control of non-linear systems with dynamic uncertainties output feedback control,” IEEE Trans. Neural Netw., vol. 18, no. 2, pp.
and input saturation,” IET Control Theory Appl., vol. 9, no. 7, pp. 1058− 482−494, Mar. 2007.
1065, Apr. 2015.
[59] S. S. Sastry and A. Isidori, “Adaptive control of linearizable systems,”
[42] S. G. Gao, B. Ning, and H. R. Dong, “Fuzzy dynamic surface control IEEE Trans. Autom. Control, vol. 34, no. 11, pp1123−1131, Nov. 1989.
for uncertain nonlinear systems under input saturation via truncated
adaptation approach,” Fuzzy Sets Syst., vol. 290, pp. 100−117, May [60] J. J. E. Slotine and W. P. Li, Applied Nonlinear Control. Englewood
2016. Cliffs, NJ, USA: Prentice-Hall, 1991.
[61] M. M. Polycarpou and P. A. Ioannou, “A robust adaptive nonlinear
[43] Q. Zhou, P. Shi, Y. Tian, and M. Y. Wang, “Approximation-based control design,” Automatica, vol. 32, no. 3, pp. 423−427, Mar. 1996.
adaptive tracking control for MIMO nonlinear systems with input
saturation,” IEEE Trans. Cybern., vol. 45, no. 10, pp. 2119−2128, Oct.
2015.

[44] Z. T. Chen, Z. J. Li, and C. L. P. Chen, “Adaptive neural control of


uncertain MIMO nonlinear systems with state and input constraints,” Zerari Nassira received the master degree in au-
IEEE Trans. Neural Netw. Learn. Syst., vol. 28, no. 6, pp. 1318−1330, tomatic control from the University of Constantine
Jun. 2017. 1, Algeria, in 2014. Currently, she is pursuing the
Ph.D. degree at the same university. Her research
[45] X. H. Yan, M. Chen, Q. X. Wu, and S. Y. Shao, “Dynamic surface interests include adaptive neural network control and
control for a class of stochastic non-linear systems with input saturation,” constrained nonlinear control.
IET Control Theory Appl., vol. 10, no. 1, pp. 35−43, Jan. 2016.

[46] Y. M. Li, S. C. Tong, and T. S. Li, “Adaptive fuzzy output-feedback


control for output constrained nonlinear systems in the presence of input
saturation,” Fuzzy Sets Syst., vol. 248, pp. 138−155, Aug. 2014.

[47] Q. Zhou, L. J. Wang, C. W. Wu, H. Y. Li, and H. P. Du, “Adaptive fuzzy


control for nonstrict-feedback systems with input saturation and output
constraint,” IEEE Trans. Syst., Man, Cybern.: Syst., vol. 47, no. 1, pp. Chemachema Mohamed received the Ph.D. degree
1−12, Jan. 2017. in automatic control from the University of Con-
stantine 1, Algeria, in 2007. He is currently a full
[48] W. H. Liu, J. W. Lu, Z. Q. Zhang, and S. Y. Xu, “Observer-based Professor at Constantine 1 University, Algeria. His
neural control for MIMO pure-feedback non-linear systems with input research interests include nonlinear robust adaptive
saturation and disturbances,” IET Control Theory Appl., vol. 10, no. 17, control, fault-tolerant control, and fuzzy/neural con-
pp. 2314−2324, Nov. 2016. trol.
[49] W. J. Si, X. D. Dong, and F. F. Yang, “Adaptive neural control for
stochastic pure-feedback non-linear time-delay systems with output
constraint and asymmetric input saturation,” IET Control Theory Appl.,
vol. 11, no. 14, pp. 2288−2298, Sep. 2017.

[50] H. Q. Wang, B. Chen, X. P. Liu, K. C. Liu, and C. Lin, “Robust adaptive


fuzzy tracking control for pure-feedback stochastic nonlinear systems
with input constraints,” IEEE Trans. Cybern., vol. 43, no. 6, pp. 2093− Essounbouli Najib received the Ph.D. degree in
2104, Dec. 2013. automatic control from the University of Reims
Champagne Ardenne, France, in 2004. Since 2010,
[51] S. Sui, S. C. Tong, and Y. M. Li, “Adaptive fuzzy backstepping output he has been a full Professor. His current research
feedback tracking control of MIMO stochastic pure-feedback nonlinear interests include fuzzy logic control, robust adaptive
systems with input saturation,” Fuzzy Sets Syst., vol. 254, pp. 26−46, control, renewable energy, and drive control.
Nov. 2014.

[52] P. A. Ioannou and J. Sun, Robust Adaptive Control. Upper Saddle River,
NJ, USA: Prentice-Hall, 1996.

You might also like