8102 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 34, NO.

10, OCTOBER 2023

Adaptive Neural Asymptotic Tracking of Uncertain Non-Strict Feedback

Systems With Full-State Constraints via Command Filtered Technique
Chun Xin, Yuan-Xin Li , and Choon Ki Ahn , Senior Member, IEEE

Abstract— This brief addresses the adaptive neural asymptotic tracking virtual controller αi−1 (i = 2, . . . , n) must meet an essential
issue for uncertain non-strict feedback systems subject to full-state feasibility condition −kai < αi−1 < kbi , where kai and kbi are
constraints. By introducing the significant nonlinear transformed function
the constraint boundaries. To eliminate these feasibility conditions,
(NTF), the command filtered technology, and the boundary estimation
method into control design, a novel command filtered backstepping Zhao and Song [23] introduced a nonlinear state-dependent function
adaptive controller is proposed. The proposed control scheme is able that completely relies on the constrained states. In [24], by integrating
to not only deal with full-state constraints but also avoid the “explosion a nonlinear transformed function (NTF) with the bound estimation
of complexity” issue. By means of a Lyapunov stability analysis, we prove methods, the proposed robust control scheme can handle full-state
that: 1) the tracking error asymptotically converges to zero; 2) all the
variables in the controlled systems are bounded; and 3) all the states constraints and reduce the computational burden.
are constrained in the asymmetric predefined sets. Finally, a numerical The “explosion of complexity” issue will often arise in the
simulation is used to demonstrate the validity of the proposed algorithm. aforementioned methods due to continuous derivation of virtual
control signals during the recursive process. To circumvent this
Index Terms— Asymptotic tracking control, command filter
disadvantage, a command filtering control via backstepping was
backstepping, full-state constraints, neural networks (NNs),
non-strict feedback systems. developed in [25] and [26]. Yu et al. [27] combined a command
filter with backstepping control to overcome model uncertainties and
reduce the computational complexity. By fusing the command filter
I. I NTRODUCTION and backstepping approach, exact adaptive tracking controllers have
Prior studies have explored the tracking control problem of been skillfully devised for nonlinear systems with uncertainties [28].
uncertain nonlinear systems with many applications related to motors, To obtain better tracking performance by virtue of the command filter,
robots, and wind turbine systems [1]. A variety of constructive finite-time controllers were then constructed in [29], [30], [30], [31].
adaptive backstepping schemes have been proposed [2]. In particular, Taking the effect of full-state constraints into account, Qiu et al. [32]
adaptive control design with fuzzy logic systems (FLSs) or neural net- designed a robust controller for nonlinear systems by associating
works (NNs) has generated an abundance of research results [3]–[12]. the BLFs and command filter technique. However, only bounded
A common characteristic of the aforementioned algorithms is that error tracking could be achieved in the aforementioned results, and
they mainly focus on strict feedback systems. However, in practical the tracking of asymptotic errors was not possible. Furthermore,
applications, controlled systems generally have a non-strict feedback feasibility conditions are needed to satisfy the full-state constraints
form. Thus, it is important to research the tracking control problem in [32]. Hence, how to develop a new adaptive control algorithm
of non-strict feedback systems. In recent decades, many excellent for nonlinear systems without the feasibility condition and explosion
studies on this problem have been produced [13]–[17]. of complexity problem is a challenging work that requires further
In many applications, state and output constraints are omnipresent investigation.
due to safety considerations, physical limitations, and/or performance To fill this gap, we investigate a novel command filtered adaptive
requirements [18]. Safety accidents and system instability may arise asymptotic tracking scheme for non-strict feedback systems under
when constraints are violated. Therefore, many control approaches the circumstance of full-state constraints. The brief’s most crucial
have been proposed to address state or output constraint prob- contributions are as follows.
lems [19]–[24]. More specifically, barrier Lyapunov functions (BLFs) 1) Unlike traditional backstepping as used in [21]–[24] for
have been introduced in stability analysis to ensure that the system’s constrained nonlinear systems, a new Command-filter-based
output is not transgressed [19]–[21]. Liu et al. [22] extended the backstepping method is developed to address the explosion of
work of [20] from output constraints to state constraints. However, complexity issue. Furthermore, the effect of filtered errors is
the main shortcoming of the BLF approach is that the designed eliminated by constructing a compensating signal βi .
2) Unlike the traditional command filtered technique used in
Manuscript received 27 August 2021; revised 30 November 2021; accepted [25]–[27], which only achieved bounded tracking control
4 January 2022. Date of publication 19 January 2022; date of current performance, the proposed control scheme in this brief
version 6 October 2023. This work was supported in part by the Funds enables the tracking error to converge to zero asymptotically
of the National Science of China under Grant 61973146, in part by the
Distinguished Young Scientific Research Talents Plan in Liaoning Province by introducing compensating signals, constructing smooth
under Grant XLYC1907077, and in part by the National Research Foundation functions, and using the bound estimation method.
of Korea (NRF) grant funded by the Korean Government (Ministry of Science 3) Compared with [32], where the virtual controllers must satisfy
and ICT) under Grant NRF-2020R1A2C1005449. (Corresponding authors: the feasibility conditions to ensure that all the states do not
Yuan-Xin Li; Choon Ki Ahn.)
transgress the prescribed constraints, a remarkable contribution
Chun Xin and Yuan-Xin Li are with the College of Science,
Liaoning University of Technology, Jinzhou 121001, China (e-mail: of this brief is that the feasibility conditions can be eliminated
[email protected]; [email protected]). using a new NTF.
Choon Ki Ahn is with the School of Electrical Engineering, Korea 4) It can be seen from (26) that the controller designed in this brief
University, Seoul 136-701, South Korea (e-mail: [email protected]). is a function of the partial state vector ζ̄i = [ζ1 , ζ2 , . . . , ζi ]T ;
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TNNLS.2022.3141091. therefore, the algebraic loop issue in (1) is circumvented by
Digital Object Identifier 10.1109/TNNLS.2022.3141091 applying the property of the radial basis function (RBF) NNs.
2162-237X © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: University of Ulsan. Downloaded on May 25,2025 at 13:54:44 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 34, NO. 10, OCTOBER 2023 8103

II. P ROBLEM S TATEMENT AND P REPARATORY W ORK In what follows, a new command filtered neural tracking controller
A. System Descriptions is established through backstepping technology for system (1). Before
moving on, the coordinate transformations are given as
This brief considers the following non-strict feedback systems:
yr
ζ̇i = f i (ζ ) + gi (ζ̄i )ζi+1 , 1 ≤ i ≤ n − 1 x1 = s1 −    (4)
ka1 + yr kb1 − yr
ζ̇n = f n (ζ ) + gn (ζ̄n )u ᾱi
xi = si − , i = 2, . . . n (5)
y = ζ1 (1) kζi

in which ζ̄i = [ζ1 , ζ2 , . . . , ζi ]T ∈ Ri , i = 1, 2, . . . , n and where ᾱi comes from the following command filter with the virtual
ζ = ζ̄n ∈ Rn denote the state variables, while u ∈ R and y ∈ R denote control input αi−1 :
the control input and output, respectively. fi (·) ∈ R is an unknown
i ᾱ˙ i + ᾱi = αi−1 , ᾱi (0) = αi−1 (0), i = 2, . . . , n (6)
nonlinear function, and gi (·) ∈ R is a known control coefficient
satisfying 0 < p1 < gi (·) < p2 . In addition, all the states ζi are with i (i = 2, . . . , n) being positive constants. To facilitate the
required to remain in the open sets ζi = {−kai (t) < ζi < kbi (t), following backstepping-based design procedure, we define θ̃i = θi −θ̂i
i = 1, . . . , n}, ∀t ≥ t0 , with kai (t) and kbi (t) being positive time- and ι̃i = ιi − ι̂i , i = 1, . . . , n as the estimation errors, with θ̂i and
varying functions. ι̂i denoting the estimations of θi and ιi , respectively.
The brief’s control goal is to design a novel command filtered Step 1: From (4) and (5), one can obtain
adaptive tracking controller for non-strict feedback systems, such
that: 1) all the signals in the controlled systems remain bounded on ẋ1 = K ζ1 f 1 (ζ ) + G 1 ζ2 − K yr ẏr
[0, +∞); 2) the output y follows a target signal yr whose trajectory = K ζ1 f 1 (ζ ) + G 1 (x2 kζ2 + ᾱ2 − α1 + α1 ) − K yr ẏr (7)
and its first time derivative ẏr (t) are continuous and bounded; and
3) the system states are always constrained in the predefined region. where K yr = (ka1 kb1 + yr2 )/((ka1 + yr )2 (kb1 − yr )2 ).
Lemma 1: Let ζ̄i = [ζ1 , ζ2 , . . . , ζi ]T and S(ζ̄i ) = To deal with the error term ᾱ1 − α1 , the compensating signal β1 is
[S1 (ζ̄i ), . . . , Sl (ζ̄i )] serve as the basis function vector of the
T expressed as
RBF NNs. For any integer 0 < k ≤ q, there exists l1 β1
  2   2 β̇1 = −k1 β1 + G 1 (ᾱ2 − α1 ) + G 1 kζ2 β2 − (8)
 S ζ̄q  ≤  S ζ̄k  . β12 + σ12

Proof: Let μi = [μi1 , . . . , μiq ]T , i = 1, . . . , l be the center of with k1 and l1 being known design constants, β1 (0) = 0 and
t
the receptive field, and η > 0 be the width of the Gaussian function. σi (i = 1, . . . , n) are selected such that limt→∞ t0 σi (s)ds ≤ σ̄i < ∞,
Using the Gaussian function definition, one has where σ̄i denotes the positive constants.
 q  2
Define the compensated tracking errors as
  2  l  1
 S ζ̄q  = exp − 2 (ζi − μit )2 ξ1 = x1 − β1 , ξi = xi − βi , i = 2, . . . , n. (9)
i=1 t=1
η
  2 The derivative of ξ1 along with (7) and (8) is given by
 l 
k
1   2
≤ exp − 2 (ζi − μit )2 =  S ζ̄k  .
i=1 t=1
η ξ̇1 = K ζ1 f1 (ζ ) + G 1 kζ2 ξ2 + G 1 α1 − K yr ẏr
l1 β1
The proof of Lemma 1 is completed. + k1 β1 + . (10)
β12 + σ12
III. M AIN R ESULTS The Lyapunov function candidate V1 is chosen as
A. Adaptive Controller Design 1 2 1 2 1 2
V1 = ξ + θ̃ + ι̃ (11)
In this section, an NTF is introduced to transform the constrained 2 1 2γ1 1 2r1 1
systems (1) into unconstrained systems. Then, a command filtered where γ1 > 0 and r1 > 0 are the design parameters.
adaptive neural tracking controller is developed to avoid the explosion Together with (10), the derivative of V1 is
of complexity issue that arises in the traditional backstepping 
framework. l1 β1
First, a novel NTF is represented as follows: V̇1 = ξ1 G 1 kζ2 ξ2 + G 1 α1 + + K ζ1 f 1 (ζ )
β12 + σ12
ζi   
si = , kζi = kai + ζi kbi − ζi (2) 1 ˙ 1
kζi − K yr ẏr + k1 β1 − θ̃1 θ̂1 − ι̃1 ι̂˙1 (12)
γ1 r1
with the initial states satisfying ζi (0) ∈ ζi , i = 1, . . . , n. From (2),
it can be seen that si → ∞ when ζi approaches the boundary of where f 1 (ζ ) is an unknown nonlinear function; therefore, for ∀ε1 > 0,
ζi ; i.e., for any initial state ζi (0) ∈ ζi , si → ±∞, if and only if there exists an NN W1T ϕ1 (ζ ) such that
ζi → −kai or ζi → kbi . Obviously, for any ζi (0) ∈ ζi ,
f 1 (ζ ) = W1T ϕ1 (ζ ) + ν1 (ζ ), ν1 (ζ ) ≤ ε1 (13)
if si ∈ L ∞ , ∀t ≥ 0, it can be guaranteed that ζi (t) ∈ ζi , ∀t ≥ 0,
that is, the full-state constraints are strictly maintained when the where ν1 (ζ ) is the approximation error.
boundedness of si on [0, +∞). In view of Lemma 1, we can obtain
Based on (1) and (2), the new systems can be described as  
K ζ1 ξ1 W1T ϕ1 (ζ ) + ν1 (ζ )
ṡi = K ζi f i (ζ ) + G i ζi+1 , i = 1, . . . , n − 1
≤ |K ζ1 ||ξ1 |W1 ||ϕ1 (ζ )|| + |K ζ1 ||ξ1 |||ν1 (ζ )||
ṡn = K ζn f n (ζ ) + G n u (3)
≤ θ1 |ξ1 ||K ζ1 |||ϕ1 (ζ1 )|| + |K ζ1 ||ξ1 |ι1
with G i = K ζi gi , and K ζi = ((kai kbi + ζi2 )/kζ2i ). ≤ η1 ξ12 θ1 + β1 ξ12 ι1 + σ1 (θ1 + ι1 ) (14)

Authorized licensed use limited to: University of Ulsan. Downloaded on May 25,2025 at 13:54:44 UTC from IEEE Xplore. Restrictions apply.
8104 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 34, NO. 10, OCTOBER 2023

where θ1 = ||W1 ||, ι1 = ε1 , η1 = ((K ζ21 ϕ1T (ζ1 )ϕ1 (ζ1 ))/ Substituting (24) into (22) leads to
(ξ12 K ζ21 ϕ1T (ζ1 )ϕ1 (ζ1 ) + σ12 )1/2 ), and β1 = (K ζ21 /(K ζ21 ξ12 + σ12 )1/2 ).

i−1 
i
Substituting(14) into (12) results in V̇i ≤ − k j ξ 2j + σ j (θ j + ι j ) + G i−1 kζi ξi−1 ξi
j =1 j =1
V̇1 ≤ ξ1 G 1 kζ2 ξ2 + G 1 α1 + k1 β1 − K yr ẏr + β1 ξ1 ι̂1 
+ ξi ki βi +G i kζi+1 ξi+1 +G i αi − K ζi ᾱ˙ i−1 + G i−1 kζi βi−1
l1 β1 1  
+ η1 ξ1 θ̂1 + + θ̃1 γ1 η1 ξ12 − θ̂˙1
β12 + σ12 γ1 li βi
+ ξi ηi θ̂i + ξi βi ι̂i +
1   βi2 + σi2
+ ι̃1 r1 β1 ξ12 − ι̂˙1 + σ1 (θ1 + ι1 ). (15)
r1 1  
+ θ̃i γi ηi ξi2 − θ̂˙i
Next, the virtual control signal α1 and the adaptive laws θ̂1 and γi
ι̂1 are constructed
 as 1    i−1  i−1

1 l1 β1 + ι̃i ri βi ξi2 − ι̂˙i + σ j ι̃ j ι̂ j + σ j θ̃ j θ̂ j . (25)

α1 = −k1 x1 − − θ̂1 ξ1 η1 − β1 ξ1 ι̂1 + K yr ẏr ri j =1 j =1
G1 β12 + σ12
The virtual control signal and adaptive laws are designed as
θ̂˙1 = γ1 η1 ξ12 − γ1 σ1 θ̂1 
1
ι̂˙1 = r1 β1 ξ12 − r1 σ1 ι̂1 . (16) αi = − ki xi − ξi ηi θ̂i − ξi βi ι̂i
Gi
Substituting (16) into (15) leads to li βi
V̇1 ≤ −k1 ξ12 + G 1 kζ2 ξ1 ξ2 + σ1 (θ1 + ι1 ) − G i−1 kζi xi−1 − + K ζi ᾱ˙ i−1
βi2 + σi2
+ σ1 θ̃1 θ̂1 + σ1 ι̃1 ι̂1 . (17)
θ̂˙i = γi ηi ξi2 − γi σi θ̂i
Step i (i = 2, . . . , n − 1): From (5) and (9), it follows that
ι̂˙i = ri βi ξi2 − ri σi ι̂i . (26)
ẋi = K ζi f i (ζ ) − K ζi ᾱ˙ i−1 + G i (xi+1 kζi+1 + ᾱi ) (18)
Putting (26) into (25), one has
where K ζi = (kai kbi + ζi2 )/kζ2i . i 
i

The compensating signal βi is designed as V̇i ≤ − k j ξ 2j + G i kζi+1 ξi ξi+1 + σ j (θ j + ι j )

j =1 j =1
β̇i = −ki βi + G i (ᾱi+1 − αi + kζi+1 βi+1 )

i 
i
li βi + σ j θ̃ j θ̂ j + σ j ι̃ j ι̂ j . (27)
− G i−1 kζi βi−1 − (19)
βi2 + σi2 j =1 j =1

where βi (0) = 0, ki , and li are known positive constants. Step n: The compensating signal βn is designed as
Based on (9), (18), and (19), the time derivative of ξi is ln βn
β̇n = −kn βn − G n−1 kζn βn−1 − (28)
li βi βn2 + σn2
ξ̇i = K ζi f i (ζ ) + G i kζi+1 ξi+1 + G i αi +
βi2 + σi2 where kn and ln are known positive constants, βn (0) = 0.
+ G i−1 kζi βi−1 − K ζi ᾱ˙ i−1 + ki βi . (20) Design the controller and adaptive laws as

1
The Lyapunov function Vi is selected as u = − kn xn − ξn ηn θ̂n − ξn βn ι̂n
Gn
1 1 2 1 2
Vi = Vi−1 + ξi2 + θ̃ + ι̃ (21) ln βn
2 2γi i 2ri i − G n−1 kζn xn−1 − + K ζn ᾱ˙ n−1
with γi > 0 and ri > 0 being known parameters. Differentiating both βn2 + σn2
sides of (21) gives θ̂˙n = γn ηn ξn2 − γn σn θ̂n

ι̂˙n = rn βn ξn2 − rn σn ι̂n . (29)
V̇i ≤ V̇i−1 + ξi G i kζi+1 ξi+1 + G i αi + ki βi
Then, construct a Lyapunov function Vn as
+ K ζi f i (ζ ) + G i−1 kζi βi−1 − K ζi ᾱ˙ i−1 1 1 2 1 2
Vn = Vn−1 + ξn2 + θ̃ + ι̃ (30)
li βi 1 ˙ 1 2 2γn n 2rn n
+ − θ̃i θ̂i − ι̃i ι̂˙i . (22)
βi2 + σi2 γi ri with γn > 0 and rn > 0 as design parameters. A similar computation
can be performed based on step i to conclude that
Using NNs, for any given εi > 0, an NN WiT ϕi (ζ ) can be used to
approximate the unknown function f i (ζ ) 
n 
n
V̇n ≤ − ki ξi2 + σi (θi + ιi + θ̃i θ̂i + ι̃i ι̂i ). (31)
f i (ζ ) = WiT ϕi (ζ ) + νi (ζ ), νi (ζ ) ≤ εi (23) i=1 i=1

By applying the square formula, we can obtain

where νi (ζ ) is the approximation error.
Invoking Lemma 1, we can obtain θi2 ι2
  θ̃i θ̂i = −θ̃i2 + θ̃i θi ≤
, ι̃i ι̂i = −ι̃i2 + ι̃i ιi ≤ i . (32)
4 4
K ζi ξi WiT ϕi (ζ ) + νi (ζ )
This together with (31) results in
≤ θi |ξi ||K ζi |||ϕi (ζ̄i )|| + |K ζi ||ξi |ιi

n 
n
≤ ηi ξi2 θi + βi ξi2 ιi + σi (θi + ιi ) (24) V̇n ≤ − ki ξi2 + σi νi (33)
where θi = ||Wi ||, ιi = εi , ηi = ((K ζ2i ϕiT (ζ̄i )ϕi (ζ̄i ))/ i=1 i=1

(ξi2 K ζ2i ϕiT (ζ̄i )ϕi (ζ̄i ) + σi2 )1/2 ), and βi = (K ζ2i /(K ζ2i ξi2 + σi2 )1/2 ). where νi = θi + ιi + (θi2 + ιi2 )/4, i = 1, . . . , n.

Authorized licensed use limited to: University of Ulsan. Downloaded on May 25,2025 at 13:54:44 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 34, NO. 10, OCTOBER 2023 8105

Based on the above analysis, the main contribution is introduced Invoking Barbalat’s Lemma [33], one has
in Theorem 1.
lim ξi (t) = 0, i = 1, . . . , n
t→∞
lim βi (t) = 0, i = 1, . . . , n. (39)
B. Stability Analysis t→∞

Theorem 1: Consider the nonlinear non-strict feedback system (1) Combining this with the definition of ξi = xi − βi , we can infer
and the controller (29). Then, all the variables in the considered that
system are bounded, all the signals remain in their constrained lim xi (t) = 0, i = 1, . . . , n. (40)
condition at all the times, and asymptotic tracking can be t→∞

achieved. Then, by the tracking error y(t)− yr (t) = ζ1 (t)− yr (t), (2), and (5),
Proof: To testify the boundedness of all the variables it holds that
in controlled systems, consider the following Lyapunov
y(t) − yr (t) = (t)x1 (t) (41)
function:

n
1 with (t) = (1 (t))/(2 (t)), where 1 (t) = (ka1 + ζ1 (t))
Vn+1 = Vn + βi2 . (34) (kb1 − ζ1 (t))(ka1 + yr (t))(kb1 − yr (t)) > 0 and 2 (t) = ka1 kb1 +
2
i=1 ζ1 (t)yr (t) > 0 are bounded functions. Based on (40) and (41), one
Recalling the results of (8), (19), (28), and (33), differentiating may conclude that limt→∞ (y(t) − yr (t)) = 0 asymptotic tracking is
Vn+1 gives achieved.
Remark 1: In contrast to the existing relevant works, our method’s
l1 β12 major distinction is as follows.
V̇n+1 = V̇n − k1 β12 + G 1 β1 (ᾱ2 − α1 + kζ2 β2 ) −
β12 + σ12 1) A modified command-filter-based backstepping control method
ln βn2 is proposed by fusing an NTF and command filtered
+ · · · − kn βn2 − G n−1 kζn βn βn−1 − technique. Unlike [21] and [22] under the condition of
βn2 + σn2
full-state constraints, both the feasibility condition of the

n
  
n
≤− ki ξi2 + βi2 + σi (νi + li ) virtual controller and the explosion of complexity problem are
i=1 i=1
circumvented.
n−1 
n 2) Compared with the command-filter-based backstepping in
+ |G i ||βi ||ᾱi+1 − αi | − li |βi |. (35) [25]–[27] where only bounded error tracking results could
i=1 i=1 be achieved, asymptotic error tracking has been enabled
It follows from [25] that |ᾱi+1 − αi | ≤ τi . (35) then becomes by introducing compensating signals, smooth functions, and
the bound estimation method in this brief. Meanwhile, the

n
  n
compensating signal βi could also be achieved in that
V̇n+1 ≤ − ki ξi2 + βi2 + σi (νi + li ) limt→∞ βi = 0.
i=1 i=1
3) The “algebraic loop problems” caused by a non-strict feedback
 n
− (li − Ḡ i τi )|βi | form have been addressed by applying the property of the RBF
i=1 NNs and bound estimation method.
n
  n Remark 2: As discussed in [11], [25], and [26], |ᾱi − αi−1 |
≤− ki ξi2 + βi2 + σi (νi + li ) (36) will converge exponentially to the set [−ε, ε] in the finite time
i=1 i=1 T1 , where ε, T1 > 0 can be made arbitrarily small by selecting
an adequately small i in (6). This implies that the command
where τi is a known constant, li satisfies li − Ḡ i τi > 0, and Ḡ i is the
filter (6) serves as the fast dynamics while the resulting close-loop
upper bound of G i . Integrating both sides of (36) gives
systems are the slow dynamics. However, decreasing i may lead

n t
to the peaking phenomenon in ᾱ˙ i and thereby initial spikes in u(t).
Vn+1 (t) ≤ Vn+1 (t0 ) − ki ξi2 (τ )dτ Moreover, increasing i may lead to a narrower filter’s bandwidth.
i=1 t0
Consequently, a tradeoff should be made in selecting the optimal

n t 
n
− ki βi2 (τ )dτ + (li + νi )σ̄i (37) parameter i in practice.
i=1 t0 i=1
IV. S IMULATION S TUDIES
which means that all the variables of the controlled system, i.e., ξi , βi ,
θ̃i , and ι̃i (i = 1, . . . , n) are bounded. According to ξi = xi − βi , xi is To demonstrate the validity of the proposed control strategy,
shown to be bounded. Due to s1 = x1 + (yr /((ka1 + yr )(kb1 − yr ))), we considered the following non-strict feedback systems:
we can obtain that s1 ∈ L ∞ . Combining this with (2), we have ζ̇1 = 1 − cos(ζ1 ζ2 ) + (2.5 + 0.5 sin ζ1 )x2
ζ1 ∈ ζ1 . Furthermore, αi and ᾱi are also uniformly ultimately
ζ̇2 = ζ12 eζ2 + (2 + sin(ζ1 ζ2 ))u
bounded. By applying an analysis similar to the above recursively, we
can obtain that si is bounded, which implies that ζi ∈ ζi . Therefore, y = ζ1 . (42)
the states never transgress their prescribed regions. The brief’s main purpose is to make the states ζi (t) satisfy the
Next, we will prove that tracking errors converge to zero following asymmetric constraints:
asymptotically. According to (37), there exists
yr (t) − ka1 (t) < ζ1 (t) < yr (t) + kb1 (t)

n t 
n t
lim ki βi2 (τ )dτ + lim ki ξi2 (τ )dτ −ka2 (t) < ζ2 (t) < kb2 (t). (43)
t→∞ t0 t→∞ t0
i=1 i=1 −0.3t
The related parameters are given as: ka1 = 1 + 2 ,

n
≤ Vn+1 (t0 ) + (li + νi )σ̄i < ∞. (38) kb1 = 0.9 + 2−0.2t , ka2 = 3 + 0.2 sin(t + 4.5), kb2 = 1.7 − 2−0.3t ,
i=1 k1 = 2, k2 = 2, γ1 = γ2 = 2, r1 = r2 = 2, l1 = 0.1,

Authorized licensed use limited to: University of Ulsan. Downloaded on May 25,2025 at 13:54:44 UTC from IEEE Xplore. Restrictions apply.
8106 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 34, NO. 10, OCTOBER 2023

of ζ2 under the state constraint (43), proving that the system state
remains in the predetermined region. Fig. 5 shows the curves of
parameters’ estimation. Overall, the simulation images indicate the
efficacy of the proposed adaptive control algorithm.

V. C ONCLUSION
This brief proposed a novel command filtered neural tracking
Fig. 1. Trajectories of states y and yr . control framework for non-strict feedback systems suffering from
full-state constraints. The proposed scheme is primarily characterized
by the application of NTF and command filtered technique to
address full-state constraints and the explosion of complexity issue
in the recursion-based design process. Through a Lyapunov stability
analysis, we were able to prove that the tracking errors asymptotically
converged to zero, every controlled system signal remained bounded,
and the time-varying constraints were satisfied. Finally, the above
simulation example demonstrated the validity of the developed
technique.

Fig. 2. Tracking error z 1 .

R EFERENCES
[1] M. Krstic, P. V. Kokotovic, and I. Kanellakopoulos, Nonlinear and
Adaptive Control Design. Hoboken, NJ, USA: Wiley, 1995.
[2] J. Zhou and C. Y. Wen, Adaptive Backstepping Control of Uncertain
Systems: Nonsmooth Nonlinearities, Interactions or Time-Variations.
New York, NY, USA: Springer, 2008.
[3] Y. Wang, C. K. Ahn, H. Yan, and S. Xie, “Fuzzy control and filtering
for nonlinear singularly perturbed Markov jump systems,” IEEE Trans.
Cybern., vol. 51, no. 1, pp. 297–308, Jan. 2021.
[4] S. S. Ge and C. Wang, “Adaptive neural control of uncertain
MIMO nonlinear systems,” IEEE Trans. Neural Netw., vol. 15, no. 3,
Fig. 3. Control input. pp. 674–692, May 2004.
[5] Z. Wang, L. Liu, and H. Zhang, “Neural network-based model-free
adaptive fault-tolerant control for discrete-time nonlinear systems with
sensor fault,” IEEE Trans. Syst., Man, Cybern., Syst., vol. 47, no. 8,
pp. 2351–2362, Apr. 2017.
[6] L. Chen, Y. Zhu, and C. K. Ahn, “Adaptive neural network-based
observer design for switched systems with quantized measurements,”
IEEE Trans. Neural Netw. Learn. Syst., early access, Dec. 10, 2021,
doi: 10.1109/TNNLS.2021.3131412.
[7] D. Wang and J. Huang, “Neural network-based adaptive dynamic
surface control for a class of uncertain nonlinear systems in strict-
feedback form,” IEEE Trans. Neural Netw., vol. 16, no. 1, pp. 195–202,
Jan. 2005.
Fig. 4. State trajectory ζ2 . [8] S. Tong, X. Min, and Y. Li, “Observer-based adaptive fuzzy tracking
control for strict-feedback nonlinear systems with unknown control
gain functions,” IEEE Trans. Cybern., vol. 50, no. 9, pp. 3903–3913,
Sep. 2020.
[9] Y. Li, Y. Liu, and S. Tong, “Observer-based neuro-adaptive optimized
control of strict-feedback nonlinear systems with state constraints,” IEEE
Trans. Neural Netw. Learn. Syst., early access, Jan. 26, 2021, doi:
10.1109/TNNLS.2021.3051030.
[10] S. Li, C. K. Ahn, M. Chadli, and Z. Xiang, “Sampled-data
adaptive fuzzy control of switched large-scale nonlinear delay
systems,” IEEE Trans. Fuzzy Syst., early access, Jan. 15, 2021, doi:
10.1109/TFUZZ.2021.3052094.
[11] J.-T. Huang and Y.-W. Jiang, “Robust composite adaptive control of
linearisable systems with improved performance,” IEEE Access, vol. 9,
Fig. 5. Adaptive parameters.
pp. 88037–88047, 2021, doi: 10.1109/ACCESS.2021.3090088.
[12] Y. Wu and Z. Wang, “Fuzzy adaptive practical fixed-time consensus for
second-order nonlinear multiagent systems under actuator faults,” IEEE
l2 = 1; σ1 = 0.1e−0.02t ; σ2 = 0.2e−0.02t ;. The target signal defined by Trans. Cybern., vol. 51, no. 3, pp. 1150–1162, Mar. 2021.
yr = 0.25 sin(t) and the initial conditions are assumed as ζ1 (0) = 0.5, [13] B. Chen, X. P. Liu, S. S. Ge, and C. Lin, “Adaptive fuzzy
control of a class of nonlinear systems by fuzzy approximation
ζ2 (0) = 0.5, θ̂1 (0) = 0.8, θ̂2 (0) = 0.8, ι̂1 (0) = 0.8, and ι̂2 (0) = 0.8. approach,” IEEE Trans. Fuzzy Syst., vol. 20, no. 6, pp. 1012–1021,
The simulation results based on Theorem 1 are shown in Figs. 1–5, Dec. 2012.
where Fig. 1 exhibits the trajectories of y and yr . It can be seen [14] S. Tong, Y. Li, and S. Sui, “Adaptive fuzzy tracking control design
that the curve of y remains in the prescribed range, while the curves for SISO uncertain nonstrict feedback nonlinear systems,” IEEE Trans.
Fuzzy Syst., vol. 24, no. 6, pp. 1441–1454, Dec. 2016.
of y and yr nearly coincide, suggesting better tracking performance. [15] Y. Sun, B. Chen, C. Lin, H. Wang, and S. Zhou, “Adaptive neural control
Fig. 2 exhibits the tracking errors y − yr , while Fig. 3 shows the for a class of stochastic nonlinear systems by backstepping approach,”
response of the controller u. Meanwhile, Fig. 4 depicts the trajectory Inf. Sci., vol. 369, pp. 748–764, Dec. 2016.

Authorized licensed use limited to: University of Ulsan. Downloaded on May 25,2025 at 13:54:44 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 34, NO. 10, OCTOBER 2023 8107

[16] J.-T. Huang, “Adaptive fuzzy state/output feedback control of nonstrict- [25] J. A. Farrell, M. Polycarpou, M. Sharma, and W. Dong, “Command
feedback systems: A direct compensation approach,” IEEE Trans. filtered backstepping,” IEEE Trans. Autom. Control, vol. 54, no. 6,
Cybern., vol. 49, no. 6, pp. 2046–2059, Jun. 2019. pp. 1391–1395, Jun. 2009.
[17] J.-T. Huang and Y.-C. Law, “Adaptive linearizing fuzzy control of [26] W. Dong, J. A. Farrell, M. M. Polycarpou, V. Djapic, and M. Sharma,
delayed MIMO nonstrict-feedback systems,” IEEE Trans. Fuzzy Syst., “Command filtered adaptive backstepping,” IEEE Trans. Control Syst.
vol. 28, no. 10, pp. 2592–2604, Oct. 2020. Technol., vol. 20, no. 3, pp. 566–580, May 2012.
[18] W. He, Y. Chen, and Z. Yin, “Adaptive neural network control of an [27] J. Yu, P. Shi, and L. Zhao, “Finite-time command filtered backstepping
uncertain robot with full-state constraints,” IEEE Trans. Cybern., vol. 46, control for a class of nonlinear systems,” Automatica, vol. 92,
no. 3, pp. 620–629, Apr. 2017. pp. 173–180, Jun. 2018.
[19] K. P. Tee, S. S. Ge, and E. H. Tay, “Barrier Lyapunov functions for the [28] Y. Li, “Command filter adaptive asymptotic tracking of uncertain
control of output-constrained nonlinear systems,” Automatica, vol. 45, nonlinear systems with time-varying parameters and disturbances,”
no. 4, pp. 918–927, Apr. 2009. IEEE Trans. Autom. Control, early access, Jun. 15, 2021, doi:
[20] K. P. Tee, B. Ren, and S. S. Ge, “Control of nonlinear systems with time- 10.1109/TAC.2021.3089626.
varying output constraints,” Automatica, vol. 47, no. 11, pp. 2511–2516, [29] H. Zhang, Y. Liu, J. Dai, and Y. Wang, “Command filter based adaptive
Nov. 2011. fuzzy finite-time control for a class of uncertain nonlinear systems with
[21] M. Lv, Y. Li, W. Pan, and S. Baldi, “Finite-time fuzzy adaptive hysteresis,” IEEE Trans. Fuzzy Syst., vol. 29, no. 9, pp. 2553–2564,
constrained tracking control for hypersonic flight vehicles with Sep. 2021.
singularity-free switching,” IEEE/ASME Trans. Mechtron., early access, [30] S. Li, C. K. Ahn, and Z. Xiang, “Command-filter-based adaptive fuzzy
Jun. 18, 2021, doi: 10.1109/TMECH.2021.3090509. finite-time control for switched nonlinear systems using state-dependent
[22] Y.-J. Liu, W. Zhao, L. Liu, D. Li, S. Tong, and C. L. P. Chen, “Adaptive switching method,” IEEE Trans. Fuzzy Syst., vol. 29, no. 4, pp. 833–845,
neural network control for a class of nonlinear systems with function Apr. 2021.
constraints on states,” IEEE Trans. Neural Netw. Learn. Syst., early [31] Y.-X. Li, “Finite time command filtered adaptive fault tolerant control
access, Sep. 14, 2021, doi: 10.1109/TNNLS.2021.3107600. for a class of uncertain nonlinear systems,” Automatica, vol. 106,
[23] K. Zhao and Y. Song, “Removing the feasibility conditions imposed on pp. 117–123, Aug. 2019.
tracking control designs for state-constrained strict-feedback systems,” [32] J. Qiu, K. Sun, I. J. Rudas, and H. Gao, “Command filter-based adaptive
IEEE Trans. Autom. Control, vol. 64, no. 3, pp. 1265–1272, Mar. 2019. NN control for MIMO nonlinear systems with full-state constraints and
[24] K. Zhao, Y. Song, and Z. Zhang, “Tracking control of MIMO nonlinear actuator hysteresis,” IEEE Trans. Cybern., vol. 50, no. 7, pp. 2905–2915,
systems under full state constraints: A single-parameter adaptation Jul. 2020.
approach free from feasibility conditions,” Automatica, vol. 107, [33] J. J. E. Slotine and W. Li, Applied Nonlinear Control. Englewood Cliffs,
pp. 52–60, Sep. 2019. NJ, USA: Prentice-Hall, 1991.

Authorized licensed use limited to: University of Ulsan. Downloaded on May 25,2025 at 13:54:44 UTC from IEEE Xplore. Restrictions apply.