Received: 6 December 2023 Revised: 12 April 2024 Accepted: 22 June 2024

DOI: 10.1002/oca.3173

RESEARCH ARTICLE

Observer-based optimal fault-tolerant tracking control for

input-constrained interconnected nonlinear systems with
mismatched disturbances

Shihui Liu1 Ning Xu2 Ning Zhao1 Liang Zhang1

1
College of Control Science and
Engineering, Bohai University, Liaoning, Summary
China This paper investigates an observer-based optimal fault-tolerant tracking con-
2
College of Information Science and trol problem for interconnected nonlinear systems with input constraints and
Technology, Bohai University, Liaoning,
mismatched disturbances via adaptive dynamic programming (ADP). Firstly, an
China
augmented system consisting of tracking error and the reference trajectory is
Correspondence constructed, and then the original optimal tracking control problem is trans-
Ning Xu, College of Information Science
and Technology, Bohai University,
formed into the optimal regulation control problem of the augmented system.
Jinzhou, Liaoning 121013, China. An integral sliding-mode-based optimal fault-tolerant control scheme is devel-
Email: [email protected] oped to eliminate the effect of actuator faults and guarantee the optimal control
Funding information performance. Furthermore, a neural network-based observer is designed to iden-
National Natural Science Foundation of tify the completely unknown system dynamics based on the input-output data,
China, Grant/Award Number: 62203064
thereby relaxing the restriction on system dynamics. Subsequently, the modi-
fied Hamilton-Jacobi-Bellman equations are solved by using the ADP algorithm
under a critic network framework. According to the Lyapunov approach, all sig-
nals in the closed-loop augmented system are uniformly ultimately bounded.
Finally, the effectiveness of the developed control scheme is demonstrated via
simulation results.

KEYWORDS
adaptive dynamic programming, decentralized tracking control, input constraints, integral
sliding-mode fault-tolerant control, interconnected nonlinear systems

1 I N T RO DU CT ION

Over the past few decades, research on interconnected systems has become an appealing subject. This is primarily because
interconnections are widely present in many practical systems, such as power systems,1 transportation systems,2 and robot
systems.3 Generally speaking, an interconnected system consists of multiple subsystems with evident interconnection
characteristics, which leads to an increasing challenge of designing stabilizing controllers using traditional centralized
control methods. Fortunately, the decentralized control method has been proposed, which partitions the control problem
of the entire system into a sequence of independently solvable subproblems. Recently, numerous remarkable research
achievements on decentralized control have been obtained. The authors in Reference 4 proposed an adaptive decentral-
ized output feedback control method for interconnected systems utilizing graph theory. In Reference 5, a decentralized
tracking control problem was investigated for a class of strong interconnected nonlinear systems under a backstepping
framework. Subsequently, the decentralized control method was applied to address interconnected nonlinear systems

2572 © 2024 John Wiley & Sons Ltd. wileyonlinelibrary.com/journal/oca Optim Control Appl Meth. 2024;45:2572–2595.
 10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU et al. 2573

with nonsmooth nonlinearities, such as dead-zone,6 saturation,7 hysteresis,8 and so forth. It is worth highlighting that
the above-mentioned control methods mainly concentrate on the stability of the considered systems. However, in practi-
cal scenarios, it is often necessary to minimize control costs while ensuring system performance, resulting in the optimal
control problem.
Optimal control, as an important branch of control theory, has received significant attention in recent years.9–11
According to optimal control theory, solutions of optimal control problems can typically be attributed to solving the
Hamilton-Jacobi-Bellman equations (HJBEs), which is a formidable challenge for nonlinear systems.12–15 Adaptive
dynamic programming (ADP) has successfully tackled this obstacle as an effective reinforcement learning algorithm by
approximating the optimal solution of an HJBE forward-in-time, and circumvents the curse of dimensionality caused by
the backward-in-time approach. Due to these advantages, the ADP algorithm has been widely applied in solving opti-
mal control problems, and a large number of outstanding results have been reported in theory and applications.16–18 As
shown in Reference 19, an ADP-based decentralized optimal control scheme was developed for interconnected nonlin-
ear systems under an actor-critic network framework. In Reference 20, an identifier-critic neural network was adopted
to solve the optimal control problem of partially unknown nonlinear systems. Xue et al.21 proposed an ADP algorithm
for completely unknown nonlinear systems under a critic framework, where a neural network-based identification tech-
nique was employed to release the dependence of system dynamics. However, it is worth noting that the above-mentioned
literatures primarily focus on designing optimal control strategies to drive the system states to a stabilization point.
In practical applications, the stability is merely a minimum standard, and it is commonly required to force the system
states to track a reference trajectory.22,23 In general, the existing solutions of ADP-based optimal tracking control problem
can be classified into two categories: In Reference 24, the tracking errors and reference trajectories were integrated to
construct an augmented system, and then the tracking control problem of the original system was transformed into a
stabilization problem of the augmented system. Another solution was presented in Reference 25, where a steady-state con-
troller and a feedback controller were designed to compensate undesired system dynamics and regulate tracking errors,
respectively. Subsequently, the authors in Reference 26 developed a novel optimal tracking control scheme for nonlinear
systems by combining these two approaches.
On the other hand, in engineering systems, actuators are prone to faults under prolonged operation, which can affect
system performances or even lead to instability.27–29 Therefore, it is necessary to explore a fault-tolerant control (FTC)
strategy to eliminate the effects of actuator faults. In the past several years, many excellent results on FTC have been pre-
sented for nonlinear systems with actuator faults.30–32 For example, an optimal FTC scheme was developed in Reference 33
for discrete-time systems under an actor-critic framework, where an auxiliary signal of the actuator fault was designed to
offset the effect of the fault. The authors in Reference 34 constructed a modified cost function to transform fault-tolerant
control into optimal control, then the solution of the modified HJBE was approximated by using a single-critic neural
network. In Reference 35, an improved discounted cost function associated with estimated actuator faults was pro-
posed to address the optimal FTC problem for interconnected nonlinear systems with input constraints. Compared with
the aforementioned approaches, sliding-mode control (SMC) holds an advantage in withstanding matched uncertain-
ties and disturbances. Based on this benefit, SMC has obtained extensive applications in fault-tolerant control design.
In Reference 36, a linear sliding-mode surface-based optimal FTC scheme was developed for nonlinear systems with
parameter perturbations. The authors in Reference 37 proposed a novel integral sliding mode fault-tolerant control
method for a class of affine nonlinear systems based on an ADP algorithm. However, few scholars have employed the inte-
gral sliding-mode FTC approach to address the optimal tracking control problem for unknown interconnected nonlinear
systems with input constraints, which motivates our research interests. Furthermore, the mismatched disturbances are
often unavoidable in practical engineering, which poses a potential risk for controlled systems. In this article, the impact
of mismatched disturbances is taken into account, thereby enhancing the feasibility of the developed control scheme.
Compared with the previous works, the main contributions can be described as follows:

1. An observer-based integral sliding-mode fault-tolerant tracking control scheme for interconnected nonlinear systems
is for the first time developed. The developed control scheme can effectively withstand the actuator faults of each
subsystem while achieving optimal tracking control with the minimum cost.
2. For sliding-mode dynamics with mismatched disturbances, unlike References 38,39, the control method developed in
this paper no longer needs to seek a saddle point solution but only needs to include a compensation term in the cost
function. Furthermore, a modified nonquadratic function is introduced to incorporate the input constraints into the
optimal control, further obtaining the modified HJBEs.
 10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2574 LIU et al.

3. Different from References 35,37, a neural network-based observer is constructed to identify the completely unknown
system dynamics via input and output data, such that the dependence of internal system dynamics is released. Then,
a single-critic network with a simpler structure compared with Reference 40 is employed to approximate the solution
of the modified HJBEs, where critic network weights are updated via a gradient descent methodology.

Notations: N denotes the set of nonnegative integers, Rn×m is the set of n × m real matrices and Rn represents the
Euclidean space of n-dimensional real vectors; ∇(⋅) is the gradient operator and −1 denotes the inverse of ; ||||
denotes the Euclidean norm of , let ⊤ represent the transpose operator and 𝜆min () denotes the minimum eigenvalue
of ; In is the n × n identity matrix, Ω represents a compact set including the origin and tr(⋅) denotes the trace operation.

2 PROBLEM STAT EMENT

Consider the completely unknown nonlinear interconnected system as

[ ]
ẋ i (t) = foi (xi (t)) + goi (xi (t)) ui (t) + Θ(t − tT )Δui + Zi (x) + hi (xi (t))𝜗i (t), (1)

where i = 1, 2, … , N, xi (t) ∈ Rni is the ith subsystem state with xi (0) = xi0 , foi (xi ) ∈ Rni , goi (xi ) ∈ Rni ×mi and hi (xi ) ∈ Rni ×qi
represent unknown Lipschitz functions with foi (0) = 0. ui (t) ∈ Rmi , Θ(t − tT )Δui ∈ Rmi , Zi (x) ∈ Rmi and 𝜗i (t) ∈ Rqi rep-
resent the control input, the actuator fault, the interconnected item and the disturbance input in sequence. Moreover,
suppose that the controlled system (1) is controllable.
Considering the nonlinear system (1), we define a reference system to investigate the trajectory tracking control
problem, which is shown as

ṙ i (t) = 𝜁i (ri (t)), (2)

where ri (t) ∈ Rni represents the desired trajectory with ri (0) = ri0 while the function 𝜁i is locally Lipschitz satisfying 𝜁i (0) =
0. The ith subsystem tracking error is defined as eti (t) = xi (t) − ri (t) with eti (0) = eti0 . Thus, the tracking error dynamic can
be depicted as
[ ]
ė ti (t) = foi (xi ) + goi (xi ) ui (t) + Θ(t − tT )Δui + Zi (x) + hi (xi )𝜗i (t) − 𝜁i (ri (t)). (3)

Noting xi (t) = eti (t) + ri (t), (3) becomes

[ ]
ė ti (t) = foi (eti (t) + ri (t)) + goi (eti (t) + ri (t)) ui (t) + Θ(t − tT )Δui + Zi (x)
(4)
+ hi (eti (t) + ri (t))𝜗i (t) − 𝜁i (ri (t)).

The augmented system state can be defined as

[ ]⊤
ℑi (t) = e⊤ti (t), ri⊤ (t) ∈ R2ni , (5)

[ ⊤ ⊤
]
where ℑi (0) = ℑi0 = e⊤ti0 , ri0 is the initial state.
Observing (2), (4), and (5), the augmented system dynamic is expressed as
[ ]
ℑ̇ i (t) = i (ℑi (t)) + Gi (ℑi (t)) ui (t) + Θ(t − tT )Δui + Zi (ℑ) + Hi (ℑi (t))𝜗i (t), (6)

where
[ ]⊤
i (ℑi (t)) = foi (eti (t) + ri (t)) − 𝜁i (ri (t)), 𝜁i (ri (t)) ∈ R2ni ,
[ ]⊤
Gi (ℑi (t)) = goi (eti (t) + ri (t)), 0ni ×mi ∈ R2ni ×mi ,
[ ]⊤
Hi (ℑi (t)) = hi (eti (t) + ri (t)), 0ni ×qi ∈ R2ni ×qi .
 10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU et al. 2575

Assumption 1 (39). Let Gi,m , H i,m and 𝜗i,m be positive constants. Gi (ℑi ) with Gi (0) ≠ 0, Hi (ℑi ) and 𝜗i (t) are
upper bounded, that is, ‖Gi (ℑi )‖ ≤ Gi,m , ‖Hi (ℑi )‖ ≤ H i,m and ‖𝜗i ‖ ≤ 𝜗i,m for every ℑi ∈ Ω. Furthermore, the
equilibrium point of the augmented system (6) is set at ℑi = 0.
Assumption 2 (41). ∀i, j = 1, 2, … , N, Zi (ℑ) is bounded as

∑
N
( )
‖Zi (ℑ)‖ ≤ aij Pij ℑj , (7)
j=1

( )
where aij ≥ 0 and Pij ℑj ∈ R are nonnegative constant and positive-definite funnction, respectively. More-
over, Zi (0) = 0 and Pij (0) = 0. Let Pi (ℑi ) = max {P1i (ℑi ), P2i (ℑi ), … , PNi (ℑi )}. Subsequently, (7) can be further
rewritten as
∑
N
( )
‖Zi (ℑ)‖ ≤ bij Pj ℑj (8)
j=1

( ) ( )
with bij ≥ aij Pij ℑj ∕Pj ℑj being also the nonnegative constant.
( ) ( )−1
Assumption 3 (42). Let dui = Ii − Gi G+i Hi 𝜗i and the funcion G+i (ℑi ) = G⊤i (ℑi )Gi (ℑi ) G⊤i (ℑi ) is
‖ ‖ ‖ + ‖
bounded by ‖G+i (ℑi )‖ ≤ d+G with d+G > 0. Furthermore, ‖ ‖
‖Θ(t − tT )Δui ‖ ≤ Δui and ‖ G H 𝜗 ‖ ≤ 𝓁0,i with Δui >
‖ ‖ i i ‖ i i i‖
0 and 𝓁0,i > 0.
‖ ‖
Assumption 4 (38). There exist some positive constants dVi , D𝜀u∗ , 𝜎M,i and d𝜀ci meeting ‖∇Vi∗ ‖ ≤ dVi ,
0,i ‖ ‖
‖ ∗‖
‖𝜀u0,i ‖ ≤ D𝜀u∗ , ‖∇𝜎ci ‖ ≤ 𝜎M,i and ‖∇𝜀ci ‖ ≤ d𝜀ci .
‖ ‖ 0,i

3 DECENTRALIZED TRAC KING CO NTRO LLER DESIGN BY MEANS OF

OPTIMAL FAU LT-TOLERANT CO NTRO L SCHEME

The objective of this paper is to develop a decentralized optimal control strategy for the augmented system (6) with actu-
ator faults and mismatched disturbances based on Assumptions 1 and 2, thereby ensuring that the entire closed-loop
augmented system is uniformly ultimately bounded (UUB). However, it is typically challenging to directly design such
a decentralized optimal control strategy. Inspired by References 41 and 43, we can partition the decentralized control
problem into N optimal control problems of isolated subsystems associated with the augmented system (6), such that the
state xi (t) of the controlled system (1) tracks the reference trajectory ri (t) as precise as possible. The isolated augmented
subsystem of (6) is given by
[ ]
ℑ̇ i (t) = i (ℑi (t)) + Gi (ℑi (t)) ui (t) + Θ(t − tT )Δui + Hi (ℑi (t))𝜗i (t). (9)
Remark 1. For ith isolated augmented subsystem, the composite control strategy ui consists of uF,i and u0,i that
is, ui = uF,i + u0,i , where uF,i is the SMC strategy to eliminate the affect of actuator faults, and u0,i is the con-
tinuous control strategy to maintain sliding-mode stability while ensuring an optimal control performance.
uF,i and u0,i collaborate to achieve tracking control for interconnected system (1) with actuator faults.

3.1 Integral sliding-mode FTC design

Here, an integral SMC method is proposed to derive the control strategy uF,i . Furthermore, the sliding mode surface can
be described as

( ) t ( )
si (ℑi ) = sfi (ℑi ) − sfi ℑ0,i − Ti (ℑi ) i (ℑi ) + Gi (ℑi )u0,i d𝜏, (10)
∫0

where sfi (ℑi ) ∈ Rmi and Ti (ℑi ) = 𝜕sfi (ℑi )∕𝜕ℑi .

 10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2576 LIU et al.

Then, the sliding mode dynamic can be derived as

( )
ṡ i = Ti (ℑi )ℑ̇ i − Ti (ℑi ) i (ℑi ) + Gi (ℑi )u0,i
( ) (11)
= Ti Gi uF,i + Gi Θ(t − tT )Δui + Hi 𝜗i .

From Assumption 3, it is obtained that Hi 𝜗i = Gi G+i Hi 𝜗i + dui and we assume that Ti Gi is invertible, then (11) can be
written as:
( )
ṡ i = Ti Gi uF,i + Gi Θ(t − tT )Δui + G+i Hi 𝜗i + Ti dui
( ) (12)
= Ti Gi uF,i + Gi Θ(t − tT )Δui + G+i Hi 𝜗i + (Ti Gi )−1 Ti dui .

To ensure the augmented subsystem states on integral sliding manifold si (ℑi ) = 0, the SMC strategy is
designed as:
( )
uF,i = −Gfi sgn G⊤i Ti⊤ si , (13)

where Gfi denotes the gain and sgn(⋅) stands for the sign function.
Theorem 1. For the ith augmented subsystem (9) with sliding manifold given by (10). Then, uF,i can drive the
system trajectories to reach the sliding mode surface if Gfi > Δui + 𝓁0,i holds.

Proof. Choose a Lyapunov function candidate as:

∑
N
Lsi = Vsi , (14)
i=1

where Vsi = (1∕2)s⊤i si . Based on (13) and assuming Ti = G+i ,42 one gets:
[ ]
V̇ si = s⊤i Ti Gi uF,i + Θ(t − tT )Δui + G+i Hi 𝜗i + (Ti Gi )−1 Ti dui
[ ( ) ]
≤ −s⊤i Ti Gi −Gfi sgn G⊤i Ti⊤ si + Θ(t − tT )Δui + G+i Hi 𝜗i + (Ti Gi )−1 Ti dui
‖ ‖ ‖ ‖ ‖ ‖ (15)
≤ −Gfi ‖s⊤i Ti Gi ‖ + ‖s⊤i Ti Gi ‖ ∗ ‖Θ(t − tT )Δui + G+i Hi 𝜗i + (Ti Gi )−1 Ti dui ‖
‖ ‖ ‖ ‖ ‖ ‖
‖ ⊤‖ ‖ ⊤‖
≤ −Gfi ‖si ‖ + (Δui + 𝓁0,i )‖si ‖.
‖ ‖ ‖ ‖
∑N
Then, we can get L̇ si = ̇
i=1 V si < 0 based on Theorem 1. The proof is concluded. ▪

According to the principle of equivalent control, the equivalent controller is obtained as

ufi = −Θ(t − tT )Δui − G+i Hi 𝜗i − (Ti Gi )−1 Ti dui . (16)

Substituting ufi into (9), we can get

ℑ̇ i = i + Gi u0,i − Gi G+i Hi 𝜗i − Gi (Ti Gi )−1 Ti dui + Hi 𝜗i

= i + Gi u0,i − Gi (Ti Gi )−1 Ti dui + dui
[ ( )−1 ] (17)
= i + Gi u0,i + Ii − Gi G+i Gi G+i dui
( )
= i + Gi u0,i + Ii − Gi G+i Hi 𝜗i .
( )
Letting IGi = Ii − Gi G+i Hi and assuming ‖ ‖
‖IGi ‖ ≤ ĪGi with ĪGi > 0, we have equivalent sliding mode dynamic:

ℑ̇ i (t) = i (ℑi (t)) + Gi (ℑi (t))u0,i + IGi 𝜗i (t). (18)

 10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU et al. 2577

3.2 Decentralized optimal control design

In this subsection, continuous optimal control strategy are designed to guarantee optimal control performance. The
dynamic (18) is converted into the following auxiliary system

ℑ̇ i (t) = i (ℑi (t)) + Gi (ℑi (t))u0,i + IGi ai . (19)

Based on considerations of input constraints and mismatched disturbances, a modified cost function is defined
as follows:
∞[ ( ) ]
2
Vi (ℑi (t)) = qai 𝜗i,m + ℑ⊤i Qi ℑi + Ui u0,i + a⊤i Qai ai + 𝜂i Pi2 (ℑi ) , (20)
∫t

2
where ai ∈ Rqi is the auxiliary control, qai 𝜗i,m represents the compensation term related to the disturbances, qai
( )
denotes the maximum eigenvalue of the matrix Qai with Qai ∈ Rqi ×qi being a positive definite matrix. Ui u0,i =
u ( )⊤
2𝜆i ∫0 0,i tanh−1 (𝛿i ∕𝜆i ) Ri d𝛿i represents the nonquadratic cost function with constrained bound 𝜆i ∈ Rmi , designed
[ ]
parameters 𝛿i ∈ Rmi and Ri ∈ Rmi ×mi . In this paper, Ri is regarded as an identity matrix and Qi = Qi , 0ni ×ni ; 0ni ×ni , 0ni ×ni
with Qi being a symmetric positive definite matrix.39
The Hamiltonian function associated with (20) is derived as
( )
Hi ℑi , u0,i , ai , ∇Vi
( )
= ∇Vi⊤ i (ℑi (t)) + Gi (ℑi (t))u0,i + IGi ai (21)
2 ( )
+ qai 𝜗i,m + ℑ⊤i Qi ℑi + Ui u0,i + a⊤i Qai ai + 𝜂i Pi2 (ℑi ),

where ∇Vi = 𝜕Vi ∕𝜕ℑi .

Substituting the optimal control law u∗0,i and optimal auxiliary control law a∗i into (20) yields
∞[ ( ) ]
2
Vi∗ = qai 𝜗i,m + ℑ⊤i Qi ℑi + Ui u∗0,i + a∗⊤
i Qai ai + 𝜂i Pi (ℑi ) dt.
∗ 2
(22)
∫t

Then, we can obtain the HJBE as follows:

( )
Hi ℑi , u∗0,i , a∗i , ∇Vi∗
( )
= ∇Vi∗⊤ i (ℑi (t)) + Gi (ℑi (t))u∗0,i + IGi a∗i (23)
2
( )
+ qai 𝜗i,m + ℑ⊤i Qi ℑi + Ui u∗0,i + a∗⊤ i Qai ai + 𝜂i Pi (ℑi ) = 0.
∗ 2

The optimal solutions u∗0,i and a∗i satisfy the following two fixed conditions:
( )
𝜕Hi ℑi , u∗0,i , a∗i , ∇Vi∗
= 0, (24)
𝜕u∗0,i
( )
𝜕Hi ℑi , u∗0,i , a∗i , ∇Vi∗
= 0. (25)
𝜕a∗i

According to (23), we can solve Equations (24) and (25) yields

u∗0,i = −𝜆i tanh(B∗ ), (26)

1
a∗i = − Q−1 I ⊤ ∇V ∗ , (27)
2 ai Gi i
 10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2578 LIU et al.

where B∗ = (1∕2𝜆i )R−1

i
G⊤i (ℑi )∇Vi∗ and Q−1
ai
is bounded by the positive constant qai .
Based on (26) and (27), (23) can be rewritten as
( )
Hi ℑi , u∗0,i , a∗i , ∇Vi∗
= ∇Vi∗⊤ i (ℑi (t)) − ∇Vi∗⊤ Gi (ℑi (t))𝜆i tanh(B∗ )
1 ⊤ 2 ⊤ (28)
− ∇Vi∗⊤ IGi Q−1 ai IGi ∇Vi + qai 𝜗i,m + ℑi Qi ℑi
∗
4
−𝜆i tanh(B∗ ) ( )⊤
+ 2𝜆i tanh−1 (𝛿i ∕𝜆i ) Ri d𝛿i + 𝜂i Pi2 (ℑi ) = 0
∫0
Theorem 2. Consider N augmented subsystems (9) with actuator faults and external disturbances. If Assump-
tions1, 2, 3, and 4 are satisfied, then we can find a sequence of positive constants 𝜂i∗ , such that, for each 𝜂i ≥ 𝜂i∗ ,
i ∈ 1, … , N, the optimal composite control strategies u1 , u2 , … , uN make sure that the augmented system (6)
can maintain UUB.

Proof. Consider a Lyapunov function candidate as:

∑
N
Li,0 (t) = Vi∗ (ℑi (t)). (29)
i=1

The derivative of Vi∗ can be expressed as

( )⊤ ( [ ] )
V̇ i = ∇Vi∗
∗
i (ℑi (t)) + Gi (ℑi (t)) ui (t) + Θ(t − tT )Δui + Zi + Hi (ℑi (t))𝜗i (t)
( )⊤ ( [ ] (30)
= ∇Vi∗ i (ℑi (t)) + Gi (ℑi (t)) ui (t) + Θ(t − tT )Δui +Gi (ℑi (t))Zi + Hi (ℑi (t))𝜗i (t)).

Substituting u∗0,i and equivalent sliding mode control strategy (16) into (30), we have:

( )⊤ ( )
V̇ i = ∇Vi∗
∗
i (ℑi (t)) + Gi (ℑi (t))u∗0,i + IGi 𝜗i (t) + Gi (ℑi )Zi . (31)

By observing (23), (26) and (27), we can get the following facts:

( )⊤ 2
( )
∇Vi∗ i (ℑi ) = −qai 𝜗i,m − ℑ⊤i Qi ℑi − Ui u∗0,i − a∗⊤ ∗
i Qai ai
( )⊤ ( )⊤
− 𝜂i Pi2 (ℑi ) − ∇Vi∗ Gi u∗0,i − ∇Vi∗ IGi a∗i
( ∗ )
( ∗ )⊤ u0,i (32)
−⊤
∇Vi Gi (ℑi ) = −2𝜆i tanh
𝜆i
( ∗ )⊤
∇Vi IGi = −2a∗⊤ i Qai .

Substituting (32) into (31), we can obtain

N [ ( )
∑ 2
L̇ i,0 (t) = −qai 𝜗i,m − ℑ⊤i Qi ℑi − Ui u∗0,i − a∗⊤
i Qai ai − 𝜂i Pi (ℑi )
∗ 2

i=1
( )⊤ ( )⊤ ( )⊤ ]
− ∇Vi∗ IGi a∗i + ∇Vi∗ IGi 𝜗i (t) + ∇Vi∗ Gi (ℑi )Zi
(33)
N [ ( )
∑ 2 ( )⊤
= −qai 𝜗i,m − ℑ⊤i Qi ℑi − Ui u∗0,i − 1∕2 ∇Vi∗ IGi a∗i
i=1
( )⊤ ( )⊤ ]
−𝜂i Pi2 (ℑi ) + ∇Vi∗ IGi 𝜗i + ∇Vi∗ Gi (ℑi )Zi .
 10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU et al. 2579

( )
According to the fact that Ui u∗0,i > 0 and Assumption 4, one has

N [
∑ 2 ( )⊤
L̇ i,0 (t) ≤ −ℑ⊤i Qi ℑi − qai 𝜗i,m − 𝜂i Pi2 (ℑi ) + ∇Vi∗ Gi (ℑi )Zi
i=1 (34)
]
1 ‖( )⊤ ‖2 2
+ d2Vi ĪG2 qai + dVi ĪGi 𝜗i,m − ‖ ∇Vi∗ Gi (ℑi )‖ + d2Vi Gi,m .
4 i ‖ ‖

From Assumption 2, it follows that

( )⊤ ‖( )⊤ ‖
∇Vi∗ Gi (ℑi )Zi ≤ 2‖ ∇Vi∗ Gi (ℑi )‖‖Zi ‖
‖ ‖
‖∑
N
‖( )⊤ ( ) (35)
≤ 2‖ ∇Vi∗ Gi (ℑi )‖ bij Pj ℑj .
‖ ‖
j=1

According to (5), we derive that ℑ⊤i Qi ℑi = e⊤ti Qi eti and combining (34) and (35) yields

N [ ( ) ]
∑
L̇ i,0 (t) ≤ −𝜆min Qi ‖eti ‖2 + Δi − ℘ , (36)
i=1

where

1 2 ̄2 2
Δi = dVi IG qai + dVi ĪGi 𝜗i,m + d2Vi Gi,m
4 i

and

‖∑
N
‖( )⊤ ‖2 ‖( )⊤ ( )
℘ = 𝜂i Pi2 (ℑi ) + ‖ ∇Vi∗ Gi (ℑi )‖ − 2‖ ∇Vi∗ Gi (ℑi )‖ bij Pj ℑj .
‖ ‖ ‖ ‖
j=1

Meanwhile, we let

𝜂̃ = diag{𝜂1 , 𝜂2 , … , 𝜂N }
1̃ = diag{11 , 12 , · · · , 1N }(1i = 1, i ∈ 1, … , N)
[ ]⊤ (37)
Y (ℑ) = −P1 (ℑ1 ), −P2 (ℑ2 ), … , −PN (ℑN ), Γ1 , Γ2 , … , ΓN
‖( )⊤ ‖
Γi = ‖ ∇Vi∗ Gi (ℑi )‖(i ∈ 1, … , N).
‖ ‖

Then, combining (37), (36) can further be derived as

N [ ( ) ]
∑
L̇ i,0 (t) ≤ −Y ⊤ (ℑ)Y (ℑ) − 𝜆min Qi ‖eti ‖2 + Δi , (38)
i=1

[ ] [ ]
b11 b12 ··· b1N
𝜂̃ ⊤
where  = with  = ⋮ ⋮ ⋱ ⋮ . It can be seen from  and  that 𝜂̃ is distributed
 1̃ b bN2 · · · bNN
N1
on the principal diagonal of the block matrix  ∈ R2N×2N . Thus, we can make  positive definite by choosing
appropriate 𝜂i . Put differently, we can find N positive constants 𝜂i∗ , such that 𝜂i ≥ 𝜂i∗ , i ∈ 1, … , N, ensures the
validity of −Y ⊤ (ℑ)Y (ℑ) < 0. Then, (38) further yields

N [ ( ) ]
∑
L̇ i,0 (t) ≤ −𝜆min Qi ‖eti ‖2 + Δi . (39)
i=1
 10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2580 LIU et al.

According to (39), we have L̇ i,0 (t) < 0 only if, for each i ∈ 1, … , N, the tracking error eti is out of the set

⎧ √ ⎫
⎪ √ ⎪
√
Ωeti = ⎨eti ∶ ‖eti ‖ ≤ √
Δi
( ) ⎬. (40)
⎪ 𝜆min Qi ⎪
⎩ ⎭

Based on the Lyapunov extension theorem, this proves that the interconnected augmented system (6) is
UUB. This completes the proof. ▪

4 NEURAL NET WO RK-BASED OBSERVER DESIGN

In this section, the input-output data are utilized to construct a neural network (NN) observer, such that the unknown
system can be reconstructed. The approximated system dynamics are as follows:

i (ℑi ) = w⊤ i 𝜎 i (ℑi ) + 𝜀 i (t)

Gi (ℑi )uIi = w⊤Gi 𝜎Gi (ℑi , uIi ) + 𝜀Gi (t) (41)
Hi (ℑi )𝜗i = w⊤Hi 𝜎Hi (ℑi , 𝜗i ) + 𝜀Hi (t),

where w i ∈ Rl i ×2ni , wGi ∈ RlGi ×2ni and wHi ∈ RlHi ×2ni are the ideal weights, and l i , lGi and lHi are the number of neurons
in hidden layers. 𝜀 i (t) ∈ R2ni , 𝜀Gi (t) ∈ R2ni and 𝜀Hi (t) ∈ R2ni are approximation errors. 𝜎 i (ℑi ) ∈ Rl i , 𝜎Gi (ℑi , uIi ) ∈ RlGi
and 𝜎Hi (ℑi , 𝜗i ) ∈ RlHi are activation functions. Furthermore, uIi = ui (t) + Θ(t − tT )Δui .
̂  i ∈ Rl i ×2ni , w
Let w ̂ Gi ∈ RlGi ×2ni and w
̂ Hi ∈ RlHi ×2ni be the estimated weights. We get
( ) ( )
̂ i ℑ ̂i =w ̂ ⊤ i 𝜎 i ℑ̂i
( ) ( )
̂i ℑ
G ̂ i uIi = w
̂ ⊤Gi 𝜎Gi ℑ ̂ i , uIi (42)
( ) ( )
̂i ℑ
H ̂ i 𝜗i = w
̂ ⊤Hi 𝜎Hi ℑ ̂ i , 𝜗i ,

̂ i represents the estimated state. Thus, the unknown interconnected augmented subsystem can be observed as
where ℑ
( ) ( ) ( ) ( )
̂̇ i (t) = ̂ i ℑ
ℑ ̂i +G ̂ i uIi + H
̂i ℑ ̂ i 𝜗i + Goi ℑi − ℑ
̂i ℑ ̂i , (43)

̂ i represents observation error. Considering (9) and (43), the error

where Goi is a positive definite matrix and eoi = ℑi − ℑ
dynamic can be obtained as
( ) ( ( )) ( ( ))
̂ i + Gi (ℑi ) − G
ė oi = i (ℑi ) − ̂ i ℑ ̂ i uIi + Hi (ℑi ) − H
̂i ℑ ̂ i 𝜗i − Goi eoi .
̂i ℑ (44)

Combining (41) and (42), we have

( ) ( )
i (ℑi ) − ̂ i ℑ̂ i = w⊤ 𝜎̃  i + w
̃ ⊤
𝜎  i ℑ ̂ i + 𝜀 i (t)
i i
( ( )) ( )
̂i ℑ
Gi (ℑi ) − G ̂ i uIi = w⊤ 𝜎̃ Gi + w ̃ ⊤Gi 𝜎Gi ℑ ̂ i , uIi + 𝜀Gi (t) (45)
Gi
( ( )) ( )
Hi (ℑi ) − Ĥi ℑ ̂ i 𝜗i = w⊤ 𝜎̃ Hi + w ̃ ⊤
𝜎 Hi ℑ ̂ i , 𝜗i + 𝜀Hi (t),
Hi Hi

where
( )w̃  i = w i − w
̂  i, w
̃ Gi = w(Gi − ŵ Gi) and w
̃ Hi = wHi − w
̂ Hi are the ( weight) estimation errors, and 𝜎̃  i = 𝜎 i (ℑi ) −
̂ ̂ ̂
𝜎 i ℑi , 𝜎̃ Gi = 𝜎Gi (ℑi , uIi ) − 𝜎Gi ℑi , uIi and 𝜎̃ Hi = 𝜎Hi (ℑi , 𝜗i ) − 𝜎Hi ℑi , 𝜗i are the estimation errors of activation
functions.
 10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU et al. 2581

Theorem 3. Considering the interconnected augmented subsystem (9), the NN observer presented in (43) can
̂  i, w
ensure that the observation error eoi to be UUB, if w ̂ Gi and w
̂ Hi are updated with the following rules
( )
̂̇ ℱ i (t) = 𝛼ℱ i 𝜎ℱ i ℑ
w ̂ i e⊤
oi
( )
ŵ̇ Gi (t) = 𝛼Gi 𝜎Gi ℑ̂ i , uIi e⊤
oi (46)
( )
ŵ̇ Hi (t) = 𝛼Hi 𝜎Hi ℑ̂ i , 𝜗 i e⊤
oi

where 𝛼 i , 𝛼Gi and 𝛼Hi are positive NN learning rates.

Proof. Select a Lyapunov function candidate as Ls (t) = Ls1 (t) + Ls2 (t) with

1∑ ⊤
N
Ls1 (t) = e eoi (47)
2 i=1 oi

and
N [ { } { } { }]
1∑ 1 ⊤ 1 ⊤ 1 ⊤
Ls2 (t) = tr ̃  iw
w ̃  i + tr ̃ Gi w
w ̃ Gi + tr ̃ Hi w
w ̃ Hi (48)
2 i=1 𝛼 i 𝛼Gi 𝛼Hi

Differentiating Ls1 (t) with respect to t, and combining (44) and (45), we can obtain

N [ ( ( ) ( ) ( ) ) ]
∑
L̇ s1 (t) = e⊤oi w ̂i +w
̃ ⊤ i 𝜎 i ℑ ̃ ⊤Gi 𝜎Gi ℑ̂ i , uIi + w
̃ ⊤Hi 𝜎Hi ℑ̂ i , 𝜗i + Θi − e⊤ Goi eoi , (49)
oi
i=1

where Θi = w⊤ i 𝜎̃  i + 𝜀 i (t) + w⊤Gi 𝜎̃ Gi + 𝜀Gi (t) + w⊤Hi 𝜎̃ Hi + 𝜀Hi (t). Differentiating Ls2 (t) with respect to t, and
considering (46), we get

∑
N { ( ) } ∑ N { ( ) }
L̇ s2 (t) = − tr w ̂ i e⊤ −
̃ ⊤ i (t)𝜎 i ℑ tr w̃ ⊤
(t)𝜎 Gi ℑ̂ i , uIi e⊤
oi Gi oi
i=1 i=1
(50)
∑
N { ( ) }
− tr w̃ ⊤Hi (t)𝜎Hi ℑ̂ i , 𝜗 i e⊤ .
oi
i=1

According to tr{ℏ𝜆} = tr{𝜆ℏ} = 𝜆ℏ for 𝜆ℏ ∈ R, we can derive

∑
N ( ) ∑ N ( )
L̇ s2 (t) = − e⊤oi w ̂i −
̃ ⊤ i (t)𝜎 i ℑ e⊤oi w
̃ ⊤Gi (t)𝜎Gi ℑ̂ i , uIi
i=1 i=1
(51)
∑N ( )
− e⊤oi w
̃ ⊤Hi (t)𝜎Hi ℑ̂ i , 𝜗i .
i=1

Combining (49) with (51), we can get

∑
N
[ ⊤ ]
L̇ s (t) = eoi Θi − e⊤oi Goi eoi . (52)
i=1

Assuming that Θi ≤ Θi , Θi > 0,44 we have

N [ ( )]
∑
L̇ s (t) ≤ −‖eoi ‖ 𝜆min (Goi )‖eoi ‖ − Θi , (53)
i=1
 10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2582 LIU et al.

where 𝜆min (Goi ) is the minimum eigenvalue of Goi . Therefore, L̇ s (t) ≤ 0 when the following condition
is violated
{ }
Θi
Ωeoi = eoi ∶ ‖eoi ‖ ≤ . (54)
𝜆min (Goi )

Then, the observation error eoi can be guaranteed to be UUB. This ends the proof. ▪

5 CRITIC NN DESIGN

By employing the NN approximation theorem, Vi∗ is approximated as:

Vi∗ = Wci⊤ 𝜎ci (ℑi ) + 𝜀ci , (55)

where Wci ∈ RKci , 𝜎ci (ℑi ) ∈ RKci and 𝜀ci represent the ideal weight, the activation function and the approximate error,
respectively.
Further, we can deduce that

∇Vi∗ = ∇𝜎ci⊤ Wci + ∇𝜀ci , (56)

where ∇𝜎ci = 𝜕𝜎ci ∕𝜕ℑi and ∇𝜀ci = 𝜕𝜀ci ∕𝜕ℑi .

Substituting (56) into (26) and (27) yields

u∗0,i = −𝜆i tanh(C∗ ) + 𝜀u∗0,i (57)

and

1 ⊤
( ⊤ )
ai IGi ∇𝜎ci Wci + ∇𝜀ci ,
a∗i = − Q−1 (58)
2
( )
where C∗ = (1∕2𝜆i )G⊤i (ℑi )∇𝜎ci⊤ Wci , 𝜀u∗0,i = −(1∕2) Ii − tanh2 (Λ) G⊤i (ℑi )∇𝜀ci with Ii = [1, 1, … , 1]⊤ ∈ Rmi . The value of Λ
( )
is selected between C∗ and (1∕2𝜆i )G⊤i (ℑi ) ∇𝜎ci⊤ Wci + ∇𝜀ci .
Since Wci in (57) and (58) is unknown, u∗0,i and a∗i cannot be obtained directly. Here, an estimation of the critic weight
is introduced so that the value function Vi∗ can be approximated as

̂i = W
V ̂⊤ 𝜎ci (ℑi ), (59)
ci

̂ci represents the estimation of Wci and W

where W ̃ci = Wci − W
̂ci is the estimation error. Similarly, we can obtain

̂i = ∇𝜎 ⊤ W
̂
∇V ci ci . (60)

Substituting (60) into (26) and (27) yields

̂ ̂
u0,i = −𝜆i tanh(C) (61)

and

1 ̂ci ,
â i = − Q−1 I ⊤ ∇𝜎 ⊤ W (62)
2 ai Gi ci

̂ = (1∕2𝜆i )G⊤ (ℑi )∇𝜎 ⊤ W

where C ̂ .
i ci ci
 10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU et al. 2583

The approximate HJBE is given by:

( ) ( )
u0,i , ̂
Hi ℑi , ̂ ̂i = W
ai , ∇V ̂⊤ ∇𝜎ci i (ℑi (t)) + Gi (ℑi (t))̂
u0,i + IGi ̂
ai
ci
( ) (63)
2
+ qai 𝜗i,m + ℑ⊤i Qi ℑi + Ui ̂
u0,i + ̂ a⊤i Qai ̂
ai + 𝜂i Pi2 (ℑi ).

By combining (23) and (63), the residual error can be obtained as

( ) ( )
u0,i , ̂
𝜃ci = Hi ℑi , ̂ ̂i − Hi ℑi , u∗ , a∗ , ∇V ∗
ai , ∇V 0,i i i
( ) (64)
u0,i , ̂
= Hi ℑi , ̂ ̂i − 0.
ai , ∇V

In order to make the error 𝜃ci small enough, the following critic weight updating law is given to minimize the objection
function Ei = (1∕2)𝜃ci⊤ 𝜃ci :

̇
̂ lci 𝜕Ei
W ci = − ( )2
1+ 𝜙⊤i 𝜙i 𝜕Ŵci
lci 𝜙i [ 2 ( ) ] (65)
= −( )2 Ŵ⊤ 𝜙i + qai 𝜗i,m + ℑ⊤ Qi ℑi + Ui û 0,i + â ⊤i Qai â i + 𝜂i P2 (ℑi ) ,
ci i i
1 + 𝜙⊤i 𝜙i
( )
u0,i + IGi ̂
where lci is the critic NN learning rate and 𝜙i = ∇𝜎ci i (ℑi (t)) + Gi (ℑi (t))̂ ai .
By using the definition of W ̃ci , one gets

lci 𝜙i ( )
̇
̃ ⊤̃
W ci = − ( )2 𝜙i Wci − 𝜃cHi , (66)
1 + 𝜙⊤i 𝜙i
( )
where 𝜃cHi = −(∇𝜀ci )⊤ i (ℑi (t)) + Gi (ℑi (t))̂
u0,i + IGi ̂
ai .
Remark 2. A single-critic network is adopted to approximate the gradient term ∇Vi∗ according to the critic
weights updating law in (65), by which the optimal tracking FTC scheme can be obtained for intercon-
nected nonlinear systems. Compared with conventional actor-critic networks, the design process is simplified,
and the computational workload is reduced to some extent. Furthermore, the control process is displayed
in Figure 1.

6 STABILITY A NA LYSIS

In this section, the stability of the i equivalent augmented subsystem is proven under the proposed control scheme.
Lemma 1 (35). ∀(𝜛1 , 𝜛2 ) ∈ R2 , we have the following condition

𝜉 𝜌1 1
𝜛1 𝜛2 ≤ |𝜛1 |𝜌1 + |𝜛2 |𝜌2 ,
𝜌1 𝜌2 𝜉 𝜌2

where 𝜉 > 0, (𝜌1 − 1)(𝜌2 − 1) = 1 and 𝜌1 , 𝜌2 > 1.

Theorem 4. For the equivalent augmented subsystem (18), supposing that Assumptions 1, 3 and 4 hold, the
approximate optimal control strategy, the approximate optimal auxiliary control strategy and the critic weight
̃ci and the tracking
updating law are given by (61), (62) and (65), respectively. Then the weight estimation error W
error eti are UUB when
( ) ( ) 1
1
lci − 𝜆min 𝜙i1 𝜙⊤i1 − > 0 (67)
2 4

holds.
 10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2584 LIU et al.

FIGURE 1 Block diagram of the proposed optimal fault-tolerant tracking control scheme.

Proof. Select the candidate of the Lyapunov function as:

⎛ ⎞
N ⎜ ⎟
∑ 1
Li,1 (t) = ⎜V ∗ (ℑi ) + W̃⊤ W ̃ci ⎟. (68)
⎜ i
i=1 ⏟⏟⏟
2 ci ⎟
⎜ L (t) ⏟⏞⏟⏞⏟⎟
⎝ i,11 Li,12 (t) ⎠

Taking the time derivative of Li,11 (t) in (68) yields

( )⊤ [ ]
L̇ i,11 (t) = ∇Vi∗ i (ℑi (t)) + Gi (ℑi (t))̂
u0,i + IGi ̂
ai . (69)

Combining (32), (69) can be rewritten as

2 ( ∗ )⊤
L̇ i,11 (t) = −qai 𝜗i,m − ℑ⊤i Qi ℑi − Ui (u∗0,i ) − a∗⊤
i Qai ai − 𝜂i Pi (ℑi ) − ∇Vi
∗ 2
Gi (ℑi (t))u∗0,i
( )⊤ ( )⊤ ( )⊤
− ∇Vi∗ IGi a∗i + ∇Vi∗ Gi (ℑi (t))̂ u0,i + ∇Vi∗ IGi ̂ ai
( ) ( ) (70)
2 1 ⊤
= −ℑ⊤i Qi ℑi − Ui u∗0,i − qai 𝜗i,m − 𝜂i Pi2 (ℑi ) − ∇Vi∗ IGi Q−1 ⊤
ai IGi ∇Vi
∗

( ) ( 4
( )⊤ )⊤ ( )
+ ∇Vi∗ Gi (ℑi (t)) ̂ u0,i − u∗0,i + ∇Vi∗ IGi ̂ ai − a∗i .

It should be pointed that:

( ∗ )⊤ ( )
∇Vi IGi ̂
ai − a∗i
= 2a∗⊤ ∗ ∗⊤
i Qai ai − 2ai Qai ai
̂
[ ]
1 −1 ⊤ ⊤̃ 1 −1 ⊤
= 2a∗⊤
i Qai − Qai IGi ∇𝜎ci Wci − Q I ∇𝜀ci (71)
2 2 ai Gi
⊤ ⊤̃ ∗ ⊤
= −a∗⊤
i IGi ∇𝜎ci Wci − ai IGi ∇𝜀ci
1( )⊤ ⊤ ⊤̃ 1 ( ∗ )⊤ ⊤
= ∇Vi∗ IGi Q−1 ai IGi ∇𝜎ci Wci + ai IGi ∇𝜀ci .
∇Vi IGi Q−1
2 2
 10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU et al. 2585

By utilizing Assumption 4 and Young’s inequality,45 one gets

( )⊤ ( ) 1 1 ‖̃ ‖2 1 ̄ 2
ai − a∗i ≤ d2Vi ĪG4 qai 𝜎M,i
2 2
∇Vi∗ IGi ̂ + ‖Wci ‖ + dVi IGi qai d𝜀ci . (72)
4 i 4‖ ‖ 2

According to Lemma 1, Assumptions 1 and 4, we can get

( )⊤ ( )
∇Vi∗ Gi (ℑi (t)) ̂u0,i − u∗0,i
‖ ( ∗ )‖2
‖ u0,i ‖
2‖
≤ 𝜆i ‖tanh −1 ‖ +‖ ‖̂
‖2
u0,i − u∗0,i ‖ (73)
𝜆i ‖ ‖ ‖ ‖
‖
‖ ‖
1 2 ‖ ‖2
≤ d2Vi Gi,m + ‖̂ u − u∗0,i ‖ .
4 ‖ 0,i ‖
√ √
Based on ‖tanh(C∗ )‖ ≤ ̂ ≤ mi 43 and Assumption 4, we can obtain
mi , || tanh(C)||

( )
‖ ‖2 ‖ ‖2
u0,i − u∗0,i ‖ = ‖
‖̂ ‖
̂ − 𝜀u∗ ‖
𝜆i tanh(C∗ ) − tanh(C) 0,i ‖
‖ ‖ ‖ ‖
2‖ ̂ ‖2 ‖ ∗ ‖2
≤ 2𝜆i ‖tanh(C ) − tanh(C)‖ + 2‖𝜀u0,i ‖
∗
‖ ‖ ‖ ‖ (74)
( )
2 ̂ 2 + 2‖
≤ 4𝜆2i ‖tanh(C∗ )‖ + || tanh(C)||
‖2
‖𝜀u∗0,i ‖
‖ ‖
≤ 8𝜆2i mi + 2D2𝜀u∗ .
0,i

Observing (73) and (74), the following conclusion can be easily drawn:

( ∗ )⊤ ( )
1 2
∇Vi Gi (ℑi (t)) ̂u0,i − u∗0,i ≤ d2Vi Gi,m + 8𝜆2i mi + 2D2𝜀u∗ . (75)
4 0,i

( )
By using the conclusions Ui u∗0,i ≥ 0 and 𝜂i Pi2 (ℑi ) ≥ 0 in References 41 and 43, it follows that

( )
1 1 ‖̃ ‖2
L̇ i,11 (t) ≤ −𝜆min Qi ‖eti ‖2 + d2Vi ĪG4 qai 𝜎M,i
2 2
+ ‖W ci ‖
4 i 4‖ ‖ (76)
1 1 2
+ dVi ĪG2 qai d𝜀ci + d2Vi Gi,m + 8𝜆2i mi + 2D2𝜀u∗ .
2 i 4 0,i

Taking the time derivative of Li,12 (t) in (68) yields

̃ci 𝜙i1
W
L̇ i,12 (t) = −lci W
̃⊤ 𝜙i1 𝜙⊤ W
̃
i1 ci + lci 𝜃cHi , (77)
ci
𝜙i2
( )
where 𝜙i1 = 𝜙i ∕ 1 + 𝜙⊤i 𝜙i and 𝜙i2 = 1 + 𝜙⊤i 𝜙i > 1.
Based on Young’s inequality and assuming |𝜃cHi | < 𝜆𝜃i with 𝜆𝜃i > 0, we get

( ) l2 𝜃 2
1 ̃⊤
L̇ i,12 (t) ≤ − lci − ̃ci + ci cHi
Wci 𝜙i1 𝜙⊤i1 W
2 2𝜙2i2 (78)
( ) ( )‖ ‖2 1 2 2
1 ̃ ‖ + l 𝜆 .
≤ − lci − 𝜆min 𝜙i1 𝜙⊤i1 ‖W
2 ‖ ci ‖ 2 ci 𝜃i

Substituting (76) and (78) into (68) yields

N [ ( ) [( )
∑ 1 ( ) 1 ]‖ ‖2 ]
L̇ i,1 (t) ≤ −𝜆min Qi ‖eti ‖2 − lci − 𝜆min 𝜙i1 𝜙⊤i1 − ‖W ̃ci ‖ + Ξi , (79)
i=1
2 4 ‖ ‖
 10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2586 LIU et al.

2
where Ξi = (1∕2)l2ci 𝜆2𝜃i + (1∕4)d2V ĪG4 qai 𝜎M,i + (1∕2)dVi ĪG2 qai d𝜀ci + (1∕4)d2V Gi,m + 8𝜆2i mi + 2D2𝜀u∗ .
2 2
i i i i 0,i
The inequality (67) and (79) implies L̇ i,1 (t) < 0 only if one of the following conditions is violated:

⎧ √ ⎫
⎪ √ ⎪
Ωeti ∶ ⎨‖eti ‖ ≤ √ √ Ξi
( ) ⎬, (80)
⎪ 𝜆min Qi ⎪
⎩ ⎭
⎧ √ ⎫
⎪‖ ‖ √ ⎪
ΩW ̃ci ‖ ≤ √
∶ ⎨‖W √ ( )
Ξi
⎬. (81)
̃ci
‖ ‖ ( )
⎪ lci − 12 𝜆min 𝜙i1 𝜙⊤i1 − 1
⎪
⎩ 4⎭

̃ci are ensured via the Lyapunov exten-

Therefore, the boundedness of eti and the weight estimation error W
sion theorem. Furthermore, the ultimate bounds of eti and W ̃ci are presented in (80) and (81), respectively.
The proof ends here. ▪

7 SIMULATION

In this section, simulation examples are given to confirm the effectiveness of the developed control scheme for nonlinear
interconnected systems with faults and disturbances. Consider the following nonlinear system, which is composed of two
interconnected subsystems:
[ ] [ ] [ ]
−x11 + x12 cos(2x11 ) 0 −1
ẋ 1 = 𝜇1 (t) +
+ 𝜗1 (t) (82)
−x12 − x12 sin2 (x11 )2.5 0
[ ] [ ] [ ]
3
−x21 + x22 0 −1.5
ẋ 2 = + 𝜇2 (t) + 𝜗2 (t) (83)
−x22 − cos(x21 ) sin(x21 ) 1.5 0

with 𝜇1 (t) = u1 (t) + Θ(t − tT )Δu1 + Z1 (x) and 𝜇2 (t) = u2 (t) + Θ(t − tT )Δu2 + Z2 (x). x1 = [x11 , x12 ]⊤ ∈ R2 and
x2 = [x21 , x22 ]⊤ ∈ R2 are system states. The interconnection Z1 (x) = (x11 + x22 )sin2 (x12 ) cos(0.5x21 ) and Z2 (x) =
2
0.5(x12 + x22 ) cos(ex21 ). Let P1 (x1 ) = ‖x1 ‖ and P2 (x2 ) = ‖x2 ‖. Then, the reference system dynamic is defined as
[ ]
ri2 cos(ri2 ) − ri1
ṙ i = , (84)
−sin2 (ri1 ) − 0.3ri2

where r1 = [r11 , r12 ]⊤ ∈ R2 and r2 = [r21 , r22 ]⊤ ∈ R2 are reference states. We define the tracking errors as eti1 = xi1 −
[ ]⊤
ri1 and eti2 = xi2 − ri2 , so the augmented state vector can be expressed as ℑi = ℑi1 , ℑi2 , ℑi3 , ℑi4 = [eti1 , eti2 , ri1 , ri2 ]⊤ ,
i = 1, 2. Therefore, the initial value of the augmented states are selected as ℑ10 = [2.5, 0.5, 0.5, −0.5]⊤ and ℑ20 =
[0.5, −0.5, 0.5, 2.5]⊤ . In addition, let the abrupt fault occurs at tT = 3 and tT = 10 for subsystems (82) and (83), and the
following actuator faults are considered:
{
0.2 cos(4t), tT < 3
Θ(t − tT )Δu1 = , (85)
0.1 cos(t), tT ≥ 3
{
0.3 sin(3t), tT < 10
Θ(t − tT )Δu2 = . (86)
0.2 sin(t), tT ≥ 10

We select Q1 = Q2 = In , Qa1 = Qa2 = 1 and 𝜆1 = 𝜆2 = 2. The random external disturbances are defined as 𝜗1 (t) =
0.5e−t cos(2t) sin(t) and 𝜗2 (t) = 0.8e−t cos(2t) sin(t). The integral slide mode surface and the slide mode control strategy
are presented as (10) and (13), respectively, where the slide mode gain Gf 1 = Gf 2 = 10. Inspired by Reference 46, a hyper-
( )
bolic tangent function tanh G⊤i Ti⊤ si ∕𝛾i is selected to replace sgn(⋅) to weaken the chattering phenomenon, where 𝛾i is
chosen as 0.09, i = 1, 2. In the NN observer, the observation error eoi = [eoi1 , eoi2 , eoi3 , eoi4 ]⊤ , i = 1, 2 and the initial weights
 10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU et al. 2587

2.5

1.5

0.5

-0.5
0 10 20 30 40 50
Time(sec)

FIGURE 2 The observation error eo1 .

2.5

1.5

0.5

-0.5
0 10 20 30 40 50
Time(sec)

FIGURE 3 The observation error eo2 .

̂  1, w
w ̂  2, w
̂ G1 , w
̂ G2 , w
̂ H1 and w ̂ H2 are selected within [−1, 1]. Let the observer parameters 𝛼 i = 𝛼Gi = 𝛼Hi = 0.1, i = 1, 2
and Go1 = Go2 = In . The activation functions 𝜎 i (ℑ ̂ i ) = tanh(ℑ ̂ i ), 𝜎Gi (ℑ
̂ i , uIi ) = tanh(ℑ̂ i )uIi and 𝜎Hi (ℑ
̂ i , 𝜗i ) = tanh(ℑ
̂ i )𝜗i ,
i = 1, 2. For critic NN, the critic activation function, the weight vector, and the critic learning rate are selected as 𝜎ci (ℑi ) =
[ 2 ]⊤ [ ]⊤
ℑi1 , ℑi1 ℑi2 , ℑi1 ℑi3 , ℑi1 ℑi4 , ℑ2i2 , ℑi2 ℑi3 , ℑi2 ℑi4 , ℑ2i3 , ℑi3 ℑi4 , ℑ2i4 , W ̂ci = W ̂(1) , … , W ̂(10) and lci = 0.75, i = 1, 2. During
ci ci
the learning process, a small decreasing probing noise is introduced into the control input, such that the persistence of
excitation condition can be ensured.
Figures 2–13 present the simulation results. The evolutions of the observation errors for the augmented sub-
systems associated with (82) and (83) are shown in Figures 2 and 3, respectively, from which it can be seen that
the observation errors gradually converge to a small domain around zero. The SMC strategies used to counter-
act the effects of faults for subsystems (82) and (83) are depicted in Figures 4 and 5, respectively. From Figure 6,
the continuous optimal control strategies u0,1 and u0,2 gradually reach the equilibrium point within the con-
strained range. The evolutions of the composite control strategies u1 and u2 are displayed in Figures 7 and 8,
respectively. It is evident from Figure 9 that the tracking errors of subsystems (82) and (83) gradually tend to
around zero, which also means that each subsystem state gradually tracks the reference trajectory given in (84).
Figures 10–13 demonstrate the progression of the critic network learning weights reaching an ideal value and
the finally converged values are W ̂c1 = [2.618, 1.977, 1.849, 1.514, 0.894, 0.694, 0.145, 0.533, −0.059, −0.751] and W ̂c2 =
[1.753, 1.037, 1.084, 4.175, −0.296, −0.407, 3.876, −0.667, 4.493, 5.720].
10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU et al.

50
50

50

40
40

40

30
30

30

Time(sec)
Time(sec)

Time(sec)

20
20

20
The sliding mode control input of subsystem (82).

The sliding mode control input of subsystem (83).

The optimal control strategies u0,1 and u0,2 .

10
10
10

0
0

-0.1
0.02

-0.02

-0.04

-0.06

-0.08

-0.12
-0.1

-0.2

-0.3

-0.1

-0.2

-0.3

-0.4
0

0
0.2

0.1

0.4

0.3

0.2

0.1
FIGURE 4

FIGURE 5

FIGURE 6
2588
10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2589

50

50

50
40

40

40
30

30

30
Time(sec)

Time(sec)

Time(sec)
20

20

20

Tracking errors of subsystems (82) and (83).

10

10

10
The composite control strategy u2 .
The composite control strategy u1 .
0

0
-0.1

-0.2

-0.3

-0.5

-0.5
0

0
0.2

0.1

0.5

2.5

1.5

0.5
FIGURE 8

FIGURE 9
FIGURE 7
LIU et al.
10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU et al.

50
50

50
40
40

40
30
30

30
Time(sec)
Time(sec)

Time(sec)
20
20

20
10
10

10
̂(10) .
̂(5) .

̂(5) .
c1

c1

c2
̂(1) − W

̂(6) − W

̂(1) − W
0
0

-0.5
1

-1
0.5

0
3

0
3.5

2.5

1.5

0.5

-1
c1

c1

c2
The critic weights W

The critic weights W

The critic weights W

F I G U R E 10

F I G U R E 11

F I G U R E 12
2590
 10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU et al. 2591

7
6
5

2
1
0

-1

-2
0 10 20 30 40 50
Time(sec)

F I G U R E 13 ̂(6) − W
The critic weights W ̂(10) .
c2 c2

F I G U R E 14 Tracking error under the fault-tolerant control scheme in Reference 35.

In order to further highlight the advantages of the control scheme developed in this paper, we apply the fault-tolerant
control scheme developed in Reference 35 and the one in this paper to the investigated optimal tracking control problem,
respectively, and consider the following interconnected nonlinear system:
[ ] [ ] [ ]
−x11 + x12 cos(2x11 ) 0 + Δ1 −1
ẋ 1 = 2
+ 𝜇1 (t) + 𝜗1 (t) (87)
−x12 − x12 sin (x11 ) 2.5 0
[ ] [ ] [ ]
3
−x21 + x22 0 + Δ2 −1.5
ẋ 2 = + 𝜇2 (t) + 𝜗2 (t) (88)
−x22 − cos(x21 ) sin(x21 ) 1.5 0
 10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2592 LIU et al.

F I G U R E 15 Tracking error under the integral sliding-mode fault-tolerant scheme in this paper.

with Δ1 = 0.3 and Δ2 = 0.2. We assume that subsystems (87) and (88) are obtained after the parameters of subsystems
(82) and (83) are changed due to certain external factors. The reference system dynamic (84) continues to be used. In
addition, actuator faults similar to those in Reference 35 are considered, as follows:
{
0, tT < 40
Θ(t − tT )Δu1 = , (89)
1, tT ≥ 40
{
0, tT < 60
Θ(t − tT )Δu2 = . (90)
2, tT ≥ 60

By comparing Figures 14 and 15, we can observe that after the faults occur, the effects of the faults can be withstood
more effectively under the optimal fault-tolerant tracking control scheme developed in this paper, and the fluctuation of
tracking errors is even weaker during the process of withstanding the faults, which also means that the fluctuation of the
subsystem states is even weaker. Furthermore, the tracking performance as shown in Figure 15 can still be obtained when
the parameters of the system itself are changed while other design parameters remain unchanged, which demonstrates
the strong robustness of the control scheme proposed in this paper.

8 CO N C LU S I O N

In this paper, an optimal fault-tolerant tracking control scheme was proposed for constrained interconnected nonlin-
ear systems with unmatched disturbances via ADP. By constructing an augmented system consisting of tracking error
and the reference trajectory, the original optimal tracking control problem was transformed into the optimal regulation
control problem for the augmented system. The effect of actuator faults was effectively eliminated, and optimal con-
trol performance can be guaranteed under the developed integral sliding-mode fault-tolerant tracking control scheme. A
NN-based observer was designed to identify completely unknown system dynamics, such that the restriction on system
dynamics was relaxed. Then, the solution of modified HJBEs was obtained by using a critic network combined with the
gradient descent methodology. Based on the Lyapunov approach, all signals in the closed-loop augmented system were
 10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU et al. 2593

UUB. Finally, the effectiveness of the developed decentralized fault-tolerant tracking control scheme was shown via sim-
ulation results. However, the optimal control scheme proposed in this paper is based on a time-triggered mechanism,
which may result in the waste of communication resources. In future research, we will consider introducing the dynamic
event-triggered mechanism developed in Reference 47 into the optimal fault-tolerant tracking control scheme proposed in
this paper, aiming at conserving appropriate communication resources. Additionally, we will also be dedicated to design-
ing a novel fault-tolerant control method to address the optimal control problem of interconnected nonlinear systems
with multiple faults.

AU THOR CONTRIBUTIONS
Shihui Liu: Conceptualization; methodology; writing original draft. Ning Xu: Conceptualization; visualization.
Ning Zhao: Conceptualization; visualization. Liang Zhang: Conceptualization; visualization.

ACKNOWLEDGMENTS
This work was partially supported by the National Natural Science Foundation of China, under grant 62203064.

DATA AVAILABILITY STATEMENT

No data was used for the research described in the article.

ORCID
Shihui Liu https://orcid.org/0009-0008-2810-6452
Ning Xu https://orcid.org/0000-0002-0717-1713
Liang Zhang https://orcid.org/0000-0002-1907-016X

REFERENCES
1. Chung IY, Park SW, Kim HJ, et al. Operating strategy and control scheme of premium power supply interconnected with electric power
systems. IEEE Trans Power Deliv. 2005;20(3):2281-2288.
2. Zhao N, Zhao X, Chen M, Zong G, Zhang H. Resilient distributed event-triggered platooning control of connected vehicles under
denial-of-service attacks. IEEE Trans Intell Transp Syst. 2023;24(6):6191-6202.
3. Wu X, Zhao N, Ding S, Wang H, Zhao X. Distributed event-triggered output-feedback time-varying formation fault-tolerant control for
nonlinear multi-agent systems. IEEE Trans Autom Sci Eng. 2024. doi:10.1109/TASE.2024.3400325
4. Li Y, Yang G. Graph theory-based decentralized adaptive output-feedback control for a class of nonlinear interconnected systems. IEEE
Trans Cybern. 2019;49(7):2444-2453.
5. Ma H, Xu L. Decentralized adaptive fault-tolerant control for a class of strong interconnected nonlinear systems via graph theory. IEEE
Trans Automat Contr. 2021;66(7):3227-3234.
6. Huang S, Niu B, Wang H, Xu N, Zhao X. Prescribed performance-based low-complexity adaptive 2-bit-triggered control for unknown
nonlinear systems with actuator dead-zone. IEEE Trans Circuits Syst Express Briefs. 2024;71(2):762-766.
7. Wang C, Wen C, Lin Y, Wang W. Decentralized adaptive tracking control for a class of interconnected nonlinear systems with input
quantization. Automatica. 2017;81:359-368.
8. Liu Y, Fang Z, Li Y. Global decentralized output feedback control for interconnected time delay systems with saturated PI hysteresis.
J Franklin Inst. 2020;357(9):5460-5484.
9. Rigatos G, Abbaszadeh M. Nonlinear optimal control for multi-DOF robotic manipulators with flexible joints. Optim Control Appl Methods.
2021;42(6):1708-1733.
10. Zhao H, Zhao N, Zong G, Zhao X, Xu N. Sliding-mode surface-based approximate optimal control for nonlinear multiplayer
stackelberg-nash games via adaptive dynamic programming. Commun Nonlinear Sci. 2024;132:107928. doi:10.1109/TASE.2023.3317902
11. Malisani P, Chaplais F, Petit N. An interior penalty method for optimal control problems with state and input constraints of nonlinear
systems. Optim Control Appl Methods. 2016;37(1):3-33.
12. Li X, Dong L, Sun C. Data-based optimal tracking of autonomous nonlinear switching systems. IEEE/CAA J Automat Sin.
2021;8(1):227-238.
13. Zhao J, Yang C, Gao W, Zhou L. Reinforcement learning and optimal setpoint tracking control of linear systems with external disturbances.
IEEE Trans Industr Inform. 2022;18(11):7770-7779.
14. Cheng W, Zhang K, Jiang B. Hierarchical structure-based fixed-time optimal fault-tolerant time-varying output formation control for
heterogeneous multiagent systems. IEEE Trans Syst Man Cybern Syst. 2023;53(8):4856-4866.
15. Shen Z, Zhang L, Niu B, Zhao N. Event-based reachable set synthesis for delayed nonlinear semi-Markov systems. Chaos Solitons Fract.
2023;177:114284.
16. Liu S, Wang H, Liu Y, Xu N, Zhao X. Sliding-mode surface-based adaptive optimal nonzero-sum games for saturated nonlinear multi-player
systems with identifier-critic networks. Neurocomputing. 2024;584:127575.
 10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2594 LIU et al.

17. Liu S, Niu B, Xu N, Zhao X. Zero-sum game-based decentralized optimal control for saturated nonlinear interconnected systems via a data
and event driven approach. IEEE Syst J. 2024;18(1):758-769.
18. Ye J, Bian Y, Luo B, Hu M, Xu B, Ding R. Costate-supplement ADP for model-free optimal control of discrete-time nonlinear systems.
IEEE Trans Neural Netw Learn Syst. 2024;35(1):45-59.
19. Zhao H, Wang H, Niu B, Zhao X, Xu N. Adaptive fuzzy decentralized optimal control for interconnected nonlinear systems with unmodeled
dynamics via mixed data and event driven method. Fuzzy Set Syst. 2024;474:108735.
20. Na J, Lv Y, Zhang K, Zhao J. Adaptive identifier-critic-based optimal tracking control for nonlinear systems with experimental validation.
IEEE Trans Syst Man Cybern Syst. 2022;52(1):459-472.
21. Xue S, Luo B, Liu D, Yang Y. Constrained event-triggered H∞ control based on adaptive dynamic programming with concurrent learning.
IEEE Trans Syst Man Cybern Syst. 2022;52(1):357-369.
22. Wang D, Gao N, Liu D, Li J, Lewis FL. Recent Progress in reinforcement learning and adaptive dynamic programming for advanced control
applications. IEEE/CAA J Automat Sin. 2024;11(1):18-36.
23. Wang D, Wang J, Zhao M, Xin P, Qiao J. Adaptive multi-step evaluation design with stability guarantee for discrete-time optimal learning
control. IEEE/CAA J Automat Sin. 2023;10(9):1797-1809.
24. Ding L, Li S, Gao H, Liu Y, Huang L, Deng Z. Adaptive neural network-based finite-time online optimal tracking control of the nonlinear
system with dead zone. IEEE Trans Cybern. 2021;51(1):382-392.
25. Tu Y, Fang H, Wang H, Shi K, He S. Reinforcement learning-based adaptive optimal tracking algorithm for Markov jump systems with
partial unknown dynamics. Optim Control Appl Methods. 2022;43(5):1435-1449.
26. Zhu B, Xu N, Zong G, Zhao X. Adaptive optimized backstepping tracking control for full-state constrained nonlinear strict-feedback
systems without using barrier Lyapunov function method. Optim Control Appl Methods. 2024;45(5):2051-2075. doi:10.1002/oca.3136
27. Zhang L, Zhang X, Chang X, Zhao N. Adaptive fault-tolerant control-based real-time reachable set synthesis of heterogeneous nonlinear
singular multiagent systems with uncertain parameters. IEEE Trans Autom Sci Eng. 2023. doi:10.1109/TASE.2023.3298343
28. Zhang L, Wang J, Zhao X, Zhao N, Sharaf S. Event-based reachable set synthesis for continuous delayed fuzzy singularly perturbed systems.
IEEE Trans Circuits Syst II Express Briefs. 2024;71(1):246-250.
29. Zhao N, Zhao X, Zong G, Xu N. Resilient event-triggered filtering for networked switched T-S fuzzy systems under denial-of-service attacks.
IEEE Trans Fuzzy Syst. 2024;32(4):2140-2152.
30. Wei F, Zhang L, Niu B, Zong G. Adaptive decentralized fixed-time neural control for constrained strong interconnected nonlinear systems
with input quantization. Int J Robust Nonlinear Control. 2024. doi:10.1002/rnc.7497
31. Li K, Li Y. Fuzzy adaptive optimal consensus fault-tolerant control for stochastic nonlinear multiagent systems. IEEE Trans Fuzzy Syst.
2022;30(8):2870-2885.
32. Zhao N, Tian Y, Zhang H, Herrera-Viedma E. Fuzzy-based adaptive event-triggered control for nonlinear cyber-physical systems against
deception attacks via a single parameter learning method. Inform Sci. 2024;657:119948.
33. Wang Z, Liu L, Zhang H, Xiao G. Fault-tolerant controller design for a class of nonlinear MIMO discrete-time systems via online
reinforcement learning algorithm. IEEE Trans Syst Man Cybern Syst. 2016;46(5):611-622.
34. Zhao B, Liu D, Li Y. Observer based adaptive dynamic programming for fault tolerant control of a class of nonlinear systems. Inform Sci.
2017;384:21-33.
35. Zhao Y, Wang H, Xu N, Zong G, Zhao X. Reinforcement learning-based decentralized fault tolerant control for constrained interconnected
nonlinear systems. Chaos Solitons Fract. 2023;167:113034.
36. Zhao Y, Wang J, Yan F, Shen Y. Adaptive sliding mode fault-tolerant control for type-2 fuzzy systems with distributed delays. Inform Sci.
2019;473:227-238.
37. Fan Q, Yang G. Nearly optimal sliding mode fault-tolerant control for affine nonlinear systems with state constraints. Neurocomputing.
2016;216:78-88.
38. Wang D, Mu C, Liu D, Ma H. On mixed data and event driven design for adaptive-critic-based nonlinear H∞ control. IEEE Trans Neural
Netw Learn Syst. 2018;29(4):993-1005.
39. Qiao J, Li M, Wang D. Asymmetric constrained optimal tracking control with critic learning of nonlinear multiplayer zero-sum games.
IEEE Trans Neural Netw Learn Syst. 2024;35(4):5671-5683.
40. Biswas A, Anselma PG, Emadi A. Real-time optimal energy management of multimode hybrid electric powertrain with online trainable
asynchronous advantage actor–critic algorithm. IEEE Trans Transp Electr. 2021;8(2):2676-2694.
41. Yang X, He H. Decentralized event-triggered control for a class of nonlinear-interconnected systems using reinforcement learning. IEEE
Trans Cybern. 2021;51(2):635-648.
42. Zhao H, Wang H, Niu B, Zhao X, Alharbi KH. Event-triggered fault-tolerant control for input-constrained nonlinear systems with
mismatched disturbances via adaptive dynamic programming. Neural Netw. 2023;164:508-520.
43. Liu D, Wang D, Li H. Decentralized stabilization for a class of continuous-time nonlinear interconnected systems using online learning
optimal control approach. IEEE Trans Neural Netw Learn Syst. 2014;25(2):418-428.
44. Zhang S, Zhao B, Liu D, Zhang Y. Observer-based event-triggered control for zero-sum games of input constrained multi-player nonlinear
systems. Neural Netw. 2021;144:101-112.
45. Zhang L, Liang J, Feng Z, Zhao N. Improved results of asynchronous mixed H∞ and passive control for discrete-time linear switched
system with mode-dependent average dwell time. Chaos Solitons Fract. 2024;178:114401.
 10991514, 2024, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.3173 by DHA Suffa University, Karachi, Wiley Online Library on [24/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU et al. 2595

46. Yang D, Li T, Xie X, Zhang H. Event-triggered integral sliding-mode control for nonlinear constrained-input systems with disturbances
via adaptive dynamic programming. IEEE Trans Syst Man Cybern Syst. 2020;50(11):4086-4096.
47. Girard A. Dynamic triggering mechanisms for event-triggered control. IEEE Trans Automat Contr. 2014;60(7):1992-1997.

How to cite this article: Liu S, Xu N, Zhao N, Zhang L. Observer-based optimal fault-tolerant tracking control
for input-constrained interconnected nonlinear systems with mismatched disturbances. Optim Control Appl Meth.
2024;45(6):2572-2595. doi: 10.1002/oca.3173