Aerospace Science and Technology 144 (2024) 108803

Contents lists available at ScienceDirect

Aerospace Science and Technology

journal homepage: www.elsevier.com/locate/aescte

Optimized intelligent tracking control for a quadrotor unmanned aerial

vehicle with actuator failures
Bo Li a , Hui Liu a , Choon Ki Ahn b,∗ , Wenquan Gong a,b
a Institute of Logistics Science and Engineering, Shanghai Maritime University, Shanghai, 201306, China
b School of Electrical Engineering, Korea University, Seoul 136-701, South Korea

A R T I C L E I N F O A B S T R A C T

Communicated by Qinglei Hu This work investigates the implementation of tracking control for a quadrotor unmanned aerial vehicle exposed
to disturbances and actuator failures. Cost reduction in control processes is not usually considered in conventional
Keywords:
control methods. To overcome this limitation, this work proposes an optimized intelligent control scheme
Quadrotor
Adaptive dynamic programming that utilizes adaptive dynamic programming. Firstly, a critic-only structure is employed to learn the nominal
Fault-tolerant control control strategies, including two critic neural network update laws, lifting the persistent excitation condition.
Radial basis function neural network Subsequently, two specifically designed observers are developed on the radial basis function neural network
(RBFNN) to reconstruct the disturbances and actuator failures and compensate for the nominal strategies. The
observers achieve fixed-time convergence by employing the fixed-time update laws of the RBFNNs. Moreover,
two variable vectors using the Gaussian error function are introduced to reduce the impact caused by the
observation errors. Additionally, the superior performance of the proposed control scheme is validated through
numerical simulations.

1. Introduction Because of its effectiveness in reducing the control cost, optimal con-
trol has been employed to overcome this limitation. However, a crucial
challenge to be overcome in optimal control is solving the intractable
Thanks to their superior ability to take off and land vertically, Hamilton-Jacobi-Bellman (HJB) equation [8]. In this case, the adap-
quadrotor unmanned aerial vehicles (UAVs) are widely used in fields tive dynamic programming (ADP) techniques adopt NNs to provide a
such as delivery logistics [1], mapping [2], maritime cruising [3], etc. feasible approach. Therefore, the ADP technique has drawn significant
Trajectory tracking is crucial for quadrotors during task execution in attention from scholars [9][10][11]. As one of the intelligent methods
different scenarios. However, it is challenging to develop effective track- integrated with reinforcement learning technology, the ADP includes
ing control strategies due to the underactuation and strongly coupled two common structures. The traditional structure of the ADP is the
dynamics of quadrotor UAVs. In practical environments, the presence actor-critic (AC) structure, using which Vamvoudakis and Lewis [9]
of disturbances (e.g., wind) can lead to poor performance and dam- dealt with the optimal control problem of nonlinear systems without
age to the quadrotor’s actuators, resulting in the failure of tracking unknown system dynamics. To solve the tracking control problem, a
tasks. Moreover, the control cost (e.g., electrical energy) consumed in control algorithm integrating optimization and backstepping control,
quadrotor flight is very high. Thus, it is vital to develop an efficient involving an actor NN and a critic NN devoted to improving and eval-
control scheme that strikes a balance between control cost and con- uating the control strategy, respectively, was introduced [12]. Wang et
trol performance. Various methods have been developed to handle the al. [13] built a critic-only structure that was superior to AC in com-
control problem of nonlinear systems (e.g., the quadrotor UAV), includ- putation reduction. Thereafter, the critic-only structure of the ADP
ing sliding mode control (SMC), backstepping control, adaptive control, technique has been widely applied. Yang et al. [14] used the critic-
[4][5][6][7] etc. However, none of these schemes achieves a balance only structure to develop an ADP-based control strategy that can be
between control cost and control performance, thus leading to needless employed to solve the attitude reorientation control problem. Neverthe-
energy waste. less, the ADP performs poorly in terms of tackling time-varying signals

* Corresponding author.
E-mail addresses: nemo127@163.com (B. Li), liuh_587@163.com (H. Liu), hironaka@korea.ac.kr (C.K. Ahn), wqgong902@gmail.com (W. Gong).

https://doi.org/10.1016/j.ast.2023.108803
Received 19 May 2023; Received in revised form 9 November 2023; Accepted 30 November 2023
Available online 6 December 2023
1270-9638/© 2023 Elsevier Masson SAS. All rights reserved.
B. Li, H. Liu, C.K. Ahn et al. Aerospace Science and Technology 144 (2024) 108803

[15], and thus fails to meet the requirement of disturbance attenua- Notation: The special subscript 𝜁 denotes the scalar or vector of the
tion when UAVs fly in complex environments. This indicates that the position subsystem (𝜁 = 𝑤) or attitude subsystem (𝜁 = 𝑧). 𝑰 𝑛 denotes a unit
high-quality control of practical UAV systems cannot be realized only matrix with n dimensions. ⋅ denotes the lower bound of ⋅. 𝟏6 is a column
through the ADP technique. vector expressed by 𝟏6 = [1, 1, 1, 1, 1, 1]T . In addition, ⌈⋅⌋𝛼 = ‖ ⋅ ‖𝛼 sign (⋅)
To extend the practical applications of the ADP, control frameworks with ⋅ being an arbitrary vector and 𝛼 a positive constant.
that combine the ADP with other technologies have been proposed to
handle the control problem of the system in disturbance or actuator fail-
2. Problem formulation
ure conditions [15][16][17][18]. Furthermore, the ADP has been inte-
grated with other techniques, such as the super twisting algorithm [19]
and observers [20], in the control design of the quadrotor, thus proving 2.1. Dynamic model of a quadrotor UAV
superior to using the ADP technique alone [21]. To handle the track-
ing control problem of a quadrotor flying in actual environments, many A quadrotor UAV comprises a rigid body and four motors. The body
disturbance attenuations [22][23][24] and fault-tolerant control (FTC) fixed frame  = {𝑥𝑏 , 𝑦𝑏 , 𝑧𝑏 } and the inertial frame  = {𝑥𝑒 , 𝑦𝑒 , 𝑧𝑒 } are
schemes [25][26][27] have been proposed. Notably, observer-based defined to describe the quadrotor conveniently. 𝑿 = [𝑥, 𝑦, 𝑧]T denotes
control schemes generally yield good performance in handling lumped the center position of the quadrotor in  . Ω𝑖 is the speed of the ith mo-
disturbances mainly induced by actuator failures. However, many of the tor (i=1, 2, 3, 4), which is described in [27]. To illustrate the attitude
existing observers require common assumptions (e.g., bounded distur- information in  , the Euler angle vector 𝚯 = [𝜙, 𝜃, 𝜓]T is introduced.
bances and bounded disturbance derivatives) related to disturbances. The dynamic model of the quadrotor can be established in the follow-
Developing observers based on NNs is a feasible approach inspired by ing form [34]:
the approximation of unknown dynamics or states using NNs [28][29].
Due to their strong fitting abilities, radial basis function NNs 𝑚𝑿 ̈ + 𝑚𝑔𝒛𝐼 = 𝜌1 𝑢𝑡 + 𝒅, (1)
(RBFNNs) are usually employed to develop observers. An adaptive ( )
FTC scheme, which cannot achieve fast convergence characteristics 𝑱 (𝚯) 𝚯̈ + 𝑪 𝚯, 𝚯̇ 𝚯̇ = 𝝆𝑎 𝒖𝑎 + 𝝉 𝑑 , (2)
[30], was developed on an RBFNN and L2 gain for a morphing air-
where 𝑚 is the total mass of the quadrotor and 𝑔 denotes the accelera-
craft. To improve the convergence response, Lin et al. [31] proposed an
tion of gravity. A unit vector is defined as 𝒛𝐼 = [0, 0, 1]T . The coupling
event-based neural control scheme with finite-time convergence char-
term is denoted as  = [cos 𝜙 sin 𝜃 cos 𝜓 + sin 𝜙 sin 𝜓, cos 𝜙 sin 𝜃 sin 𝜓 −
acteristics for a six-rotor UAV. Moreover, Liu et al. [32] developed a
sin 𝜙 cos 𝜓, cos 𝜙 cos 𝜃]T . 𝝆𝑎 = diag{𝜌2 , 𝜌3 , 𝜌4 } with 0 < 𝜌𝑖 ≤ 1 being the
fixed-time control scheme on RBFNN for multi-robot systems with ac-
ith control efficiency factor. The definitions of the symmetric matrix
tuator failures. Although the convergence time can be free of initial
𝑱 (𝚯) and Coriolis term 𝑪(𝚯, 𝚯) ̇ can be seen in reference [20]. 𝒅 ∈ ℝ3
conditions, all weights of the RBFNN should be positive. To address
and 𝝉 𝑑 ∈ ℝ3 are the disturbance forces and torques, respectively, im-
this issue, an RBFNN observer-based control scheme was proposed by [ ]T
Zhang et al. [33]. They implemented an adaptive law for RBFNN that posed on the position and attitude loops. 𝑢𝑡 and 𝒖𝑎 = 𝑢𝜙 , 𝑢𝜃 , 𝑢𝜓 are
achieved fixed-time convergence. However, the upper limit on conver- the thrust and torques generated by four rotors, respectively. In addi-
gence time did not eliminate the need for acquiring knowledge of the tion, the relationship between control inputs (i.e., 𝑢𝑡 and 𝒖𝑎 ) and the
ideal weight information. motor speed Ω𝑖 can be seen in [27].
Inspired by the above findings, this work addresses the tracking To implement trajectory tracking, the error subsystems should be
control problem of a quadrotor UAV exposed to disturbances and ac- established subsequently. Define the position error vector 𝝌 𝑤1 = 𝑿 −
[ ]T [ ]T
tuator failures. Intelligent control technologies, optimization ideas, and 𝑿 𝑑 = 𝑥𝑒 , 𝑦𝑒 , 𝑧𝑒 , with 𝑿 𝑑 = 𝑥𝑑 , 𝑦𝑑 , 𝑧𝑑 denoting the desired posi-
fault tolerance are integrated into the design process of the scheme for tion. Redefining 𝝂 𝑤 = 𝑢t and 𝝂 𝑧 = 𝒖𝑎 for convenience, the system (1)
solving the control problem. The main contributions of this work are can be transformed into the following position error subsystem:
outlined as follows: {
1) An ADP-based control framework is proposed for a quadrotor 𝝌̇ 𝑤1 = 𝝌 𝑤2
, (3)
UAV with lumped disturbances mainly induced by actuator failures. 𝝌̇ 𝑤2 = −𝑔𝒛I + 1
𝝂 + 1
𝒅 ̈𝑑
−𝑿
𝑚 𝑤 𝑚 𝑤
Two critic NN update laws are developed for lifting the common persis-
tent excitation (PE) condition required in [15][21] to learn the approx- with
( lumped
) disturbance force 𝒅 𝑤 being expressed by 𝒅 𝑤 =
imate optimal nominal control strategies online. 𝜌1 − 1 𝝂 𝑤 + 𝒅 .
2) Two fixed-time RBFNN observers are developed to reconstruct Generally, 𝑿 and 𝜓 are selected as the controlled states to imple-
lumped disturbances. Superior to asymptotic convergence [30] and ment the full flight control of the quadrotor, and the desired roll and
finite-time convergence [31], the observers achieve fixed-time conver- pitch angles can be obtained by utilizing the determined 𝑿 𝑑 and yaw
gence by employing the proposed fixed-time RBFNN weight update angle 𝜓𝑑 , which are formulated as follows [21][34]:
laws. The proposed fixed-time RBFNN weight update law brings ben- ( )
efits as it decouples the observation error settling time from the ideal 𝑢𝑤1 sin𝜓𝑑 − 𝑢𝑤2 cos𝜓𝑑
𝜙𝑑 = arcsin ,
RBFNN weights, which is more reasonable compared with the existing ‖𝒖𝑤 ‖
( )
research [33]. 𝑢𝑤1 cos𝜓𝑑 + 𝑢𝑤2 sin𝜓𝑑
𝜃𝑑 = arctan ,
3) Using the Gaussian error function, two variable vectors are pro- 𝑢𝑤3
posed to reduce the impact of observation errors on the nominal sys- [ ]T 1
tem. Compared with the use of the hyperbolic tangent function [20], with 𝒖𝑤 = 𝑢𝑤1 , 𝑢𝑤2 , 𝑢𝑤3 = 𝝂 𝑤 .
𝑚
[ ]T
the variable vectors provide higher approximation accuracy. Addition- The attitude error vector is denoted as 𝝌 𝑧1 = 𝚯 − 𝚯𝑑 = 𝜙𝑒 , 𝜃𝑒 , 𝜓𝑒 .
ally, the proposed control framework balances control performance and Similar to the modeling process for the position error subsystem, the
control cost while ensuring that the control system remains ultimately attitude error subsystem can be obtained using
bounded (UUB).
{
The rest of this paper is organized as follows: The control prob- 𝝌̇ 𝑧1 = 𝝌 𝑧2
lem of the quadrotor under actuator failures is formalized in Section 2. , (4)
𝝌̇ 𝑧2 = −𝑱 −1 𝑪 𝚯̇ + 𝑱 −1 𝝂 𝑧 + 𝑱 −1 𝒅 𝑧 − 𝚯
̈𝑑
Section 3 introduces the ADP-based FTC algorithm, and Section 4 pro-
vides the numerical simulation results and analysis. The conclusions are with
( lumped
) disturbance torque 𝒅 𝑧 being expressed by 𝒅 𝑧 =
drawn in Section 5. 𝝆𝑚 − 𝑰 3 𝝂 𝑧 + 𝝉 𝑑 .

2
B. Li, H. Liu, C.K. Ahn et al. Aerospace Science and Technology 144 (2024) 108803

2.2. Model transformation NN can provide a superior scheme for this problem. If NNs contain
abundant sets of basis functions, the optimal cost function can be re-
Divide the control laws 𝝂 𝜁 into the feed-forward control laws 𝝂 𝜁1 constructed as follows:
and feedback laws 𝝂 𝜁2 (e.g., 𝝂 𝜁 = 𝝂 𝜁1 + 𝝂 𝜁2 ). Two feed-forward con- ( ) ( ) ( )
̈ 𝑑 + 𝑪 𝚯̇ 𝑑 and 𝝂 𝑤1 = 𝑚𝑔𝒛I + 𝑚𝑿
trol laws are designed as 𝝂 𝑧1 = 𝑱 𝚯 ̈ 𝑑 to 𝐽𝜁∗ 𝝌 𝜁 = 𝝎T𝜁 𝛿𝜁 𝝌 𝜁 + 𝜖𝜁 𝝌 𝜁 , (14)
compensate for time-varying known information related to 𝑿 𝑑 and 𝚯𝑑 .
where 𝝌 𝜁 ∈ ℂ𝜁 are compact sets. Thus, the optimal control policy 𝝂 ∗𝜁𝑛
Then systems (3) and (4) can be rewritten as
{ can be converted below:
𝝌̇ 𝑤1 = 𝝌 𝑤2 1 ( ( ) ( ))
(5) 𝝂 ∗𝜁𝑛 = − 𝑹−1 𝒒 T𝜁 ∇T 𝛿𝜁 𝝌 𝜁 𝝎𝜁 + ∇𝜖𝜁 𝝌 𝜁 . (15)
𝝌̇ 𝑤2 = 𝑚1 𝝂 𝑤2 + 𝑚1 𝒅 𝑤 , 2 𝜁
{
𝝌̇ 𝑧1 = 𝝌 𝑧2 Owing to the unknown optimal weight vectors, the approximation
(6) results of (14) can be obtained by:
𝝌̇ 𝑧2 = −𝑱 −1 𝑪𝝌 𝑧2 + 𝑱 −1 𝝂 𝑧2 + 𝑱 −1 𝒅 𝑧 .
( ) ( )
For simplification, systems (5) and (6) can be rewritten as the fol- 𝐽𝜁 𝝌 𝜁 = 𝝎̂ T𝜁 𝛿𝜁 𝝌 𝜁 . (16)
lowing compact form: Similarly, the optimal control policy (15) can be approximated as:
( )
𝝌̇ 𝜁 = 𝑝𝜁 (𝝌 𝜁 ) + 𝒒 𝜁 𝝂 𝜁2 + 𝒅 𝜁 , (7) 1 ( )
𝝂 𝜁𝑛 = − 𝑹−1 𝒒 T ∇T 𝛿𝜁 𝝌 𝜁 𝝎̂ 𝜁 . (17)
( ) [ ]T 2 𝜁 𝜁
with 𝑝𝜁 and 𝑞𝜁 being expressed by 𝑝𝑤 𝝌 𝑤 = 𝝌 T𝑤2 , 𝟎T3×1 , 𝒒 𝑤 = Subsequently, the approximated nonlinear functions can be derived
[ ]T ( ) [ ( )T ] T [ ]T below:
𝟎3×3 , 𝑚1 𝑰 3 , 𝑝𝑧 𝝌 𝑧 = 𝝌 T𝑧2 , − 𝑱 −1 𝑪𝝌 𝑧2 and 𝒒 𝑧 = 𝟎3×3 , 𝑱 −T .
( )
It is noteworthy that the conditions 𝑝𝜁 (𝟎) = 𝟎 are satisfied due to the 𝐻𝜁 𝝌 𝜁 , 𝝂 𝜁𝑛 , 𝝎̂ 𝜁 = 𝝎̂ T𝜁 𝜸 𝜁 + 𝝌 T𝜁 𝑸𝜁 𝝌 𝜁 + 𝝂 𝜁𝑛 T 𝑹𝜁 𝝂 𝜁𝑛 = 𝑒𝜁1 (18)
implementation of the feed-forward control technique. ( ( ) )
Therefore, the objective is transformed into finding the feedback with 𝜸 𝜁 = ∇𝛿𝜁 𝑝𝜁 𝝌 𝜁 + 𝒒 𝜁 𝝂 𝜁𝑛 . According to [9], 𝑒𝜁1 can be rewritten
control law 𝝂 𝜁2 to stabilize the system (7). as:

̃ T𝜁 𝜸 𝜁 + 𝜖𝜁𝐻 ,
𝑒𝜁1 = 𝝎 (19)
Assumption 1. For 𝝌 𝜁 ∈ 𝚽𝜁 , there are some positive constants 𝛿̌𝜁 , 𝜎̌ 𝜁 , 𝑞̌𝜁 ,
( )
𝜀̌ 𝜁 , 𝜖̌𝜁 , and 𝜖̌𝐻𝜁 such that ‖∇𝛿𝜁 (𝝌 𝜻 )‖ ≤ 𝛿̌𝜁 , ‖𝜎𝜁 𝝌 𝜁 ‖ ≤ 𝜎̌ 𝜁 , ‖𝒒 𝜁 ‖ ≤ 𝑞̌𝜁 , where 𝝎 ̃ 𝜁 = 𝝎̂ 𝜁 − 𝝎𝜁 , and the definitions of 𝜖𝜁𝐻 are consistent with [9].
‖𝜀𝜁 (𝝌 𝜁 )‖ ≤ 𝜀̌ 𝜁 , ‖∇𝜖𝜁 (𝝌 𝜁 )‖ ≤ 𝜖̌𝜁 , and ‖𝜖𝐻𝜁 ‖ ≤ 𝜖̌𝐻𝜁 . It is noteworthy that the relationship between 𝑒𝜁1 and the weight ap-
proximation errors 𝝎 ̃ 𝜁 can be inferred from (19). Thus, 𝑒𝜁1 is generally
3. Optimized intelligent FTC algorithm design used when developing critic NN weight update laws.

3.1. ADP-based nominal control algorithm Theorem 1. For nominal system (8) with cost function (9), if control strate-
gies 𝝂 𝜁𝑛 are given as (17), and the critic NN update laws are designed as
The nominal system of (7) is defined as follows:
follows:
𝝌̇ 𝜁 = 𝑝𝜁 (𝝌 𝜁 ) + 𝒒 𝜁 𝝂 𝜁𝑛 , (8) 𝑒𝜁1 𝜸 𝜁
̂̇ 𝜁 = −𝜅𝜁1 (
𝝎 )2 − 𝜅𝜁2 𝜑𝜁 𝝎̂ 𝜁 ,
where 𝝂 𝜁𝑛 denote the approximate optimized nominal control laws to (20)
1 + 𝜸 T𝜁 𝜸 𝜁
be developed. By referring to [9][11][15], the cost functions of the
( ( ) )
nominal system (8) can be defined in the following form: with 𝜅𝜁1 , 𝜅𝜁2 ∈ ℝ+ and 𝜑𝜁 = ‖∇𝛿𝜁 𝝌 𝜁 ‖ + 𝜇𝜁 𝑙𝜁 . 𝜇𝜁 denote small posi-
( )
∞ tive constants. Besides, 𝑙𝜁 are switching quantities (e.g., if ‖∇𝛿𝜁 𝝌 𝜁 ‖ = 0,
( ) then 𝑙𝜁 = 1; otherwise, 𝑙𝜁 = 0). Then, the UUB stability of 𝝎 ̃ 𝜁 and 𝝌 𝜁 in (8)
𝐽𝜁 𝝌 𝜁 = 𝝌 T𝜁 𝑸𝜁 𝝌 𝜁 + 𝝂 T𝜁𝑛 𝑹𝜁 𝝂 𝜁𝑛 𝑑𝜗. (9)
∫ is guaranteed.
𝑡

Assuming that the optimal control strategies 𝝂 ∗𝜁𝑛 exist, the corre-
( ) Proof. See Subsection 6.1 for details.
sponding 𝐽𝜁∗ 𝝌 𝜁 can be yielded as:
∞ One can conclude that the proposed control policies (17) can guar-
( )
𝐽𝜁∗ 𝝌𝜁 = 𝝌 T𝜁 𝑸𝜁 𝝌 𝜁 + 𝝂 ∗𝜁𝑛 T 𝑹𝜁 𝝂 ∗𝜁𝑛 𝑑𝜗. (10) antee the UUB stability of the nominal systems (8). The nominal control
∫ laws are baseline control laws that can stabilize nominal systems. How-
𝑡
ever, to tackle the optimized tracking problem of practical systems
Taking the time derivative of (10), one has: modeled in (7), it is necessary to carry out the FTC scheme in the fol-
( ) ( ( ) )
lowing design process.
𝐻𝜁 𝝌 𝜁 , 𝝂 ∗𝜁𝑛 , ∇𝐽𝜁∗ (𝝌 𝜁 ) = ∇T 𝐽𝜁∗ 𝑝𝜁 𝝌 𝜁 + 𝒒 𝜁 𝝂 ∗𝜁𝑛 + 𝝌 T𝜁 𝑸𝜁 𝝌 𝜁
(11)
+𝝂 ∗𝜁𝑛 T 𝑹𝜁 𝝂 ∗𝜁𝑛 = 0. 3.2. Fixed-time RBFNN observer for the lumped disturbances
Then, the optimal control policies 𝝂 ∗𝜁𝑛 can be calculated using:
The ADP technique lacks the ability to handle time-varying sig-
1 ( ) nals, resulting in poor tracking performance when the quadrotor faces
𝝂 ∗𝜁𝑛 = − 𝑹−1 𝒒 T ∇𝐽 ∗ 𝝌 𝜁 . (12)
2 𝜁 𝜁 𝜁 lumped disturbances. In order to overcome this defect, two distur-
Substituting 𝝂 ∗𝜁𝑛 into (11), the HJB equations can be yielded as: bance observers are developed to compensate for the lumped dis-
turbances. Owing to the strong approximation characteristics of the
( ) 1 ( )
𝝌 T𝜁 𝑸𝜁 𝝌 𝜁 + ∇T 𝐽𝜁∗ 𝑝𝜁 𝝌 𝜁 − ∇T 𝐽𝜁∗ (𝝌 𝜁 )𝒒 𝜁 𝑹−1 RBFNN, lumped disturbance terms 𝒅 𝜁 can be reconstructed using the
𝜁 𝒒 𝜁 ∇𝐽𝜁 𝝌 𝜁 = 0.
T ∗ (13)
4 following equations:
Obtaining analytically exact solutions for the HJB equations is ( ) ( )
widely acknowledged as very challenging. It is worth noting that the 𝒅 𝜁 = 𝑾 T𝜁 𝜎𝜁 𝝌 𝜁 + 𝜀𝜁 𝝌 𝜁 , (21)

3
B. Li, H. Liu, C.K. Ahn et al. Aerospace Science and Technology 144 (2024) 108803

where 𝑾 𝜁 ∈ ℝ𝑓 ×𝑡 denote the ideal weight matrices of two RBFNNs. performs better when it is used for smooth approximations of | ⋅ | with
The radial basis functions are denoted as respect to accuracy [35]. Moreover, using erf (⋅) to replace a discontin-
( ) [ ( ) ( ) ( )]T uous robust term (e.g., sign (⋅)) can effectively improve the chattering
𝜎𝜁 𝝌 𝜁 = 𝜎𝜁1 𝝌 𝜁 , 𝜎𝜁2 𝝌 𝜁 , ⋯ , 𝜎𝜁𝑓 𝝌 𝜁 phenomenon.
( ) ‖𝝌 𝜁 −Υ𝜁 ‖2
(e.g., 𝜎𝜁𝑓 𝝌 𝜁 = exp( − ), 𝑓 = 1, ⋯ , 6). 𝜂𝜁𝑓 refer to the width 4. Simulation results and analysis
𝜂𝜁2f
[ ]T
of the Gaussian function, 𝚼𝜁 = Υ𝜁1 , Υ𝜁2 , ⋯ , Υ𝜁6 represent the center
( ) Numerical simulation results will be presented in this section to val-
vectors of the receptive field, and 𝜀𝜁 𝝌 𝜁 ∈ ℝ denote the approxima-
6
idate the effectiveness of the proposed control scheme (called ADPNN).
tion errors.
Due to the unknown ideal weights 𝑾 𝜁 , the fault information can be
4.1. Simulation initial conditions setting
approximated as follows:
( )
𝒅̂ 𝜁 = 𝑾̂ 𝜁 𝜎𝜁 𝝌 𝜁 ,
T Simulations are performed on a quadrotor UAV. The physical param-
(22)
eters of the quadrotor, the initial conditions, and the numerical values
where the estimation results for 𝑾 𝜁 are represented by 𝑾 ̂ 𝜁 . The weight of external disturbances were defined the same as [20]. To simulate
approximation errors, 𝑾̃ 𝜁 , and observation errors, 𝒅̃ 𝜁 , can be expressed actuator failures, mutation-induced control efficiency factors are set,
as 𝑾̃ 𝜁 = 𝑾̂ 𝜁 − 𝑾 𝜁 and 𝒅̃ 𝜁 = 𝒅 𝜁 − 𝒅̂ 𝜁 , respectively. Define the error which are described by
vector 𝒆𝜁 = 𝝌 𝜁 − 𝝌̂ 𝜁 with two auxiliary systems being designed as: { {
1, 𝑡 < 20 1, 𝑡 < 15
⌈ ⌋1 ⌈ ⌋3 𝜌1 = , 𝜌2 = ,
𝝌̂̇ 𝜁 = 𝜆𝜁1 𝒆𝜁 + 𝜆𝜁2 𝒆𝜁 2 + 𝜆𝜁3 𝒆𝜁 + 𝑝𝜁 (𝝌 𝜁 ) + 𝒒 𝜁 𝝂 𝜁2 + 𝒒 𝜁 𝒅̂ 𝜁 . (23) 0.9, 𝑒𝑙𝑠𝑒 0.9, 𝑒𝑙𝑠𝑒
{ {
Taking the derivative of the aforedefined 𝒆𝜁 and recalling (7), one 1, 𝑡 < 25 1, 𝑡 < 35
𝜌3 = and 𝜌4 = .
has 0.9, 𝑒𝑙𝑠𝑒 0.9, 𝑒𝑙𝑠𝑒
⌈ ⌋1 ⌈ ⌋3 ( ( ) ( ))
𝒆̇ 𝜁 = −𝜆𝜁1 𝒆𝜁 −𝜆𝜁2 𝒆𝜁 2 −𝜆𝜁3 𝒆𝜁 −𝒒 𝜁 𝑾̃ T 𝜎𝜁 𝝌 𝜁 − 𝜀𝜁 𝝌 𝜁 . (24) The critic NNs are employed to learn the nominal control strategy, and
𝜁
RBFNNs are to reconstruct lumped simulations. The basis functions of
the critic NNs are referred to [20]. The initial weights of critic NN are
Theorem 2. Consider the RBFNN observers designed in (22), if the weight
𝝎̂ 𝑤 (0) = [25, 25, 30, 25, 25, 30]T and 𝝎
̂ 𝑧 (0) = 𝟏6 . In addition, the main
adaptive laws of RBFNNs are designed below:
parameters are 𝑹𝑤 = 2.5𝑰 3 , 𝑹𝑧 = 19𝑰 3 , 𝑸𝑤 = 5𝑰 6 , 𝑸𝑧 = 20𝑰 6 , 𝜆𝑤1 =
( ( ) )
̂̇ 𝜁 = −𝑎𝜁 −𝜎𝜁 𝝌 𝜁 𝒆T 𝒒 𝜁 + 𝑏𝜁 𝑾̂ 𝜁 + 𝑐𝜁 𝑾̂ 𝜁 𝑾̂ T 𝑾̂ 𝜁 , 𝜆𝑧1 = 50, 𝜆𝑤2 = 0.001, 𝜆𝑧2 = 0.2, 𝜆𝑤3 = 0.01, 𝜆𝑧3 = 0.3, 𝜅𝑤1 = 0.5, 𝜅𝑧1 =
𝑾 𝜁 𝜁 (25)
0.05, 𝑎𝑤 = 600, 𝑎𝑧 = 5, 𝑘𝑤 = 0.5, and 𝑘𝑧 = 2.
with 𝑎𝜁 ∈ ℝ+ . 𝑏𝜁 and 𝑐𝜁 are positive constants. Then, the observation errors
can achieve practical fixed-time stability. 4.2. Numerical simulations implementation driven by different schemes

Proof. See Subsection 6.2 for details. To highlight the superiority and effectiveness of the proposed
ADPNN scheme, two comparison schemes are implemented. One com-
bines the ADP-based control strategy proposed in [36] with fixed-time
3.3. ADP-based optimized intelligent FTC algorithm
NN observers proposed in [33] (called ADPFT), and the other is pre-
sented in the Appendix of [37] (called PIDSMC). For a fairer compari-
Integrating the ADP-based nominal controllers and RBFNN ob-
son, the convergence time of the tracking errors belonging to different
servers, the optimized intelligent FTC scheme will be developed in this
control schemes is set to an almost equal level by adjusting the param-
subsection.
eters. Notably, in the comparison simulation figures, the red, purple,
and green curves represent the proposed ADPNN, PIDSMC, and ADPFT
Theorem 3. For practical systems in (7), if two control laws are designed schemes, respectively. Figs. 1 and 2 illustrate the tracking errors of the
as follows: quadrotor. It can be seen that the tracking errors can converge into a
1 small region when 𝑡 = 5 s, which indicates that the proposed fixed-time
𝝂 𝜁2 = 𝝂 𝜁𝑛 − 𝒅̂ 𝜁 − 𝜾𝜁 , (26)
2 RBFNN observers possess fine performance in estimating lumped dis-
( ) turbances. The nominal control strategies can guarantee the stability of
where variable vectors 𝜾𝜁 are expressed as 𝜾𝜁 = 𝜛𝜁 𝒒 †𝜁 erf ∇𝐽𝜁 ∕𝜚𝜁 , with
the system when 𝑡 > 5 s. The 3D tracking results shown in Fig. 3 are em-
𝜛𝜁 denoting the positive constants. Then, the tracking errors 𝝌 𝜁 in practical
ployed to describe the tracking effect more vividly. From Figs. 1 and 2,
systems (7) and weight errors 𝝎̃ 𝜁 are UUB stable.
one can see that the numerical values of the red curves are the smallest,
which indicates that the proposed ADPNN possesses the highest accu-
Proof. See Subsection 6.3 for details. racy compared with PIDSMC and ADPFT.
Figs. 4 and 5 show the time responses of the observation errors.
Remark 1. Since the approximation errors the critic ( NNs
) cannot con- The NN observers proposed in both ADPNN and ADPFT schemes can
verge at the origin, it is almost impossible for ‖∇𝛿𝜁 𝝌 𝜁 ‖ = 0 in Theo- estimate lumped disturbances effectively. But the observers in ADPNN
rem 1 to hold. However, once the above conditions hold, the bounds of possess higher accuracy, with 𝑑̃w1 –𝑑̃w3 being less than 4 × 10−3 N and
approximation errors cannot be obtained. To solve this problem, switch- 𝑑̃z1 –𝑑̃z3 being less than 4 × 10−4 N ⋅ m. It is evident that the numeri-
ing quantities are designed. To be specific, if the conditions hold, the cal values of the ADPFT scheme are greater than the proposed ADPNN
bounds of approximation
( ) errors can be obtained. On the other hand scheme. Therefore, the RBFNN observers proposed in ADPNN can per-
(e.g., ‖∇𝛿𝜁 𝝌 𝜁 ‖ ≠ 0), the approximation errors can converge simulta- form better in observation accuracy compared with the observer in
neously with the convergence of the system states. ADPFT. The approximation results of the critic NNs are depicted in
( ) Fig. 6. It can be found that all the curves of critic NN weights tend to
Remark 2. The well-designed continuous terms erf ∇𝐽𝜁 ∕𝜚𝜁 are em- remain stable with almost no fluctuation after 5 s. This is because actua-
ployed to attenuate the system impact because of the observation errors. tor failures have been compensated by the RBFNN observers. As shown
Superior to the continuous robust term constructed by tanh (⋅), erf (⋅) in Fig. 7, the red curves fluctuate faster than both the purple and green

4
B. Li, H. Liu, C.K. Ahn et al. Aerospace Science and Technology 144 (2024) 108803

Fig. 1. Tracking error results of the position loop driven by different schemes. (For interpretation of the colors in the figure(s), the reader is referred to the web
version of this article.)

Fig. 2. Tracking error results of the attitude loop driven by different schemes.
( )
∞
curves when t = 15, 20, 25, 35 s. This is because the proposed ADPNN ∫0 𝝌 T𝜁 𝑸𝜁 𝝌 𝜁 + 𝝂 T𝜁 𝑹𝜁 𝝂 𝜁 𝑑𝜗. Notably, 𝝂 𝜁 are the total control laws
can respond quickly to actuator failures under the fixed-time RBFNN belonging to the two subsystems, rather than the nominal control strate-
observers. Moreover, the values of the purple curves in the initial period gies. The histograms in Fig. 8 illustrate the comparison results of control
are much larger than the red and green curves, demonstrating that tra- cost driven by ADPNN and PIDSMC. One can see that the total cost
ditional control frameworks (i.e., such as PIDSMC) may result in higher values under the proposed ADPNN are smaller than those under the
control cost. PIDSMC used for comparison quantitatively. Moreover, the cost val-
While ADPNN and ADPFT based on optimal control methods can ues of the position loop are much larger than the attitude loop; this
make a trade-off between control cost and control performance, the is because the control inputs shown in Fig. 7 of the position loop are
comparison in cost reduction between ADPNN and ADPFT is not imple- more significant. To be specific, the cost values 𝐽𝜁 under the proposed
mented herein. To highlight the scheme’s compelling performances in ADPNN are smaller in the interval 0–15 s, while they are almost equal
( )
reducing control cost, let us define two novel cost functions 𝐽𝜁 𝝌 𝜻 = in the other two intervals (e.g., 15–30 s and 30–50 s) under different

5
B. Li, H. Liu, C.K. Ahn et al. Aerospace Science and Technology 144 (2024) 108803

𝜅𝜁 1 𝜖̌𝜁2𝐻 𝜅𝜁 2 𝛿̌𝜁 𝜖̌𝜁2 ( )

where 𝜖𝜁𝑑 = ( )2 + ‖𝝎𝜁 ‖2 + 𝜆max 𝑨𝜁 . Define 𝑪 𝜁 =
2 2
2 1+𝜸 T 𝜸
𝜁 𝜁
𝜅𝜁 1 𝜸 𝜁 𝜸 T
𝜁
𝜅𝜁2 𝜑 𝑰 6 + ( )2 − 𝑮 𝜁 , then
𝜁
1+𝜸 T 𝜁
𝜸 𝜁

1 T
𝑉̇ 𝜁1 ≤ −𝝌 T𝜁 𝑸𝜁 𝝌 𝜁 − 𝝎 ̃ + 𝜖𝜁𝑑 .
̃ 𝑪 𝝎 (29)
2 𝜁 𝜁 𝜁
By referring to Assumption 1, it can be concluded that ‖𝑮𝜁 ‖ is
bounded, such that ‖𝑮𝜁 ‖ ≤ 𝐺̌ 𝜁 . Also, if 𝜆min (𝑪 𝜁 ) > 0 can be satisfied
by selecting appropriate parameters 𝜅𝜁1 and 𝜅𝜁2 , then 𝝌 𝜁 and 𝝎 ̃ 𝜁 are
guaranteed to achieve UUB. Therefore, the proposed novel weight up-
date laws demonstrably relax the common PE condition.

6.2. Proof of Theorem 2

Fig. 3. 3D tracking results driven by different schemes. Consider Lyapunov functions as follows:

( 3 )
1∑ 1 ∑ ̃ T ̃
6
control schemes, as shown in Fig. 8. Because the control inputs of the 1
𝑉𝜁2 = 𝜁1 + 𝜁2 = 𝒆T𝜁 𝒆𝜁 + 𝑊 𝑊 . (30)
initial stage under the proposed ADPNN are at a higher level, the system 2 2 𝑓 =1 𝑎𝜁 𝑡=1 𝜁𝑓 𝑡 𝜁𝑓 𝑡
error states can converge into a small neighborhood of the origin. From
If positive constants 𝜆𝜁1 satisfy 𝜆𝜁1 > 0.5, then the time derivative
the above discussions and analysis, we can conclude that the proposed
of 𝜁 can be obtained in the following form:
ADPNN yields superior performance in striking a balance between con-
trol cost and control performance when the quadrotor is exposed to
( ) 𝑞̌
2
actuator failures. ( )
̃ T 𝜎𝜁 𝝌 𝜁 + 𝜁 ‖𝜀𝜁 𝝌 𝜁 ‖2 .
3
̇ 𝜁1 ≤ −𝜆𝜁2 ‖𝒆𝜁 ‖ 2 − 𝜆𝜁3 ‖𝒆𝜁 ‖4 − 𝒆T𝜁 𝒒 𝜁 𝑾 𝜁 2
5. Conclusion (31)

This work developed an optimized tracking control scheme for a Taking the time derivative of 𝜁2 and according to Lemma 2 in [38],
quadrotor UAV with actuator failures, which ensures UUB stability of the following inequalities hold:
the control system. The critic-only structure was utilized to learn the
6 (∑
∑ 3 (
nominal control strategies through the proposed critic NN-based update ( ) 𝑏 𝑏
̇ 𝜁2 ≤ ̃ T 𝜎𝜁𝑓 𝝌 𝜁 𝑒T 𝑞𝜁𝑓 𝑡 − 𝜁 ‖𝑊
𝑊 ̃ 𝜁𝑓 𝑡 ‖2 + 𝜁 ‖𝑊𝜁𝑓 𝑡 ‖2
laws without the PE condition. On this basis, two RBFNN observers were 𝜁𝑓 𝑡 𝜁𝑓 2 2
𝑓 =1 𝑡=1
designed to reconstruct lumped disturbances while guaranteeing fixed-
4 4
time convergence characteristics. The observers feature upper bounds 3𝑐𝜁 𝑠 3 𝑐𝜁 9𝑐𝜁 𝑠 3
on the RBFNN weight errors that no longer rely on their ideal weights. + ‖𝑊𝜁𝑓 𝑡 ‖ + 4 ̃ 𝜁𝑓 𝑡 ‖4 +
‖𝑊 ̃ 𝜁𝑓 𝑡 ‖4
‖𝑊
4 4𝑠4 4
Two specifically designed continuous variable vectors with the intro-
duction of the Gaussian error function were employed to enhance the 3𝑐𝜁 𝛿̌𝜁 ))
̃ 𝜁𝑓 𝑡 ‖4 +
− 𝑐𝜁 ‖𝑊 ‖𝑊𝜁𝑓 𝑡 ‖4
system’s control performance. A comparison of the simulation results 4𝑠4
demonstrated the effectiveness of the proposed scheme in reducing 6 (∑
3 (
∑ ( ) 𝑏 𝑏
̃ T 𝜎𝜁𝑓 𝝌 𝜁 𝑒T 𝑞𝜁𝑓 𝑡 − 𝜁 ‖𝑊
̃ 𝜁𝑓 𝑡 ‖2 − 𝜁 ‖𝑊
3
control cost. In the future, the authors will focus on improving their ≤ 𝑊 ̃ 𝜁𝑓 𝑡 ‖ 2
𝜁𝑓 𝑡 𝜁𝑓 8 2
research on fixed-time ADP and its applications to game-based control 𝑓 =1 𝑡=1
frameworks.
4 4
( 9𝑐𝜁 𝑠 3 )
𝑐𝜁 ( 3𝑐𝜁 𝑠 3 3𝑐𝜁 )
6. Appendix − 𝑐𝜁 − − ̃ 𝜁𝑓 𝑡 ‖4 +
‖𝑊 + ‖𝑊𝜁𝑓 𝑡 ‖4
4𝑠 4 4 4 4𝑠4
𝑏𝜁 ) 𝑏𝜁 )
6.1. Proof of Theorem 1
+ ‖𝑊𝜁𝑓 𝑡 ‖2 + . (32)
2 8
Choose Lyapunov functions as follows: Referring to Lemmas 3 and 4 in [39], and substituting (31) and (32)
1 T back into the time derivatives of (30), the following inequalities can be
𝑉𝜁1 = 𝐽𝜁∗ (𝝌 𝜁 ) + ̃ 𝝎
𝝎 ̃ . (27)
2 𝜁 𝜁 obtained:
Notably, 𝑮𝜁 = ∇𝛿𝜁 𝒒 𝜁 𝑹−1
𝜁 𝒒 𝜁 ∇ 𝛿𝜁 and 𝑨𝜁 = 𝒒 𝜁 𝑹 𝜁 𝒒 𝜁 , with 𝑮𝜁 and 𝑨𝜁
T T −1 T 3
𝑉̇ 𝜁2 ≤ −𝜆𝜁2 ‖𝒆𝜁 ‖ 2 − 𝜆𝜁3 ‖𝒆𝜁 ‖4
being symmetric non-negative matrices. The time derivative of (27) is
6 (∑
∑ 3 (
yielded as: 𝑏𝜁 𝑏𝜁 3
+ − ̃ 𝜁𝑓 𝑡 ‖2 −
‖𝑊 ̃ 𝜁𝑓 𝑡 ‖ 2
‖𝑊
8 2
𝑉̇ 𝜁1 = −𝝌 T𝜁 𝑸𝜁 𝝌 𝜁 − 𝝂 ∗𝜁𝑛 T 𝑹𝜁 𝝂 ∗𝜁𝑛 𝑓 =1 𝑡=1

( ) ( 1 −1 T T (
4
+ 𝝎T ∇𝛿𝜁 + ∇T 𝜖𝜁 𝒒 𝜁 𝑹 𝒒 ∇ 𝛿𝜁 𝝎𝜁 𝑐𝜁 9𝑐𝜁 𝑠 3 ) ))
2 𝜁 𝜁 − 𝑐𝜁 − − ̃ 𝜁𝑓 𝑡 ‖4 + 𝜗𝜁
‖𝑊
) 4𝑠4 4
1 1
− 𝑹−1 𝜁 𝒒 T𝜁 ∇T 𝛿𝜁 𝝎̂ 𝜁 + 𝑹−1 𝜁 𝒒 T𝜁 ∇𝜖𝜁 + 𝝎 ̂̇ 𝜁
̃ T𝜁 𝝎 3
2 2
≤ −𝛼̌ 𝜁 𝑉𝜁 4 − 𝛽̌𝜁 𝑉𝜁2 + 𝜗𝜁 , (33)
1 T
≤ −𝝌 T𝜁 𝑸𝜁 𝝌 𝜁 + 𝝎 ̃ 𝑮 𝝎 ̃
2 𝜁 𝜁 𝜁 where the expressions of 𝛼̌ 𝜁 , 𝛽̌𝜁 , and 𝜗𝜁 are
( 𝜅𝜁1 𝜸 𝜁 𝜸 T𝜁 𝜅𝜁2 )
−𝝎̃ T𝜁 + 𝜑 𝑰6 𝝎 ̃ 𝜁 + 𝜖𝜁𝑑 , (28) { 3
}
3 1
2(1 + 𝜸 𝜁 𝜸 𝜁 )
T 2 𝜁
2
𝛼̌ 𝜁 = min 2 4 𝜆𝜁2 , 2− 4 𝑎𝜁4 𝑏𝜁 ,

6
B. Li, H. Liu, C.K. Ahn et al. Aerospace Science and Technology 144 (2024) 108803

Fig. 4. Observation error results of the position loop driven by different RBFNN observers.

Fig. 5. Observation error results of the attitude loop driven by different RBFNN observers.

Fig. 6. Weight results of the critic NNs driven by ADPNN.

⎧ ⎛ 4
⎞⎫ 𝜗𝜁 4 1
𝑎2𝜁 ⎜ 𝑐𝜁 9𝑐𝜁 𝑠𝜁3 with 𝑔𝜁 = in fixed time 𝑡𝜁𝑢 ≤ + 0 < 𝜃𝜁0 < 1. There-
,
⎪ ⎟⎪ 1−𝜃𝜁 0 𝛼̌ 𝜁 𝜃𝜁 0 𝛽̌𝜁 𝜃𝜁 0
𝛽̌𝜁 = min ⎨4𝜆𝜁3 , 𝑐 − √
9 ⎜⎜ 𝜁 4𝑠4 ⎟⎬
− ̃ 𝜁𝑓 𝑡 are calculated by ‖𝒆𝜁 ‖ ≤ 2𝜗𝜁
4 fore, the upper bounds of 𝒆𝜁 and 𝑊
⎪ 𝜁 ⎟⎪ √
⎩ ⎝ ⎠⎭ and ‖𝑊 ̃ 𝜁𝑓 𝑡 ‖ ≤ 2𝑎𝜁 𝜗𝜁 . Then, ‖𝒅̃ 𝜁 ‖ ≤ 𝜎̌ 𝜁 ‖𝑾
̃ ‖ + 𝜀̌ 𝜁 can be further
T
𝜁
and obtained. According to fixed-time convergence characteristics of the
RBFNN weight approximation errors 𝑾 ̃ 𝜁 , the observation errors 𝒅̃ 𝜁 can
4
converge into the upper bounds of ‖𝒅̃ 𝜁 ‖ in fixed time. This completes
6 (∑
3 (( 3𝑐 𝑠 3
∑ 𝜁 𝜁 3𝑐𝜁 ) 𝑏𝜁 ) 𝑏𝜁 )
the proof of Theorem 2.
𝜗𝜁 = + ‖𝑊𝜁𝑓 𝑡 ‖4 + ‖𝑊𝜁𝑓 𝑡 ‖2 +
4 4
4𝑠𝜁 2 8
𝑓 =1 𝑡=1
6.3. Proof of Theorem 3
𝑞̌𝜁 ( ) Consider Lyapunov functions 𝑉𝜁3 as follows:
+ ‖𝜀T𝜁 𝝌 𝜁 ‖2 .
2 ( ) 1 T
According to Lemma 2 in [4], one can obtain that ̃ 𝝎
𝑉𝜁3 = 𝐽𝜁∗ 𝝌 𝜁 + 𝝎 ̃ . (34)
2 𝜁 𝜁
For simplification, let 𝜶 𝜁 = 𝒒 𝜁 𝒅̃ 𝜁 . It can be reasonably assumed that
⎧ ⎧( ) 4 ( ) 1 ⎫⎫
⎪( ) ( ) ⎪ 𝑔𝜁 3 𝑔𝜁 2 ⎪⎪ there exist positive constants 𝑘𝜁 satisfying ‖𝜶 𝜁 ‖ ≤ 𝑘𝜁 , according to As-
̃ ̃
⎨ 𝒆𝜁 , 𝑊𝜁𝑓 𝑡 ∣ 𝑉𝜁 𝒆𝜁 , 𝑊𝜁𝑓 𝑡 ≤ min ⎨ 𝛼̌ , ⎬⎬ sumption 1, Claim in [40] and Eq. (3.1) in [35]. The time derivative of
⎪ ⎪ 𝜁 𝛽̌𝜁 ⎪⎪
⎩ ⎩ ⎭⎭ (34) yields:

7
B. Li, H. Liu, C.K. Ahn et al. Aerospace Science and Technology 144 (2024) 108803

Fig. 7. Illustration of the four control inputs driven by different schemes.

Fig. 8. The results of control cost driven by ADPNN and PIDSMC.

1 1 ( ) 𝜅𝜁 1 𝜸 𝜁 𝜸 T
𝑉̇ 𝜁3 ≤ −𝝌 T𝜁 𝑸𝜁 𝝌 𝜁 + ∇T 𝐽𝜁∗ 𝑨𝜁 ∇T 𝐽𝜁∗ − ∇T 𝐽𝜁∗ 𝑨𝜁 ∇T 𝛿𝜁 𝝎𝜁 + 𝝎
̃𝜁 tanh (⋅) function. Additionally, the expressions 𝑫 𝜁 are 𝑫 𝜁 = 𝜁
+
4 4 2(1+𝜸 T 𝜸 )2
𝜁 𝜁
( ) ( ) 𝜅𝜁 2 𝜑
𝜁 𝑮𝜁
+ ∇T 𝐽𝜁∗ 𝜶 𝜁 − 𝜛𝜁 ∇T 𝐽𝜁 tanh ∇𝐽𝜁 ∕𝜚𝜁 + 𝜛𝜁 ∇T 𝐽𝜁 erf ∇𝐽𝜁 ∕𝜚𝜁 𝑰6 − .
2 4
[ ]T
( ) ̃ 𝜁‖ .
Assuming that there exist argument vectors 𝚵𝜁 = ‖𝝌 𝜁 ‖, ‖𝝎
− 𝜛𝜁 ∇T 𝐽𝜁∗ erf ∇𝐽𝜁 ∕𝜚𝜁 +𝝎 ̂̇ 𝜁
̃ T𝜁 𝝎 The compact form of (35) can be expressed as the following inequal-
ities
≤ −𝝌 T𝜁 𝑸𝜁 𝝌 𝜁 − 𝝎 ̃ 𝜁 + 𝑃𝜁 ‖𝝎
̃ T𝜁 𝑫 𝜁 𝝎 ̃ 𝜁 ‖ + 𝜖𝜁𝑧 , (35)

( ) ( ) ̌ 𝜁 𝚵𝜁 + 𝑻 𝜁 𝚵𝜁 + 𝜖𝜁𝑧 ,
𝑉̇ 𝜁3 ≤ −𝚵𝑇𝜁 𝑫 (36)
𝜅𝜁 1 𝜖̌𝜁2𝐻 𝜅𝜁 2 𝛿̌𝜁 𝜖̌𝜁2
where 𝜖𝜁𝑧 = ( )2 + 2
‖𝝎𝜁 ‖2 + 2
𝜆max 𝑨𝜁 + 𝜛𝜁 𝜖̌𝜁 1 + 𝐽̌𝜁 + ( ) ( )
2 1+𝜸 T 𝜁 𝜁
𝜸 with the terms 𝑫 ̌ 𝜁 , 𝑻 𝜁 expressed by 𝑫̌ 𝜁 = diag{𝜆min 𝑸𝜁 , 𝜆min 𝑫 𝜁 }
( ) ( )
𝑌𝜁 and 𝑃𝜁 = 𝜛𝜁 𝛿̌𝜁 1 + 𝐽̌𝜁 with 𝐽̌𝜁 = ‖erf ∇𝐽𝜁 ∕𝜚𝜁 ‖. Moreover, and 𝑻 𝜁 = [0, 𝑃𝜁 ], respectively. It can be reasonably assumed that there
𝑌𝜁 are positive constants generated by the inequality related to the are positive constants 𝑇̄𝜁 such that ‖𝑻 𝜁 ‖ ≤ 𝑇̄𝜁 . Further, one has:

8
B. Li, H. Liu, C.K. Ahn et al. Aerospace Science and Technology 144 (2024) 108803

( )
̌ 𝜁 ‖𝚵𝜁 ‖2 + 𝑇̄𝜁 ‖𝚵𝜁 ‖ + 𝜖𝜁𝑧 .
𝑉̇ 𝜁3 ≤ −𝜆min 𝑫 (37) [15] Q. Fan, G. Yang, Adaptive actor-critic design-based integral sliding-mode control for
partially unknown nonlinear systems with input disturbances, IEEE Trans. Neural
Then, 𝑉̇ 𝜁3 < 0 hold when the inequalities Netw. Learn. Syst. 27 (1) (2016) 165–177.
√ ( )
[16] M. Ran, J. Li, L. Xie, Reinforcement-learning-based disturbance rejection control for
𝑇̄𝜁 + ̌ 𝜁 𝜖𝜁𝑧
𝑇̄𝜁2 + 4𝜆min 𝑫 uncertain nonlinear systems, IEEE Trans. Cybern. 52 (9) (2022) 9621–9633.
[17] Z. Guo, D. Yao, W. Bai, H. Li, R. Lu, Event-triggered guaranteed cost fault-tolerant
‖𝚵𝜁 ‖ > ( )
2𝜆min 𝑫 ̌𝜁 optimal tracking control for uncertain nonlinear system via adaptive dynamic pro-
gramming, Int. J. Robust Nonlinear Control 31 (7) (2021) 2572–2592.
are satisfied. This completes the proof. [18] X. Han, Z. Zheng, L. Liu, B. Wang, Z. Cheng, H. Fan, Y. Wang, Online policy iteration
ADP-based attitude-tracking control for hypersonic vehicles, Aerosp. Sci. Technol.
106 (2020) 106233.
Declaration of competing interest [19] L. Dou, X. Su, X. Zhao, Q. Zong, L. He, Robust tracking control of quadrotor via
on-policy adaptive dynamic programming, Int. J. Robust Nonlinear Control 31 (7)
The authors declare that they have no known competing financial (2021) 2509–2525.
interests or personal relationships that could have appeared to influence [20] H. Liu, B. Li, B. Xiao, D. Ran, C. Zhang, Reinforcement learning-based tracking con-
trol for a quadrotor unmanned aerial vehicle under external disturbances, Int. J.
the work reported in this paper.
Robust Nonlinear Control 33 (17) (2023) 10360–10377, https://doi.org/10.1002/
rnc.6334.
Data availability [21] C. Mu, Y. Zhang, Learning-based robust tracking control of quadrotor with time-
varying and coupling uncertainties, IEEE Trans. Neural Netw. Learn. Syst. 31 (1)
(2020) 259–273.
No data was used for the research described in the article.
[22] B. Li, H. Zhang, B. Xiao, C. Wang, Y. Yang, Fixed-time integral sliding mode control
of a high-order nonlinear system, Nonlinear Dyn. 107 (1) (2022) 909–920.
Acknowledgements [23] Y. Lian, J. Xia, J.H. Park, W. Sun, H. Shen, Disturbance observer-based adaptive neu-
ral network output feedback control for uncertain nonlinear systems, IEEE Trans.
This work was partly supported by National Natural Science Foun- Neural Netw. Learn. Syst. 34 (10) (2023) 7260–7270, https://doi.org/10.1109/
TNNLS.2021.3140106.
dation of China (62073212), Natural Science Foundation of Shang-
[24] B. Xiao, X. Wu, L. Cao, X. Hu, Prescribed time attitude tracking control of space-
hai (23ZR1426600), Innovation Fund of Chinese Universities Industry- craft with arbitrary disturbance, IEEE Trans. Aerosp. Electron. Syst. 58 (3) (2022)
University-Research (2021ZYB05004), Training Programme for the Top 2531–2540.
Innovative Talents of Shanghai Maritime University (2022YBR015), [25] B. Li, Q. Hu, Y. Yang, Continuous finite-time extended state observer based fault
tolerant control for attitude stabilization, Aerosp. Sci. Technol. 84 (2019) 204–213.
and the National Research Foundation of Korea (NRF) grant funded
[26] X. Zheng, H. Li, C.K. Ahn, D. Yao, NN-based fixed-time attitude tracking control
by the Korea government (Ministry of Science and ICT) (No. NRF- for multiple unmanned aerial vehicles with nonlinear faults, IEEE Trans. Aerosp.
2020R1A2C1005449). Electron. Syst. 59 (2) (2023) 1738–1748.
[27] W. Gong, B. Li, C.K. Ahn, Y. Yang, Prescribed-time extended state observer and
References prescribed performance control of quadrotor UAVs against actuator faults, Aerosp.
Sci. Technol. 138 (2023) 108322, https://doi.org/10.1016/j.ast.2023.108322.
[28] X. Bu, M. Lv, H. Lei, J. Cao, Fuzzy neural pseudo control with prescribed perfor-
[1] D.N. Das, R. Sewani, J. Wang, M.K. Tiwari, Synchronized truck and drone rout-
mance for waverider vehicles: a fragility-avoidance approach, IEEE Trans. Cybern.
ing in package delivery logistics, IEEE Trans. Intell. Transp. Syst. 22 (9) (2021)
53 (8) (2023) 4986–4999.
5772–5782.
[29] Y. Li, Y. Liu, S. Tong, Observer-based neuro-adaptive optimized control of strict-
[2] H.H. Kang, D.K. Lee, C.K. Ahn, Finite-memory-structured online training algo-
feedback nonlinear systems with state constraints, IEEE Trans. Neural Netw. Learn.
rithm for system identification of unmanned aerial vehicles with neural networks,
Syst. 33 (7) (2022) 3131–3145.
IEEE/ASME Trans. Mechatron. 27 (6,2) (2022) 5846–5856.
[30] L. Yuan, L. Wang, J. Xu, Adaptive fault-tolerant controller for morphing aircraft
[3] S. Abujoub, J. McPhee, R.A. Irani, Methodologies for landing autonomous aerial
based on the L2 gain and a neural network, Aerosp. Sci. Technol. 132 (2023)
vehicles on maritime vessels, Aerosp. Sci. Technol. 106 (2020) 106169.
107985.
[4] B. Li, W. Gong, Y. Yang, B. Xiao, D. Ran, Appointed fixed time observer-based sliding
[31] G. Lin, H. Li, C.K. Ahn, D. Yao, Event-based finite-time neural control for human-in-
mode control for a quadrotor UAV under external disturbances, IEEE Trans. Aerosp.
the-loop UAV attitude systems, IEEE Trans. Neural Netw. Learn. Syst. 34 (12) (2023)
Electron. Syst. 58 (1) (2022) 290–303.
10387–10397, https://doi.org/10.1109/TNNLS.2022.3166531.
[5] J. Wang, K.A. Alattas, Y. Bouteraa, O. Mofid, S. Mobayen, Adaptive finite-time back-
[32] H. Liu, X. Huang, X. Tian, J. Yuan, Fixed-time self-structuring neural network coop-
stepping control tracker for quadrotor UAV with model uncertainty and external
erative tracking control of multi-robot systems with actuator faults, Int. J. Robust
disturbance, Aerosp. Sci. Technol. 133 (2023) 108088, https://doi.org/10.1016/j.
Nonlinear Control 32 (16) (2022) 9053–9076.
ast.2022.108088.
[33] S. Zhang, S. Yuan, X. Yu, L. Kong, Q. Li, G. Li, Adaptive neural network fixed-time
[6] X. Bu, B. Jiang, H. Lei, Performance guaranteed finite-time non-affine control of
control design for bilateral teleoperation with time delay, IEEE Trans. Cybern. 52 (9)
waverider vehicles without function-approximation, IEEE Trans. Intell. Transp. Syst.
(2022) 9756–9769.
24 (3) (2023) 3252–3262.
[34] B. Xiao, S. Yin, A new disturbance attenuation control scheme for quadrotor un-
[7] L. Cao, B. Xiao, M. Golestani, D. Ran, Faster fixed-time control of flexible spacecraft
manned aerial vehicles, IEEE Trans. Ind. Inform. 13 (6) (2017) 2922–2932.
attitude stabilization, IEEE Trans. Ind. Inform. 16 (2) (2020) 1281–1290.
[35] Y.J. Bagul, C. Chesneau, Sigmoid functions for the smooth approximation to the
[8] Y. Li, X. Min, S. Tong, Observer-based fuzzy adaptive inverse optimal output feed-
absolute value function, Moroccan J. Pure Appl. Anal. 7 (1) (2021) 12–19, https://
back control for uncertain nonlinear systems, IEEE Trans. Fuzzy Syst. 29 (6) (2021)
doi.org/10.2478/mjpaa-2021-0002.
1484–1495.
[36] J. Sun, C. Liu, Disturbance observer-based robust missile autopilot design with full-
[9] K.G. Vamvoudakis, F.L. Lewis, Online actor-critic algorithm to solve the continuous-
state constraints via adaptive dynamic programming, J. Franklin Inst. Eng. Appl.
time infinite horizon optimal control problem, Automatica 46 (5) (2010) 878–888.
Math. 355 (5) (2018) 2344–2368.
[10] H. Dong, X. Zhao, H. Yang, Reinforcement learning-based approximate optimal con-
[37] M. Van, M. Mavrovouniotis, S.S. Ge, An adaptive backstepping nonsingular fast ter-
trol for attitude reorientation under state constraints, IEEE Trans. Control Syst.
minal sliding mode control for robust fault tolerant control of robot manipulators,
Technol. 29 (4) (2021) 1664–1673.
IEEE Trans. Syst. Man Cybern. Syst. 49 (7) (2019) 1448–1458.
[11] Q. Hu, H. Yang, H. Dong, X. Zhao, Learning-based 6-DOF control for autonomous
[38] H.Q. Wang, B. Chen, C. Lin, Adaptive neural tracking control for a class of stochastic
proximity operations under motion constraints, IEEE Trans. Aerosp. Electron. Syst.
nonlinear systems, Int. J. Robust Nonlinear Control 24 (7) (2014) 1262–1280.
57 (6) (2021) 4097–4109.
[39] Z. Zuo, B. Tian, M. Defoort, Z. Ding, Fixed-time consensus tracking for multiagent
[12] G. Wen, C.L.P. Chen, S.S. Ge, Simplified optimized backstepping control for a class
systems with high-order integrator dynamics, IEEE Trans. Autom. Control 63 (2)
of nonlinear strict-feedback systems with unknown dynamic functions, IEEE Trans.
(2018) 563–570.
Cybern. 51 (9) (2021) 4567–4580.
[40] M. Polycarpou, P. Ioannou, A robust adaptive nonlinear control design, Automatica
[13] D. Wang, D. Liu, H. Li, Policy iteration algorithm for online design of robust control
32 (3) (1996) 423–427.
for a class of continuous-time nonlinear systems, IEEE Trans. Autom. Sci. Eng. 11 (2)
(2014) 627–632.
[14] H. Yang, Q. Hu, H. Dong, X. Zhao, ADP-based spacecraft attitude control under
actuator misalignment and pointing constraints, IEEE Trans. Ind. Electron. 69 (9)
(2022) 9342–9352.