IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 11, NO.

10, OCTOBER 2024 2099

Fuzzy-Model-Based Finite Frequency Fault

Detection Filtering Design for Two-
Dimensional Nonlinear Systems
Meng Wang , Huaicheng Yan , Jianbin Qiu , Fellow, IEEE, and Wenqiang Ji

Abstract—This article studies the fault detection filtering design approach is to design an evaluation generator to produce a
problem for Roesser type two-dimensional (2-D) nonlinear sys- residual signal which is sensitive to faults and robust to exter-
tems described by uncertain 2-D Takagi-Sugeno (T-S) fuzzy mod-
els. Firstly, fuzzy Lyapunov functions are constructed and the 2-D
nal disturbances. Based on the residual signal, an evaluation
Fourier transform is exploited, based on which a finite frequency function with a threshold is designed and the faut is detected
fault detection filtering design method is proposed such that a once the evaluation function exceeds the threshold [7], [8]. On
residual signal is generated with robustness to external distur- the other hand, many practical processes, such as video data
bances and sensitivity to faults. It has been shown that the utiliza- processing and circuit systems, involve not only the time vari-
tion of available frequency spectrum information of faults and able but also the spatial variable and thus can not be well
disturbances makes the proposed filtering design method more
general and less conservative compared with a conventional non- described by one-dimensional (1-D) systems [9], [10]. In
frequency based filtering design approach. Then, with the pro- order to study this kind of complex processes, 2-D systems
posed evaluation function and its threshold, a novel mixed finite have been proposed [11]–[13] and many elegant results have
frequency H∞ /H− fault detection algorithm is developed, based been reported over the past few decades [14]–[17]. In [14], a
on which the fault can be immediately detected once the evalua- novel diagonally dominant matrix method was proposed, and
tion function exceeds the threshold. Finally, it is verified with sim-
less conservative stochastic stability analysis results were
ulation studies that the proposed method is effective and less con-
servative than conventional non-frequency and/or common Lya- obtained for 2-D Roesser systems with state-multiplicative
punov function based filtering design methods. noises. Based on piecewise Lyapunov functions together with
Index Terms—Fault diagnosis, finite frequency specifications,
the extended average dwell time technique, the authors in [15]
mixed H∞ /H− performance, two-dimensional nonlinear systems. derived stability analysis conditions for 2-D switched systems
under arbitrary and restricted switchings, respectively. Then,
I. Introduction an elegant linearization technique was applied and controller
ANY practical control systems suffer from inevitable design approach was obtained for 2-D switched systems. Due
M faults and/or external disturbances, which may degrade
system performance or even cause instability of the system
to their recursive form, the filters can be updated online and
produce real-time state estimation. Note that most of the exist-
[1]–[6]. The main goal of an admissible fault detection ing results [11]–[15] on 2-D systems do not consider the fact
that faults may exist. Since there is a great demand for safety
Manuscript received March 1, 2024; accepted March 26, 2024. This work
was supported in part by the National Natural Science Foundation of China
and reliability in many real-world applications, it is of much
(62373152, 62333005, U21B6001, 62073143, 62273121), in part by the significance to tackle the problem of fault detection for 2-D
Natural Science Funds for Excellent Young Scholars of Hebei Province in systems.
2022 (F2022202014), in part by Science and Technology Research Project of
Colleges and Universities in Hebei Province (BJ2020017), and in part by the Recently, few works have been reported on the issue of fault
China Postdoctoral Science Foundation (2022M711639, 2023T160320). detection for 2-D systems [18]–[20]. For example, in [18], the
Recommended by Associate Editor Peter Liu. (Corresponding author: Meng issues of fault detection, isolation, and fault tolerant control
Wang.)
for linear 2-D Roesser systems under communication
Citation: M. Wang, H. Yan, J. Qiu, and W. Ji, “Fuzzy-model-based finite
frequency fault detection filtering design for two-dimensional nonlinear resources limited network environment were investigated. By
systems,” IEEE/CAA J. Autom. Sinica, vol. 11, no. 10, pp. 2099–2110, Oct. introducing an event-triggered mechanism, a unified solution
2024.
for the robust fault detection/isolation observer and controller
M. Wang and H. Yan are with the Key Laboratory of Smart Manufacturing
in Energy Chemical Process, Ministry of Education, School of Information design was developed and network communication resources
Science and Engineering, East China University of Science and Technology, were saved. In [19], [20], a class of positive 2-D systems was
Shanghai 200237, China (e-mail: [email protected]; hcyan@ecust. considered, in which the system state is always nonnegative.
edu.cn).
J. Qiu is with the Research Institute of Intelligent Control and Systems,
Most of the results have been developed for 2-D linear sys-
School of Astronautics, Harbin Institute of Technology, Harbin 150080, tems in existing works [18]–[20] and few results have
China (e-mail: [email protected]). addressed the investigation of 2-D nonlinear systems. There
W. Ji is with the School of Artificial Intelligence, Hebei University of widely exist nonlinearities in many real dynamics including
Technology, Tianjin 300401, China (e-mail: [email protected]).
Color versions of one or more of the figures in this paper are available
those 2-D systems. Therefore, it is of utmost importance to
online at http://ieeexplore.ieee.org. study 2-D nonlinear systems.
Digital Object Identifier 10.1109/JAS.2024.124452 On the other hand, T-S fuzzy models have been well recog-
2100 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 11, NO. 10, OCTOBER 2024

nized as a powerful nonlinear function approximation tool. constraints [40]. For 2-D nonlinear systems, to the best of the
Many elegant results on T-S fuzzy model based control, filter- authors’ knowledge, few works have been reported in the
ing, and fault detection problems have been obtained for 1-D open literature to tackle the problem of finite frequency per-
nonlinear systems [21]–[29]. In [27], an adaptive compen- formance analysis and synthesis. This observation motivates
sator was introduced to eliminate the effect of Lipschitz non- the study of finite frequency fault detection filtering design for
linear terms caused by unmeasurable premise variables of 2-D nonlinear systems.
fuzzy systems. In [28], the unmeasurable premise variables of In this article, we tackle the fault detection filtering design
fuzzy systems were estimated. And then, novel fault detection problem of 2-D nonlinear systems in finite frequency domain.
observer design methods were developed in [27], [28] to Using the 2-D Fourier transform and 2-D Parseval’s relation
simultaneously reconstruct the sensor fault and the system together with fuzzy Lyapunov functions, an admissible finite
state. Due to their success in the study of 1-D nonlinear sys- frequency fault detection filtering design method is proposed
tems, T-S fuzzy models have been further generalized to to generate a residual signal with the features of sensitivity to
investigate more complex 2-D nonlinear systems in recent faults and robustness to external disturbances in given fre-
years [30]–[32]. For example, based on T-S fuzzy affine mod- quency ranges. Then, with the proposed evaluation function
els, the output feedback sliding mode control problem for 2-D and its threshold, a novel fault detection algorithm is given,
nonlinear systems was investigated in [30]. In [31], using based on which the fault can be immediately detected once the
stochastic analysis techniques and basis-dependent Lyapunov evaluation function exceeds the threshold. Finally, it is veri-
functions, the authors proposed a fuzzy model based fault fied with simulation studies that the proposed method is effec-
detection approach for discrete-time 2-D uncertain nonlinear tive and advantageous. The contributions of this paper are
systems. summarized as:
For the problem of analysis and synthesis of dynamical sys- 1) Firstly, compared with the results on 2-D linear systems
tems with disturbances/faults, it is worth mentioning that the in [39], [40], a unified framework of finite frequency analysis
frequencies of disturbances and faults are always restricted in and synthesis has been established for 2-D nonlinear systems,
many practical situations [33], [34]. However, the traditional which can be straightforwardly extended to address the finite
fault detection approaches do not fully utilize the useful fre- frequency control, estimation, model reduction problems for
quency spectrum information of faults and disturbances, 2-D nonlinear systems.
which makes these approaches tend to be conservative. The 2) Secondly, we have exploited the available frequency
authors in [35] developed a powerful finite frequency analy- information of disturbances/faults to address finite frequency
sis and synthesis tool, i.e., the generalized Kalman- H∞ and H− performances of 2-D nonlinear systems. The less
Yakubovich-Popov (GKYP) lemma, which reveals the equiv- conservatism of the finite frequency performance over the
alence between a set of linear matrix inequalities (LMIs) and conventional full frequency performance [18]–[20] is ana-
frequency domain inequalities of transfer functions. Thanks to lyzed.
the powerful tool GKYP lemma developed in [35], some fre- 3) Thirdly, for both accessible and inaccessible frequency
quency information based fault detection filtering design information cases, the proposed finite frequency fault detec-
results for 1-D systems have been reported [36]–[38]. For tion filtering design method for 2-D nonlinear systems can be
example, assuming that the faults and the unknown inputs applied, which indicates the wide applicability of the method.
reside in high-/middle-/low-frequency ranges, mixed H− /H∞ Notations: Rn denotes the Euclidean space of n dimensions.
filtering design results were obtained for T-S fuzzy systems in R n×m and Sn refer to the set of n × m matrices and the set of
[36]. In [37], by sum of squares (SOS) techniques, a finite fre- n × n symmetric matrices, respectively. For a matrix G, G∗
quency polynomial fuzzy filtering design method was pro- and GT denote its complex-conjugate transpose and transpose,
posed and applied to discrete-time polynomial fuzzy systems. respectively. diag{· · ·} refers to the block diagonal matrix. In a
The authors in [39] extended the GKYP lemma to Roesser matrix, the induced conjugate symmetric term is denoted as ⋆
type 2-D linear systems and provided a new tool for the finite
for simplicity. Sym{G} denotes G + GT . tr(G) denotes the
frequency performance analysis and synthesis of such sys-
trace of the matrix G. j is the imaginary unit. l2 {[0, ∞), [0,
tems. In [40], based on the 2-D version of the GKYP lemma,
∞)} represents the space of square-summable infinite vector
the problem of finite frequency H∞ /H− observer design and
sequences over {[0, ∞), [0, ∞)} .
its application to fault detection for linear 2-D Roesser sys-
tems was studied. II. Problem Formulation
Compared with 2-D linear systems, the finite frequency The following IF-THEN rules are utilized to describe a class
analysis and design of 2-D nonlinear systems are much more of 2-D Roesser nonlinear systems:
complex and challenging. The main difficulty comes from the
fact that the 2-D version of GKYP lemma proposed in [39] Plant Rule Rl : IF ψ1 (i, j) is Ψl1 and ψ2 (i, j) is Ψl2 and
can not be applied to 2-D nonlinear systems, since their trans-
· · · and ψφ (i, j) is Ψlφ , THEN
fer functions do not exist. In contract, for the linear case, the
 +
frequency domain inequalities of transfer functions can be 
 x (i, j) = (Al + ∆Al )x(i, j) + Bl d(i, j) + El f (i, j)

used to describe the finite frequency performance, and then, 

y(i, j) = (Cl + ∆Cl )x(i, j) + Dl d(i, j) + Fl f (i, j)
(1)
based on the 2-D GKYP lemma, these frequency domain
inequalities can be converted to linear matrix inequality based where
WANG et al.: FUZZY-MODEL-BASED FINITE FREQUENCY FAULT DETECTION FILTERING DESIGN FOR 2-D NONLINEAR SYSTEMS 2101

 h   h   +
 x (i + 1, j)   x (i, j)  
 x (i, j) = (A(µ) + ∆A(µ))x(i, j)
x (i, j) =  v
+
 , x(i, j) =  v  (2) 



x (i, j + 1) x (i, j) 

 + B(µ)d(i, j) + E(µ) f (i, j)

h 


(8)
with xh (i, j) ∈ Rnx denotes the horizontal state; xv (i, j) ∈ Rnx
v


y(i, j) = (C(µ) + ∆C(µ))x(i, j)



denotes the vertical state; x(i, j) ∈ Rnx denotes the system 
+ D(µ)d(i, j) + F(µ) f (i, j)
state; y(i, j) ∈ Rny , d(i, j) ∈ Rnd , and f (i, j) ∈ Rn f denote the
measured output, the disturbance, and the fault, respectively; where
Rl is the l-th fuzzy rule with l ∈ L = {1, 2, . . . , r} ; r is the num-  ∑ ∑




 A(µ) = µl (i, j)Al , ∆A(µ) = µl (i, j)∆Al
ber of the fuzzy rules; ψ(i, j) = [ψ1 (i, j), ψ2 (i, j), . . . , ψφ (i, j)] 



 l∈L l∈L
are premise variables; Ψlϕ are fuzzy sets; Al ∈ Rnx ×nx , Bl ∈ 



 ∑ ∑
Rnx ×nd , El ∈ Rnx ×n f , Cl ∈ Rny ×nx , Dl ∈ Rny ×nd , and Fl ∈ Rny ×n f 

 = µ , = µl (i, j)El



B(µ) l (i, j)Bl E(µ)
are the l-th local system matrices; ∆Al and ∆Cl denote the 

 l∈L l∈L
modeling uncertainties 

 ∑ ∑ (9)



    
C(µ) = µl (i, j)Cl , ∆C(µ) = µl (i, j)∆Cl
 ∆Al   Ul1  




  =   ∆l Wl , l ∈ L
l∈L l∈L
(3) 


∆Cl Ul2 

 ∑ ∑


 D(µ) = µl (i, j)Dl , F(µ) = µl (i, j)Fl .
with Ul1 , Ul2 , and Wl denote known real matrices; ∆l satisfies 

l∈L l∈L
∆Tl ∆l ≤ I p2 , l ∈ L. (4) For the Roesser type 2-D T-S fuzzy system (1) or (8), we
propose a fuzzy filter
Note that the disturbance d(i, j) ∈ l2 {[0, ∞), [0, ∞)} is energy
bounded and its frequency satisfies Filter Rule F l : IF ψ1 (i, j) is Ψl1 and ψ2 (i, j) is Ψl2 and
{
Θd = (ϑhd , ϑvd )|ϑhd ∈ [θhd , θ̄dh ], ϑvd ∈ [θvd , θ̄dv ], · · · and ψφ (i, j) is Ψlφ , THEN
}  +

 x̂ (i, j) = Gl x̂(i, j) + Hl y(i, j)

0 ≤ θ̄dh − θhd ≤ 2π, 0 ≤ θ̄dv − θvd ≤ 2π (5)


r(i, j) = Ml x̂(i, j) + Nl y(i, j), l ∈ L
(10)
where θhd , θ̄dh , θv , and θ̄v are known scalars. The fault f (i, j)
d d where
resides in a low finite frequency set
 h   h 
{ }  x̂ (i + 1, j)   x̂ (i, j) 
Θ f = (ϑhf , ϑvf )||ϑhf | ≤ θ̄hf , |ϑvf | ≤ θ̄vf , 0 ≤ θ̄hf , θ̄vf ≤ π (6) x̂ (i, j) =  v
+  , x̂(i, j) =  v  (11)
x̂ (i, j + 1) x̂ (i, j)
θ̄hf and θ̄vf are known scalars. Suppose that the system state h
with x̂h (i, j) ∈ R x̂ ( 1 ≤ nhx̂ ≤ nhx ) denotes the horizontal filter
n
∑
satisfies the boundary condition ∞ k=0 (|x (0, k)| + |x (k, 0)| ) <
h 2 v 2 v
state; x̂ (i, j) ∈ Rnx̂ ( 1 ≤ nvx̂ ≤ nvx ) denotes the vertical filter
v
∞. state; x̂(i, j) ∈ Rnx̂ ( 1 < n x̂ ≤ n x ) is the filter state; r(i, j) ∈ Rnr
Remark 1: It follows from the definitions of frequency sets
denotes the residual signal. The matrices Gl ∈ Rnx̂ ×nx̂ , Hl ∈
Θd and Θ f in (5) and (6) that by choosing θhd = θvd = −π , Rnx̂ ×ny , Ml ∈ Rnr ×nx̂ , and Nl ∈ Rnr ×ny , l ∈ L are filter gains.
θ̄dh = θ̄dv = π , and θ̄hf = θ̄vf = π , the restrictions of the frequen- The boundary of the filter satisfies x̂h (0, k) = 0, x̂v (k, 0) =
cies of disturbances and faults are removed, which indicates 0, k = 0, 1, . . . , ∞ . Similar to (8), it can be established that
that the disturbances and faults can reside in any frequency  +

 x̂ (i, j) = G(µ) x̂(i, j) + H(µ)y(i, j)

ranges. Therefore, the considered frequency information based


r(i, j) = M(µ) x̂(i, j) + N(µ)y(i, j), l ∈ L
(12)
disturbances and faults are very general and include their cor-
responding ones in the whole frequency domain as their spe-
where
cial cases.  ∑ ∑
∏ [ ] 
Define Ψl := φϕ=1 Ψlϕ . Let µl ψ(i, j) be its normalized 

G(µ) = µ l (i, j)G l , H(µ) = µl (i, j)Hl



membership function  l∈L l∈L


 ∑ ∑ (13)
∏φ [ ] 

 M(µ) = µ (i, j)M , N(µ) = µl (i, j)Nl .
µlϕ ψϕ (i, j) 
 l l
ϕ=1 l∈L l∈L
µl [ψ(i, j)] = [ ] ≥0
∑
r ∏φ Then, define two new vectors
ϕ=1 µϱϕ ψϕ (i, j)  h   h 
ϱ=1  x̄ (i + 1, j)   x̄ (i, j) 
 , x̄(i, j) =  v
+
x̄ (i, j) =  v  (14)
∑
r x̄ (i, j + 1) x̄ (i, j)
µl [ψ(i, j)] = 1 (7)
l=1
where
 h   v 
where µlϕ [ψ(i, j)] refers to the grade of membership of  x (i, j)
x̄h (i, j) =  h
 v  x (i, j)
 , x̄ (i, j) =  v

 . (15)
ψϕ (i, j) in Ψlϕ . x̂ (i, j) x̂ (i, j)
It follows from the IF-THEN rules in (1) that: Combining (8) and (12) yields
2102 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 11, NO. 10, OCTOBER 2024

 +

 x̄ (i, j) = Ā x̄(i, j) + B̄d(i, j) + Ē f (i, j)
 and fault sensitivity levels can be interpreted as the H∞ and


r(i, j) = C̄ x̄(i, j) + D̄d(i, j) + F¯ f (i, j)
(16) H− performance indices γd and γ f , respectively. It follows
from the definitions in (19) and (20) that an admissible fault
where Ā = ΠT ÃΠ , B̄ = ΠT B̃ , Ē = ΠT Ẽ , C̄ = C̃Π , D̄ = D̃ , detection filter is the one with small γd and large γ f . How-
F¯ = F˜ with ever, in practical applications, there is always a tradeoff
  between these two kind of performances to be considered.
 A(µ) + ∆A(µ) 0 
à =   
H(µ)(C(µ) + ∆C(µ)) G(µ) III. Main Results
    Firstly, the analysis results will be presented in Theorem 1.
 B(µ)   E(µ) 
B̃ =   , Ẽ =   Then, a filtering design method will be obtained and summa-
H(µ)D(µ) H(µ)F(µ) rized in Theorem 2. Finally, a novel fault detection algorithm
[ ] will be presented in Algorithm 1.
C̃ = N(µ)(C(µ) + ∆C(µ)) M(µ)
Theorem 1: The asymptotic stability with finite frequency
D̃ = N(µ)D(µ), F˜ = N(µ)F(µ) (17) H∞ performance γd > 0 and finite frequency H− perfor-
and mance γ f > 0 is guaranteed for the fault detection system
(nh +nh )
  (16), if there exist matrices 0 < Phl ∈ S x x̂ , 0 < Pvl ∈
 Inh 0 0 0  (nh +nh ) (nv +nv ) (nh +nh ) (nv +nv )
  x  S(nx +nx̂ ) , Rhl ∈ S x x̂ , Rvl ∈ S x x̂ , T lh ∈ S x x̂ , T lv ∈ S x x̂ ,
v v

 1   0
Π 0 Invx 0 
Π =   . (n +n )
h h (n +n ) (n +n )
v v h h
 =  (18) l ∈ L , 0 < Qhd ∈ S x x̂ , 0 < Qvd ∈ S x x̂ , 0 < Qhf ∈ S x x̂ ,
Π2  0 Inh 0 0 
0 < Qvf ∈ S(nx +nx̂ ) , satisfying
v v
 x̂ 
0 0 0 Invx̂
T
It should be noted that an admissible fault detection filter is A P(µ+ )A − P(µ) < 0 (23)
one that can generate a residual signal, based on which the
 T  
 Ā B̄   Ā B̄ 
 Φd 
fault is immediately detected once it appears. To this end, two
 
definitions will be introduced. I 0 I 0
Definition 1: Under zero boundary conditions, the fault  T  
detection system (16) satisfies a finite frequency H∞ perfor-  C̄ D̄   C̄ D̄ 
+   Ψd   < 0 (24)
mance γd > 0 , if 0 I 0 I
∑
∞ ∑
∞ ∑
∞ ∑
∞
 T  
r∗ (i, j)r(i, j) ≤ γd2 d∗ (i, j)d(i, j)  Ā Ē   Ā Ē 
 Φ f 
(19)
i=0 j=0 i=0 j=0
 
I 0 I 0
where the disturbance d(i, j) ∈ l2 {[0, ∞), [0, ∞)} and its fre-   T  
quency satisfying (ϑhd , ϑvd ) ∈ Θd as in (5).  C̄ F¯   C̄ F¯ 
+   Ψ f   < 0 (25)
Definition 2: Under zero boundary conditions, the fault 0 I 0 I
detection system (16) satisfies a finite frequency H− perfor- where
mance γ f > 0 , if  
 −R(µ+ ) Q(1) 
∑
∞ ∑
∞ ∑
∞ ∑
∞
Φd =   
(2) 
d
r∗ (i, j)r(i, j) ≥ γ2f f ∗ (i, j) f (i, j) (20) ⋆ R(µ) + Qd
i=0 j=0 i=0 j=0 { }
Ψd = diag I, −γd2 I
where the frequency of the fault satisfying (ϑhf , ϑvf ) ∈ Θ f as in
 
(6).  −T (µ+ ) Q(1) 
Given the finite frequency sets Θd and Θ f in (5) and (6), Φ f =   f 
⋆ T (µ) + Q(2) 
respectively, the aim is to design a filter (12) and generate a f
{ }
residual signal r(i, j) . With the residual signal r(i, j) gener- Ψ f = diag −I, γ2f I (26)
ated by the proposed filter (8), the evaluation function can be
designed as with
v ∑
t∑ ∑j R(µ+ ) = µ+l (i, j)diag{Rhl , Rvl }
i T
i−N1 +1 j−N2 +1 r (m, n)r(m, n)
Jr (i, j) = (21) l∈L
N1 N2 ∑
R(µ) = µl (i, j)diag{Rhl , Rvl }
where N1 and N2 are two positive integers of the evaluation
l∈L
window. ∑
Then, choose the following threshold for fault detection: T (µ+ ) = µ+l (i, j)diag{T lh , T lv }
l∈L
Jth = sup Jr (i, j). (22) ∑
0,d∈l2 , f =0 T (µ) = µl (i, j)diag{T lh , T lv }
Remark 2: Note that the worst case disturbance attenuation l∈L
WANG et al.: FUZZY-MODEL-BASED FINITE FREQUENCY FAULT DETECTION FILTERING DESIGN FOR 2-D NONLINEAR SYSTEMS 2103

h + v +
Q(1)
d
= diag{ej(θd ) Qhd , ej(θd ) Qvd } where
{ } h +
Q(2) = −diag 2 cos((θ h −
d ) )Q h
d , 2 cos((θ v −
d ) )Q v
d
Υhd (i, j) = ej(θd ) xh (i, j)(xh (i + 1, j))∗
d
h +
Q(1) = diag{Qhf , Qvf } + e−j(θd ) xh (i + 1, j)(xh (i, j))∗
f
{ }
Q(2) = −diag 2 cos( θ̄ h
)Q h
, 2 cos( θ̄ v
)Q v − 2 cos((θdh )− )xh (i, j)(xh (i, j))∗
f f f f f
v +
θ̄dh + θhd θ̄dh − θhd Υvd (i, j) = ej(θd ) xv (i, j)(xv (i, j + 1))∗
(θdh )+ = , (θdh )− = v +
2 2 + e−j(θd ) xv (i, j + 1)(xv (i, j))∗
θ̄dv + θvd θ̄dv − θvd − 2 cos((θdv )− )xv (i, j)(xv (i, j))∗ .
(θdv )+ = , (θdv )− = . (27) (35)
2 2
Note that the following equation holds if the system (16) is
Proof: The system (16) with no fault and external distur-
asymptotically stable:
bance reduces to
∑
∞ ∑
∞
+
x̄ (i, j) = Ā x̄(i, j). (28) −( x̄+ (i, j))∗ R(µ+ )( x̄+ (i, j))
i=0 j=0
Construct a 2-D fuzzy Lyapunov function candidate
∑
∞ ∑
∞
V(x(i, j)) = x̄T (i, j)P(µ) x̄(i, j) (29) + ( x̄(i, j))∗ R(µ) x̄(i, j) = 0. (36)
∑ i=0 j=0
where P(µ) = l∈L µ(i, j)diag{Pl , Pl } with Pl > 0 , Pvl > 0 .
h v h
Also note that in (34), the matrices Qhd and Qvd satisfy
Define ∑ ∑∞ h
Qd > 0 and Qv > 0 , respectively. If ∞
h
i=0 j=0 Υd (i, j) ≥ 0 and
∆V(x(i, j)) = V(x+ (i, j)) − V(x(i, j)). (30) ∑∞ ∑∞ d
i=0 j=0 Υd (i, j) ≥ 0 can be shown. From (34) we have
v

Note that the inequality

∑
∞ ∑
∞ ∑
∞ ∑
∞
+ r∗ (i, j)r(i, j) ≤ γd2 d∗ (i, j)d(i, j) (37)
Ā P(µ )Ā − P(µ) < 0
T
(31)
i=0 j=0 i=0 j=0
ensures ∆V(x(i, j)) < 0 .
which ensures the finite freuqncy H∞ performance. Let
For the case with no fault, the fault detection system (16)
x̄h (ϑhd , ϑvd ) denote the 2-D Fourier transform of the horizontal
reduces to ∑∞ ∑∞
 + state x̄h (i, j) . The time-domain term i=0 j=0 Υhd (i, j) is trans-

 x̄ (i, j) = Ā x̄(i, j) + B̄d(i, j)

 (32) formed into the following frequency domain term:

r(i, j) = C̄ x̄(i, j) + D̄d(i, j).
∑
∞ ∑
∞ ∑ ∞ {
∞ ∑
h +
Left- and [right-multiplying (24) with the vector [ x̄∗ (i, j) Υhd (i, j) = e j(θd ) x̄h (i, j)( x̄h (i + 1, j))∗
]
x̄(i, j) i=0 j=0 i=0 j=0
d∗ (i, j)] and d(i, j) yields h +
+ e− j(θd ) x̄h (i + 1, j)( x̄h (i, j))∗
 + ∗   + 
 x̄ (i, j)   −R(µ+ ) Q(1)   x̄ (i, j)  }
   d    −2 cos((θdh )− ) x̄h (i, j)( x̄h (i, j))∗
x̄(i, j) ⋆ R(µ) + Q(2) x̄(i, j)
1 w π w π ( j((θh )+ −ϑh ) )
d
j(ϑhd −(θdh )+ )
+ r∗ (i, j)r(i, j) − γd2 d∗ (i, j)d(i, j) < 0. (33) = 2 e d d +e − 2 cos((θdh )− )
4π −π −π
Sum up (33) from (0, 0) to (∞, ∞) × x̄h (ϑhd , ϑvd )( x̄h (ϑhd , ϑvd ))∗ dϑhd dϑvd
∑
∞ ∑
∞
− ( x̄+ (i, j))∗ R(µ+ )( x̄+ (i, j)) 1 wπ wπ ( )
= 2 cos((θdh )+ − ϑhd ) − cos((θdh )− )
i=0 j=0 2π −π −π
∑
∞ ∑
∞ × x̄h (ϑhd , ϑvd )( x̄h (ϑhd , ϑvd ))∗ dϑhd dϑvd .
+ ( x̄(i, j))∗ R(µ) x̄(i, j)
It follows 0 ≤ |(θdh )+ − ϑdh | ≤ (θdh )− ≤ π that cos((θdh )+ − ϑhd )−
i=0 j=0 ∑∞ ∑∞ h
    cos((θdh )− ) ≥ 0 , and thus i=0 j=0 Υd (i, j) ≥ 0 . Similarly,
 ∑∞ ∑
∞   ∑∞ ∑
∞  ∑∞ ∑∞ v
+ tr Qhd Υhd (i, j) + tr Qvd Υvd (i, j) i=0 j=0 Υd (i, j) ≥ 0 . Hence, it follows from the inequality
i=0 j=0 i=0 j=0 (24) that the finite frequency H∞ performance can be guaran-
∑
∞ ∑
∞ teed.
+ r∗ (i, j)r(i, j) In the case of no disturbances, the system (16) reduces to
 +
i=0 j=0

 x̄ (i, j) = Ā x̄(i, j) + Ē f (i, j)

∑
∞ ∑
∞ 

r(i, j) = C̄ x̄(i, j) + F¯ f (i, j).
(38)
− γd2 d∗ (i, j)d(i, j) < 0 (34)
[
i=0 j=0 Left- and right-multiply (38) with the vector x̄∗ (i, j) f ∗ (i, j)]
2104 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 11, NO. 10, OCTOBER 2024

[ ]
x̄(i, j) form technique is used to convert time domain inequalities
and f (i, j) , respectively. into their frequency domain counterparts, and then utilized in
the performance analysis results in Theorem 1.
∑
∞ ∑
∞
Remark 5: In practice, faults usually belong to the low fre-
− ( x̄+ (i, j))∗ T (µ+ )( x̄+ (i, j))
quency domain, for example, an incipient signal, the actuator
i=0 j=0
stuck failures, etc. [40], which motivates the study of the low
∑
∞ ∑
∞
frequency fault detection problem. Note that our proposed
+ ( x̄(i, j))∗ T (µ) x̄(i, j) method can also be extended to handle more general faults.
i=0 j=0
    Specifically, redefine the frequency set of the fault as
 ∑ ∞ ∑
∞   ∑ ∞ ∑
∞  {

+ tr Q fh  
Υ f (i, j) + tr Q f
h v
Υ f (i, j)
v Θ f = (ϑhf , ϑvf )|ϑhf ∈ [θhf , θ̄hf ], ϑvf ∈ [θvf , θ̄vf ],
i=0 j=0 i=0 j=0 }
0 ≤ θ̄hf − θhf ≤ 2π, 0 ≤ θ̄vf − θvf ≤ 2π
∑
∞ ∑
∞ ∑
∞ ∑
∞
− r∗ (i, j)r(i, j) + γ2f f ∗ (i, j) f (i, j) < 0 (39) which can represent the low, middle, and high frequency sets
i=0 j=0 i=0 j=0 by choosing the parameters θhf , θ̄hf , θvf , and θ̄vf . Then, the
(1) (2)
where matrices Q f , Q f in (27) in Theorem 1 can be rewritten as
Υhf (i, j) = xh (i, j)(xh (i + 1, j))∗ + xh (i + 1, j)(xh (i, j))∗ Q(1)
j(θh )+
= diag{e f Qhf , e f Qvf }
j(θv )+
f
h { }
− 2 cos(θ f )xh (i, j)(xh (i, j))∗ Q(2) = −diag 2 cos((θhf )− )Qhf , 2 cos((θvf )− )Qvf
f
Υvf (i, j) = xv (i, j)(xv (i, j + 1))∗ + xv (i, j + 1)(xv (i, j))∗
θ̄hf + θhf θ̄hf − θhf
v (θhf )+ = , (θhf )− =
− 2 cos(θ f )xv (i, j)(xv (i, j))∗ . (40) 2 2
Similarly, the following inequality can be shown: θ̄vf + θvf θ̄vf − θvf
(θvf )+ = , = (θvf )− .
∑
∞ ∑
∞ 2 2
− ( x̄+ (i, j))∗ T (µ+ )( x̄+ (i, j)) Then, Theorem 1 can be used to handle the faults with more
i=0 j=0 general frequency characteristics including low, middle, or
∑
∞ ∑
∞ high frequency domain. The proof procedure is similar to that
+ ( x̄(i, j))∗ T (µ) x̄(i, j) of the low frequency domain and thus omitted.
i=0 j=0 Theorem 2: Given γd > 0 and γ f > 0 , the asymptotic stabil-
  ity, the finite frequency H∞ performance γd , and the finite
 ∑ ∞ ∑
∞ 

+ tr Q fh
Υ f (i, j)
h frequency H− performance γ f is ensured for the fault detec-
(nh +nh )
i=0 j=0 tion system (16), if there exist matrices 0 < Phl ∈ S x x̂ , 0 <
  Pvl ∈ S(nx +nx̂ ) , Rhl ∈ S(nx +nx̂ ) , Rvl ∈ S(nx +nx̂ ) , T lh ∈ S(nx +nx̂ ) , T v ∈
v v h h v v h h
 ∑ ∞ ∑
∞ 

+ tr Q fv
Υ f (i, j) ≥ 0.
v (n +n ) h (n +n )
h v v
l
S(nx +nx̂ ) , l ∈ L , 0 < Qhd ∈ S x x̂ , 0 < Qvd ∈ S x x̂ , 0 < Qhf ∈
v v
(41)
i=0 j=0 (n +n )
v v
S(nx +nx̂ ) , 0 < Qvf ∈ S x x̂ , and matrices Ḡl ∈ Rnx̂ ×nx̂ , H̄l ∈
h h

Thus, the finite frequency H− can be ensured

Rnx̂ ×ny , M̄l ∈ Rnr ×nx̂ , N̄l ∈ Rnr ×ny , Yl1 ∈ R2(nx +nx̂ )×nx , Yl2 ∈
∑
∞ ∑
∞ ∑
∞ ∑
∞
R(2nx +2nx̂ +nd )×nx , Yl3 ∈ R(2nx +2nx̂ +n f )×nx , S l1 ∈ Rnx̂ ×nx̂ , and
r∗ (i, j)r(i, j) ≥ γ2f f ∗ (i, j) f (i, j). (42) (1) (2) (3)
S l2 ∈ Rnr ×nr , l ∈ L , and ϵls > 0 , ϵls > 0 , and ϵls > 0 , l, s ∈ L ,
i=0 j=0 i=0 j=0
such that
■  (1) 
Remark 3: Compared with the filtering design methods for  Σmls + Σ(1) Uls(1) U sl(1) 
 msl 
2-D systems in [31], [41], one advantage of the results in The-  ⋆ −ϵls(1) I p1 0 p1 ×p1  < 0 (43)
 
orem 1 is that the useful and available frequency spectrum  (1)
information of external disturbances and faults is exploited. ⋆ ⋆ −ϵ sl I p1
Note that by choosing Qhd = 0 , Qvd = 0 , Qhf = 0 , Qvf = 0 in  (2) 
 Σmls + Σ(2) Uls(2) U sl(2) 
Theorem 1, and observing (23)–(25) and matrix definitions in  msl 
 ⋆ −ϵls(2) I p1 0 p1 ×p1  < 0 (44)
(26), (27), it can been seen that no frequency spectrum of dis-  

turbances/faults is used in the performance analysis results. ⋆ ⋆ −ϵ sl(2) I p1
For this case, our results reduce to the non-frequency perfor-  (3) 
mance analysis ones.  Σmls + Σ(3) Uls(3) U sl(3) 
 msl 
Remark 4: Due to the 2-D structure and nonlinearity of the  ⋆ −ϵls(3) I p1 0 p1 ×p1  < 0 (45)
 
2-D nonlinear system (16), the commonly used GKYP lemma 
[35] in relevant finite frequency filtering design works [36], ⋆ ⋆ −ϵ sl(3) I p1
[37] can not be adopted in this case. The 2-D Fourier trans- where l, s, m ∈ L and
WANG et al.: FUZZY-MODEL-BASED FINITE FREQUENCY FAULT DETECTION FILTERING DESIGN FOR 2-D NONLINEAR SYSTEMS 2105

{ } −1 −1
Σ(1) = diag ΠP m Π T
, −P l + ϵ (1) T T
Π 1 W s W s Π 1 Gl = S l1 Gl , Hl = S l1 Hl
mls ls
{[ −1 −1
+ Sym −Yl1 −e1 S l1 Ml = S l2 M l , Nl = S l2 Nl. (47)
]}
(Yl1 A s + e1 H̄lC s )Π1 + e1Ḡl Π2 Proof: In the first step, it will be shown that the condition
(43) guarantees (23). Substituting Ā = ΠT ÃΠ into (23), one
Uls(1) = Yl1 U s1 + e1 H̄l U s2 has
{ }  T   
Σ(2) = diag Ξd , I, −γd2 I  ÃΠ   ΠP(µ+ )ΠT 0   ÃΠ 
      < 0.
mls
{ } (48)
I 0 −P(µ) I
+ Sym Yl2 Φ(1) d
+ e2 Φ(2)
d
+ e3 Φ(3)
d
Assign
Uls(2) = Y s2 Ul1 + e2 H̄ s Ul2 + e3 N̄ s Ul2
 
   ΠP(µ+ )ΠT 0 
 −ΠRm ΠT ΠQ(1)  Σ →  
Ξd =   d  0 −P(µ)
⋆ Rl + Q(2)d
+ ϵ (2) T T
ls
Π1 W s W s Π1 [ ]
h + v +
Z → −I(nx +nx̂ ) ÃΠ
Q(1)
d
= diag{e j(θd ) Qhd , e j(θd ) Qvd }
{ } X → I. (49)
Q(2)
d
= −diag 2 cos((θdh )− )Qhd , 2 cos((θdv )− )Qvd
[ ] It can be checked that
Φ(1) = −I 0 A s Π 1 0 B s  
d [ ]  ÃΠ 
[ ] −I(nx +nx̂ ) ÃΠ    = 0
Φ(2)
d
= 0 −S l1 H̄lC s Π1 + Ḡl Π2 0 H̄l D s | {z } I
[ ] Z | {z }
Φ(3) = 0 0 N̄lC s Π1 + M̄l Π2 −S l2 N̄l D s NZ
d
{ } where NZ denotes the right null spaces Z and satisfy
Σ(3)
mls
= diag Ξ f , −I, γ2f I ZNZ = 0 . Then, (48) can be rewritten as NZT ΣNZ < 0 . Using
{ }
+ Sym Yl3 Φ(1) + e4 Φ(2) + e5 Φ(3) the Projection Lemma shown in the Appendix, (48) can be
f f f
implied by
Uls(3) = Y s3 Ul1 + e4 H̄ s Ul2 + e5 N̄ s Ul2 { } { }
Σ + Sym XT YZ = diag ΠP(µ+ )ΠT , −P(µ)
 
 −ΠTm ΠT ΠQ(1)  { [ ]}
Ξ f =  f 
 + Sym Y −I(nx +nx̂ ) ÃΠ < 0 (50)
⋆ Tl + Q(2)
f
+ ϵ (3) T T
ls
Π 1 W s W s Π1
where Y is a variable to be assigned. In order to facilitate the
Q(1) = diag{Qhf , Qvf }
f fault detection filtering design, the variable Y is assigned as
{ } [ ]
Q(2) = −diag 2 cos( θ̄ h
)Q h
, 2 cos( θ̄ v
)Q v Y → Yl1 e1 S l1 , l ∈ L (51)
f f f f f
[ ]
(1)
Φ f = −I 0 A s Π1 0 E s where Yl1 ∈ R2(nx +nx̂ )×nx and S l1 ∈ Rnx̂ ×nx̂ , and e1 = [0nx̂ ×nx
[ ] In x̂ 0nx̂ ×(nx −nx̂ ) Inx̂ 0nx̂ ×n x̂ ]T . By assigning the slack matrix Y
Φ(2)
f
= 0 −S l1 H̄lC s Π1 + Ḡl Π2 0 H̄l F s in (51), the matrix S l1 can be absorbed in the filter gains by
[ ] defining
Φ(3)
f
= 0 0 N̄lC s Π1 + M̄l Π2 −S l2 N̄l F s
[ ]T Ḡl = S l1Gl , H̄l = S l1 Hl , l ∈ L. (52)
e1 = 0nx̂ ×nx Inx̂ 0nx̂ ×(nx −nx̂ ) In x̂ 0nx̂ ×nx̂ Substituting à in (17) and Y in (51) into (50), and taking
[ ]T (9), (13), and (52) into consideration, one has
e2 = 0nx̂ ×nx Inx̂ 0nx̂ ×nx Inx̂ 0nx̂ ×(nr +nd )
∑∑ ∑ { { }}
[ µl µ s µ+m Γ(1) + Sym Uls(1) ∆ s W(1)
s
e3 = 0nr ×(nx −nr ) Inr 0nr ×(nx̂ −nr ) Inr 0nr ×nx mls
l∈L s∈L m∈L
]T
0nr ×(nx̂ −nr ) Inr Inr 0nr ×nd ∑
r ∑
r { { }}
= µ2l µ+m Γ(1)
mll
+ Sym Ull(1) ∆l Wl(1)
[ ]T l=1 m=1
e4 = 0nx̂ ×(nx −nx̂ ) 0nx̂ ×(nx̂ +nx ) Inx̂ 0nx̂ ×(nr +n f )
[ ∑
r−1 ∑
r ∑
r {
e5 = 0nr ×(n x −nr ) Inr 0nr ×(nx̂ −nr ) Inr 0nr ×nx + µl µ s µ+m Γ(1)
mls
+ Γ(1)
msl
l=1 s=l+1 m=1
]T { } { (1) }}
0nr ×(nx̂ −nr ) Inr Inr 0nr ×n f . (46) + Sym Uls(1) ∆ s W(1) (1)
s + Sym U sl ∆l Wl < 0 (53)
Moverover, the filter gains can be obtained by where
2106 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 11, NO. 10, OCTOBER 2024

{ } {[  
Γ(1) = diag ΠPm ΠT , −Pl + Sym −Yl1 −e1 S l1  −ΠRm ΠT ΠQ(1) 0 0 
mls  d 
]}  Q(2) ΠT Rl + Q(3) 0 0 
(Yl1 A s + e1 H̄lC s )Π1 + e1Ḡl Π2 Γmls = 
(2)  d d 

 0 0 I 0 

Uls(1) = Yl1 U s1 + e1 H̄l U s2 0 0 0 −γd2 I
[ ] { }
W(1)s = 0 p2 ×n x 0 p2 ×n x̂ W s Π1 . (54) + Sym Yl2 Φ(1) + e2 Φ(2) + e3 Φ(3)
d d d
[ ]
Since µl > 0 , µ s > 0 , and µ+m > 0 , the following inequality Φ(1) = −I 0 A s Π1 0 Bs
d
(55) holds for 1 ≤ l ≤ s ≤ r , 1 ≤ m ≤ r such that (53) can be [ ]
guaranteed: Φ(2)
d
= 0 −S l1 H̄ l C s Π 1 + Ḡ l Π 2 0 H̄ l D s
{ } [ ]
Γ(1)
mls
+ Γ(1)
msl
+ Sym Uls(1) ∆ s W(1)
s Φ(3)
d
= 0 0 N̄lC s Π1 + M̄l Π2 −S l2 N̄l D s
{ }
+ Sym U sl(1) ∆l Wl(1) < 0. (55) Uls(2) = Yl2 U s1 + e2 H̄l U s2 + e3 N̄l U s2
[ ]
Using Lemma 2 in the Appendix, (44) is implied by W(2)s = 0 p2 ×n x 0 p2 ×n x̂ W s Π1 0 p2 ×(nd +nr ) . (62)

Γ(1) + Γ(1) + ϵls(1) (W(1) (1) (1) (s) T

s ) W s + ϵ sl (Wl ) Wl
T (1) Similar to the procedure as (55) and (56) in the first step,
mls msl
one has that the LMIs in (44) imply the H∞ performance
1 1 index γd . Similar to the first and second steps, we have (45).
+ Uls(1) (Uls(1) )T + U sl(1) (U sl(1) )T < 0. (56)
ϵls(1) ϵ sl(1) ■
Remark 6: Note that Theorem 2 covers conventional full
By Schur complement, the LMIs in (43) imply (56). The frequency filtering as a special case. Specifically, with
asymptotic stability has then been shown. Qhf = 0 , Qv = 0 / Qhd = 0 , Qv = 0 , in Theorem 2, the frequency
In the second step, substituting Ā = ΠT ÃΠ , B̄ = ΠT B̃ , f d
spectrum of faults/disturbances are not utilized. Then, Theo-
C̄ = C̃Π , and D̄ = D̃ into (44), it can be established that
rem 2 will be reduced to the conventional full frequency filter-
 T   ing design method, which is conservative.
 ÃΠ B̃   ÃΠ B̃ 
 I 0 
  I 0 
 Remark 7: In practical applications of our method, one
   <0
  Φ̄ , Ψ }
d  should consider the tradeoff between the two performances.
 C̃Π D̃ 
diag{ (57)
 C̃Π D̃ 
d
    Given the weighting parameters κd and κ f , which satisfy
0 I 0 I 0 ≤ κd , κ f ≤ 1 and κd + κ f = 1 , the minimized mixed H∞ /H−
where performance index κd γd2 − κ f γ2f with its corresponding filter
  gains can be obtained by Algorithm 1.
 −ΠR(µ+ )ΠT ΠQ(1) 
Φ̄d =  d  Algorithm 1:
⋆ R(µ) + Q(2)
{ }
d min κd γd2 − κ f γ2f s.t.
Ψd = diag I, −γd2 I . (58)
(43), (44), and (45) (Theorem 2).
Similar to the steps in (48)−(50), by Projection Lemma with With the proposed results, the fault is detected by the fol-
a new variable M , it can be established that lowing steps.
   Step 1: Given the weighting parameters κd and κ f , which


  −I ÃΠ 0 B̃  

diag{Φ̄d , Ψd } + Sym  M   < 0 (59) satisfy 0 ≤ κd , κ f ≤ 1 and κd + κ f = 1 , by Algorithm 1, the

 0 C̃Π −I D̃  
mixed performance index κd γd2 − κ f γ2f with its corresponding
which implies (44). The variable M is assigned as fault detection filter gains can be obtained.
[ ] Step 2: Choose two positive integers N1 and N2 . In the
M → Yl2 e2 S l1 e3 S l2 (60) fault-free √ case, calculate the threshold Jth (i, j) =
∑i ∑j
rT (m,n)r(m,n)
where Yl2 ∈ R(2nx +2nx̂ +nw )×nx S l1 ∈ Rnx̂ ×nx̂ , and S l2 ∈ Rnx̂ ×nr̂ sup
i−N1 +1 j−N2 +1
N1 N2 .
are variables to be designed, and e2 = [0nx̂ ×nx Inx̂ 0nx̂ ×nx 0,d∈l2 , f =0 √∑
i ∑j
In x̂ 0nx̂ ×(nr +nd ) ]T , e3 = [0nr ×(nx −nr ) Inr 0nr ×(nx̂ −nr ) Inr 0nr ×nx
T
i−N1 +1 j−N2 +1 r (m,n)r(m,n)
Step 3: Calculate Jr = N1 N2 . If Jr (i,
0nr ×(nx̂ −nr ) Inr Inr 0nr ×nd ]T .
j) > Jth , then the fault occurs.
Substituting à , B̃ , C̃ , D̃ in (17) and M in (60) into (59)
and considering the definitions Ḡl = S l1Gl , H̄l = S l1 Hl , IV. Simulation Studies
M̄l = S l2 Ml , and N̄l = S l2 Nl , one has In this section, two industrial 2-D processes, i.e., a partial
∑∑ ∑ { { (2) }} differential equation used to describe a thermal process, and a
µl µ s µ+m Γ(2)
mls
+ Sym Uls
∆ s W (2)
s <0 (61) Darboux equation process, are presented.
l∈L s∈L m∈L Example 1: Consider the following partial differential equa-
where tion [31]:
WANG et al.: FUZZY-MODEL-BASED FINITE FREQUENCY FAULT DETECTION FILTERING DESIGN FOR 2-D NONLINEAR SYSTEMS 2107

∂Λ(s, t) ∂Λ(s, t) where ∆l = diag{ν1 , ν2 } and

=− − (1 + 0.1ν1 (t))h1 (s, t)Λ(s, t)
∂s ∂t   

  0 0  [ ]
+ h2 (s, t)d(s, t) + h3 (s, t) f (s, t) 

U = 
  , Ul2 = 0 0.05



l1
0.1 0

y(s, t) = (1 + 0.05ν2 (t))Λ(s, t) + 0.1d(s, t) 

    

  0 −∆tψmin   0 −∆tψmax 
Λ(0, t) = g1 (t), Λ(s, t) = g2 (s)


W1 =  0
  , W2 =   .
1 0 1
s ∈ [0, s f ], t ∈ [0, ∞] (63) By Algorithm 1 with κd = 1 , κ f = 0 , and θhd = −θ̄dh = −0.2π ,
where |ν1 | ≤ 1 , |ν2 | ≤ 1 , and Λ(s, t) refers to the temperature in θvd = −θ̄dv = −0.2π , θ̄hf = 0.1π , and θ̄vf = 0.1π , one has the opti-
space s at time t. mized κd γd2 − κ f γ2f = 0.0428 , γd = 0.2069 , and γ f = 0.0089
Defining Λ(i, j) = Λ(i∆s, j∆t) , d(i, j) = d(i∆s, j∆t) , and with filter gains
f (i, j) = f (i∆s, j∆t) yields  
   −0.6221 0.4815 −0.6117 
∂Λ(i, j) Λ(i, j + 1) − Λ(i, j)  G1 H1   
≈   =  −0.5615 −0.3261 −0.1365 
∂t ∆t M1 N1  
1.0875 −0.0928 −0.3194
∂Λ(i, j) Λ(i, j) − Λ(i − 1, j)
≈ .  
∂s ∆s    −0.2459 0.4232 −0.6656 
 G2 H2   
Choose   =  −0.3924 −0.3522 0.0388  .
 h    M2 N2  
 x (i, j)   Λ(i − 1, j)  0.9623 1.3468 −0.3821
x(i, j) =  v  =   .
x (i, j) Λ(i, j) As mentioned in Remark 3, choosing Qhd = 0 , Qvd = 0 ,
Then, one has Qhf = 0 , Qv = 0 , Theorem 2 becomes a conventional filtering
f
 
 0 1  design method, which can deal with the case of unrestricted

x (i, j) =  ∆t
+  x(i, j)
∆t  frequency faults and disturbances. Now, we will make a com-
1− − (1 + 0.1ν1 (t))h1 (i, j)∆t  parison between the finite frequency filtering design method
∆s ∆s
    over the conventional full frequency one. Applying Algo-
 0   0 
+   d(i, j) +   f (i, j). rithm 1 (Theorem 2) with Qhd = 0 , Qvd = 0 , Qhf = 0 , Qvf = 0 ,
h2 (i, j)∆t h3 (i, j)∆t κd = 1 , and κ f = 0 , yields no feasible solutions. As shown
Let h1 (i, j) = 7 + sin(xv (i, j)/10) , h2 (i, j) = 1 , h3 (i, j) = 1 , above, the finite frequency filtering method yields feasible
∆s = 0.4 , ∆t = 0.1 . Define solutions γd = 0.2069 and γ f = 0.0089 . It is thus validated
that the proposed fault detection filtering approach is less
ψ(i, j) = 7 + sin(xv (i, j)/10). conservative over the conventional non-frequency informa-
Then, one has ψ(i, j) ∈ [ψmin , ψmax ] , where ψmin = 6 and tion based one [18]–[20], [31].
ψmax = 8 . In addition, the following equations hold: As shown in Fig. 1, choose

µ1 (xv ) × ψmin + µ2 (xv ) × ψmax = ψ(i, j) d(i, j) = exp(−0.02(i + j)) sin(0.02i + 0.01 j).
µ1 (xv ) + µ2 (xv ) = 1
where membership functions are defined as
ψ − ψ(i, j) 0.8
µ1 = max 0.6
ψmax − ψmin
Disturbance d(i, j)

0.4
ψ(i, j) − ψmin
µ2 = . 0.2
ψmax − ψmin
0
The system in (63) is described by (1) with −0.2
     −0.4


  0 1   0 
   200

 =   , =  



A1  ∆t ∆t  Bl
∆t 150 200

 1− − ∆tψmin 150


 ∆s ∆s 100
100


   50 50
  0 1    i 0 0 j


    0 


 A2 =  ∆t ∆t  , El =  


 1− − ∆tψmax ∆t Fig. 1. Disturbance d(i, j) .


 ∆s ∆s



C = [ 0 1 ], D = 0.1, F = 0 Then, as shown in Fig. 2, a combined fault is considered as
l l l

and the parameter uncertainties satisfy 

 1 30 ≤ i ≤ 60, 40 ≤ j ≤ 80



    
 ∆Al   Ul1  f (i, j) = 
sin(0.1i + 0.2 j) 30 ≤ i ≤ 60, 100 ≤ j ≤ 160
  =   ∆l Wl , l ∈ {1, 2} 



0
∆Cl Ul2 otherwise.
2108 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 11, NO. 10, OCTOBER 2024

0.30
Jr(10, j) Jr(120, j)

Evaluation function Jr(*, j) and Threshold Jth

Jr(40, j) Jr(150, j)
1.0 0.25 Jr(70, j) Jth
Jr(100, j)
0.5
0.20
Fault f(i, j)

0
0.15
−0.5
0.10
−1.0
200
150 200 0.05
100 150
50 100
i 50 0
0 0 j 0 50 100 150 200
j
Fig. 2. Fault f (i, j) .
Fig. 4. Residual evaluation function Jr (∗, j) andthreshold Jth .
Choose the parameters of the evaluation window in (21) as
N1 = 10 and N2 = 10 . Under initial condition x(i, j) = [0.2 t ∈ [0, ∞) , a1 , a2 , b1 , and b2 denote real constant coefficients.
∂g(s,t)
0.1] and x̂(i, j) = [0 0] , simulation results can be obtained. Choose h(s, t) = ∂t − a2 g(s, t) , xh (i, j) = h(s, t) = h(i∆s, j∆t) ,
Figs. 3 and 4 show the residual evaluation fuction Jr (i, j) and xv (i, j) = g(s, t) = g(i∆s, j∆t) , y(i, j) = y(i∆s, j∆t) , d(i, j) =
the threshold Jth = 0.1097 . It can be observed from the fault in d(i∆s, j∆t) , and f (i, j) = f (i∆s, j∆t) .
Fig. 2, the residual evaluation fuction and threshold in Fig. 3, Choose the parameters as a1 = −9 , a2 = −2 , b1 = 20 ,
that the fault can be immediately detected once it occurs. Note b2 = 10 , ∆t = 0.1 , and ∆s = 0.1 . Then,
that Figs. 1−3 are all 3-D graphs. We also show a 2-D version [ ]
∆sυ + 1 0
of the simulation results in Fig. 4, where the residual evalua- x+ (i, j) = x(i, j)
( ∆t ∆ta2 + 1
tion fuction and threshold Jr (i, j) , where i = 10, 40, 70, [ ] [ ]
)
100, 120, 150 ) are depicted. From Fig. 4, we can also ∆sb1 ∆sb2
+ d(i, j) + f (i, j)
observe that the fault can be immediately detected once it 0 0
occurs. The effectiveness of the proposed fault detection fil- [ ]
tering method can be verified from the simulation results in y(i, j) = 0 1 x(i, j) + d(i, j) + 0.1 f (i, j)
Figs. 3 and 4. where υ = a1 + arctan(xv (s, t)) . Define
ψ(i, j) = arctan(xv (s, t))
where ψ(i, j) ∈ [ψmin , ψmax ] , with ψmin = − π2 and ψmax = π2 .
Evaluation function Jr(i, j)
Evaluation function Jr(i, j)

Note that the following equations hold:

0.25
Threshold Jth
and threshold Jth

0.20 µ1 (xv ) × ψmin + µ2 (xv ) × ψmax = ψ(i, j)

0.15 µ1 (xv ) + µ2 (xv ) = 1
0.10
where membership functions are chosen as
0.05
ψ − ψ(i, j)
0 µ1 = max
ψmax − ψmin
150 200
100 150 ψ(i, j) − ψmin
50 100 µ2 = .
i 0 50 j
ψmax − ψmin
The system (1) with the following matrices can be used to
Fig. 3. Residual evaluation function Jr (i, j) and threshold Jth . describe (64):
 [ ]
Example 2: This example is used to show the less conser- 

 ∆s(a1 + ψmin ) + 1 0
vatism of our proposed fuzzy Lyapunov function based 

 A =



1
∆t ∆ta2 + 1
method than the common Lyapunov function based one. Con- 



 [ ]
sider a nonlinear Darboux equation [41] 

 ∆s(a1 + ψmax ) + 1 0


 A =



2
∆t ∆ta2 + 1
∂2 g(s, t) ∂g(s, t) ∂g(s, t) 

= (a1 + arctan(g(s, t))) + a2 
 [ ] [ ]
∂s∂t ∂t ∂s 
 ∆sb1 ∆sb2


 B = , E =
+ (−a1 a2 − a2 arctan(g(s, t))g(s, t) 


l
0
l
0




 [ ]
+ b1 d(s, t) + b2 f1 (s, t) 

Cl = 0 1 , Dl = 1, Fl = 0.1




 [ ] [ ]
y(s, t) = g(s, t) + d(s, t) + 0.1 f (s, t) (64) 

 0.2

Ul1 =
 , Ul2 = 0.5, Wl = 0.1 0.1 .
where g(s, t) is a function at space s ∈ [0, sg ] and time interval 0.1
WANG et al.: FUZZY-MODEL-BASED FINITE FREQUENCY FAULT DETECTION FILTERING DESIGN FOR 2-D NONLINEAR SYSTEMS 2109

By Algorithm 1 (Theorem 2) with κd = 0.2 , κ f = 0.8 , and where NY and NZ denote the right null spaces of Y and Z
θhd = −1.0 , θ̄dh = 1.0 , θvd = 0.1 , θ̄v = 2.0 , θ̄hf = 0.1 , and θ̄vf = 0.3 , and satisfy YNY = 0 and ZNZ = 0 , respectively.
d
one has the minimized κd γd2 − κ f γ2f = 2.3191 , with the follow- Lemma 2 [42]: Given matrices X, U, W, and ∆ subject to
∆T ∆ ≤ I , then X + Sym {U∆W} < 0 is solvable if there exists
ing fuzzy Lyapunov function matrices:
  δ > 0 satisfying X + δ−1 WT W + δUUT < 0.
 5.3242 −0.3189 
P1 = 
h

−0.3189 3.2822 References
  [1] C. Deng, W. Gao, and W. Che, “Distributed adaptive fault-tolerant
 15.6965 2.5541 
Pv1 =   output regulation of heterogeneous multi-agent systems with coupling
2.5541 2.8418 uncertainties and actuator faults,” IEEE/CAA J. Autom. Sinica, vol. 7,
no. 4, pp. 1098–1106, Jul. 2020.
 
 4.7436 −0.2639  [2] M. Zhou, Z. Cao, and Y. Wang, “Robust fault detection and isolation
P2 = 
h
 based on finite-frequency H− /H∞ unknown input observers and
−0.2639 3.2774 zonotopic threshold analysis,” IEEE/CAA J. Autom. Sinica, vol. 6, no. 3,
  pp. 750–759, May 2019.
 15.8045 2.4963 
Pv2 =   . [3] S. Yan, W. Sun, F. He, and J. Yao, “Adaptive fault detection and
isolation for active suspension systems with model uncertainties,” IEEE
2.4963 2.8745
Trans. Reliability, vol. 68, no. 3, pp. 927–937, Sept. 2019.
Note that by choosing Phl = Ph , Pvl = Pv , Theorem 2 will be [4] X. Xie, D. Ma, D. Yue, and J. Xia, “Gain-scheduling fault estimation
reduced to a common Lyapunov function based method. for discrete-time Takagi-Sugeno fuzzy systems: A depth partitioning
approach,” IEEE Trans. Circuits and Systems I: Regular Papers,
Applying Algorithm 1 (Theorem 2) with the above parame- vol. 69, no. 4, pp. 1693–1703, 2022.
ters, one has the minimized κd γd2 − κ f γ2f = 2.5284 with the fol- [5] S. Xiao and J. Dong, “Distributed fault-tolerant containment control for
lowing common Lyapunov function matrices: nonlinear multi-agent systems under directed network topology via
hierarchical approach,” IEEE/CAA J. Autom. Sinica, vol. 8, no. 4,
  pp. 806–816, Apr. 2021.
 5.7564 −0.1177 
P =
h   [6] D. Zhang, Z. Ye, and X. Dong, “Co-design of fault detection and
−0.1177 3.3312 consensus control protocol for multi-agent systems under hidden DoS
  attack,” IEEE Trans. Circuits and Systems I: Regular Papers, vol. 68,
 16.1361 2.4981  no. 5, pp. 2158–2170, May 2021.
P = 
v
 . [7] L. Zhang, H.-K. Lam, Y. Sun, and H. Liang, “Fault detection for fuzzy
2.4981 2.2048
semi-Markov jump systems based on interval type-2 fuzzy approach,”
It has been verified by the optimized performance index IEEE Trans. Fuzzy Systems, vol. 28, no. 10, pp. 2375–2388, Oct. 2020.
κd γd2 − κ f γ2f ( 2.3191 < 2.5284 ) that the fuzzy Lyapunov func- [8] H. Li, Z. Chen, L. Wu, H.-K. Lam, and H. Du, “Event-triggered fault
detection of nonlinear networked systems,” IEEE Trans. Cybernetics,
tion based fault detection approach is less conservative than vol. 47, no. 4, pp. 1041–1052, Apr. 2017.
the common Lypunov function based approach. [9] H.-N. Wu and Z.-P. Wang, “Observer-based H∞ sampled-data fuzzy
control for a class of nonlinear parabolic PDE systems,” IEEE Trans.
V. Conclusions Fuzzy Systems, vol. 26, no. 2, pp. 454–473, Apr. 2018.
[10] J. Zeng and X. Zhang, “An adaptive large neighborhood search for
This article has developed a fuzzy model based filtering single-machine batch processing scheduling with 2-D rectangular bin-
synthesis method for Roesser type 2-D nonlinear systems with packing constraints,” IEEE Trans. Reliability, vol. 71, no. 1, pp. 139–
disturbances and faults, whose frequencies are in a rectangu- 148, Mar. 2022.
lar finite frequency set and a low frequency set, respectively. [11] V. Kostyuk. Two-Dimensional Linear Systems. Berlin: Springer, 1985.
Based on the 2-D Fourier transform, a frequency based fault [12] R. P. Roesser, “A discrete state-space model for linear image
processing,” IEEE Trans. Automatic Control, vol. AC-20, no. 1, pp. 1–
detection filtering design method is proposed to generate a 10, Feb. 1975.
residual signal with both sensitivity to faults and robustness to [13] E. Fornasini and G. Marchesini, “Doubly-indexed dynamical systems:
external disturbances. With the generated residual signal, an State-space models and structural properties,” Mathematical Systems
Theorey, vol. 12, pp. 59–72, Dec. 1978.
evaluation function together with its threshold has been
[14] C. K. Ahn, L. Wu, and P. Shi, “Stochastic stability analysis for 2-D
designed, and then a finite frequency fault detection algo- Roesser systems with multiplicative noise,” Automatica, vol. 69,
rithm has been developed, based on which the fault is immedi- pp. 356–363, Jul. 2016.
ately detected once the evaluation function exceeds the thresh- [15] L. Wu, R. Yang, P. Shi, and X. Su, “Stability analysis and stabilization
of 2-D switched systems under arbitrary and restricted switchings,”
old. Automatica, vol. 59, pp. 206–215, Sept. 2015.
Appendix [16] Z. Q. Ge, “Exact controllability and exact observability of descriptor
infinite dimensional systems,” IEEE/CAA J. Autom. Sinica, vol. 8,
Lemma 1 (Projection Lemma) [42]: Given matrices W, Y, no. 12, pp. 1956–1963, Dec. 2021.
and Z, there exists a variable X satisfying [17] M. Wang, J. Qiu, H. Yan, Z. Li, and Y. Li, “Static output feedback
control for uncertain Roesser-type continuous-time two-dimensional
W + YT XT Z + ZT XY < 0 piecewise affine systems,” SCIENCE CHINA Information Sciences,
vol. 65, no. 11, p. 219204, Sept. 2022.
if and only if [18] H. Alikhani and N. Meskin, “Event-triggered robust fault diagnosis and
 control of linear Roesser systems: A unified framework,” Automatica,


 NYT WNY < 0, if NZ = 0, NY , 0



vol. 128, p. Article No. 109575, 2021.
 T [19] Z. Duan, J. Fu, C. K. Ahn, Z. Xiang, and I. Ghous, “Mixed l1/l- fault

 NZ WNZ < 0, if NY = 0, NZ , 0



detection for positive 2-D systems with distributed delays,” IEEE Trans.

N T WNY < 0, N T WNZ < 0, if NZ , 0, NY , 0 Systems, Man, and Cybernetics: Systems, vol. 52, no. 11, pp. 790–7100,
Y Z Nov. 2022.
2110 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 11, NO. 10, OCTOBER 2024

[20] J. Fu, Z. Duan, and Z. Xiang, “On mixed l1/l- fault detection observer Yakubovich-Popov lemma for discrete Roesser model,” IEEE Trans.
design for positive 2D Roesser systems: Necessary and sufficient Circuits and Systems-I: Regular Paper, vol. 55, no. 10, pp. 3223–2233,
conditions,” Journal of the Franklin Institute, vol. 359, pp. 160–177, Nov. 2008.
2022. [40] D.-W. Ding, H. Wang, and X. Li, “ H∞ /H∞ fault detection observer
[21] W. Ji, J. Qiu, and H.-K. Lam, “A new sampled-data output-feedback design for two-dimensional Roesser systems,” Systems & Control
controller design of nonlinear systems via fuzzy affine models,” IEEE Letters, vol. 82, pp. 115–120, 2015.
Trans. Cybernetics, vol. 52, no. 3, pp. 1681–1690, Mar. 2022. [41] L. V. Hien, H. Trinh, and N. T. Lan-Huong, “Delay-dependent energy-
[22] X. Xie, C. Wei, Z. Gu, and K. Shi, “Relaxed resilient fuzzy stabilization to-peak stability of 2-D time-delay Roesser systems with multiplicative
of discrete-time Takagi-Sugeno systems via a higher order time-variant stochastic noises,” IEEE Trans. Automatic Control, vol. 64, no. 12,
balanced matrix method,” IEEE Trans. Fuzzy Systems, vol. 30, no. 11, pp. 5066–5073, Dec. 2019.
pp. 5044–5050, Nov. 2022. [42] S. Boyd, L. El Ghaoui, F. Feron, and V. Balakrishnan. Linear Matrix
[23] J. Mao, X. Meng, and D. Ding, “Fuzzy set-membership filtering for Inequality in Systems and Control Theory. Philadelphia, PA: SIAM,
discrete-time nonlinear systems,” IEEE/CAA J. Autom. Sinica, vol. 9, 1994.
no. 6, pp. 1026–1036, Jun. 2022.
[24] P. Cheng, S. He, V. Stojanovic, X. Luan, and F. Liu, “Fuzzy fault Meng Wang received the B.Eng. degree in automa-
detection for Markov jump systems with partly accessible hidden tion from Northeastern University at Qinhuangdao,
information: An event-triggered approach,” IEEE Trans. Cybernetics, in 2011, and the M.Eng. degree in control science
vol. 52, no. 8, pp. 7352–7361, Aug. 2022. and engineering from Harbin Institute of Technol-
[25] Z. Zhang, H. Li, C. Wu, and Q. Zhou, “Finite frequency fuzzy H∞ ogy, in 2013, and the Ph.D. degree from the Depart-
control for uncertain active suspension systems with sensor failure,” ment of Mechanical and Biomedical Engineering,
IEEE/CAA J. Autom. Sinica, vol. 5, no. 4, pp. 777–786, July 2018. City University of Hong Kong, Hong Kong, China,
[26] X. Zhang, H. Wang, V. Stojanovic, P. Cheng, S. He, X. Luan, and F. in 2018. He is currently an Associate Professor at the
Liu, “Asynchronous fault detection for interval type-2 fuzzy School of Information Science and Engineering, East
nonhomogeneous higher level Markov jump systems with uncertain China University of Science and Technology.
transition probabilities,” IEEE Trans. Fuzzy Systems, vol. 30, no. 7, His research interests include robust control and filtering, fuzzy systems
pp. 2487–2499, Jul. 2022. and control, and their engineering applications.
[27] J. Han, X. Liu, X. Wei, and X. Hu, “Adaptive adjustable dimension
observer based fault estimation for switched fuzzy systems with
Huaicheng Yan received the B.Sc. degree in auto-
unmeasurable premise variables,” Fuzzy Sets and Systems, vol. 452,
matic control from Wuhan University of Technology,
pp. 149–167, Jan. 2023.
in 2001, and the Ph.D. degree in control theory and
[28] Y. Mu, H. Zhang, Z. Gao, and S. Sun, “Fault estimation for discrete- control engineering from Huazhong University of
time T-S fuzzy systems with unmeasurable premise variables based on Science and Technology, in 2007. In 2011, he was a
fuzzy Lyapunov functions,” IEEE Trans. Circuits and Systems-II, Research Fellow with the University of Hong Kong,
Express Briefs, vol. 69, no. 3, pp. 1297–1231, Mar. 2022. Hong Kong, China, for three months, and also a
[29] J. Han, X. Liu, X. Gao, and X. Wei, “Intermediate observer-based Research Fellow with the City University of Hong
robust distributed fault estimation for nonlinear multiagent systems with Kong, Hong Kong, China, in 2012, for six months.
directed graphs,” IEEE Trans. Industrial Informatics, vol. 16, no. 12, Currently, he is a Full Professor with the School of
pp. 7426–7436, Dec. 2020. Information Science and Engineering, East China University of Science and
[30] J. Qiu, W. Ji, and M. Chadli, “A novel fuzzy output feedback dynamic Technology. He is an Associate Editor of IEEE Transactions on Neural Net-
sliding mode controller design for two-dimensional nonlinear systems,” works and Learning Systems, International Journal of Robotics and Automa-
IEEE Trans. Fuzzy Systems, vol. 29, no. 10, pp. 2869–2877, Oct. 2021. tion, and IEEE Open Journal of Circuits and Systems.
His research interests include networked control systems, multi-agent sys-
[31] Y. Luo, Z. Wang, G. Wei, and F. E. Alsaadi, “ H∞ fuzzy fault detection tems and robotics.
for uncertain 2-D systems under Round-Robin scheduling protocol,”
IEEE Trans. Systems, Man, and Cybernetics: Systems, vol. 47, no. 8,
pp. 2172–2184, Aug. 2017. Jianbin Qiu (Fellow, IEEE) received the B.Eng. and
[32] X. Wang, Z. Fei, P. Shi, and J. Yu, “Zonotopic fault detection for 2-D Ph.D. degrees in mechanical and electrical engineer-
systems under event-triggered mechanism,” IEEE Trans. Cybernetics, ing from the University of Science and Technology
vol. 52, no. 5, pp. 3510–3518, May 2022. of China, in 2004 and 2009, respectively. He also
[33] Z. Hu, S. Liu, W. Luo, and L. Wu, “Intrusion-detector-dependent received the Ph.D. degree in mechatronics engineer-
distributed economic model predictive control for load rrequency ing from the City University of Hong Kong, Hong
regulation with PEVs under cyber attacks,” IEEE Trans. Circuits and Kong, China, in 2009.
Systems I: Regular Papers, vol. 68, no. 9, pp. 3857–3868, Sept. 2021. He is currently a Full Professor at the School of
Astronautics, Harbin Institute of Technology. He was
[34] R. Zhao, L. Liu, and G. Feng, “ H− /H∞ memory fault detection
an Alexander von Humboldt Research Fellow at the
filtering design for uncertain systems with finite frequency
Institute for Automatic Control and Complex Systems, University of Duis-
specifications,” Int. Journal of Robust and Nonlinear Control, vol. 31,
burg-Essen, Germany. His current research interests include intelligent and
no. 11, pp. 5381–5403, Jul. 2021.
hybrid control systems, signal processing, and robotics.
[35] T. Iwasaki and S. Hara, “Generalized KYP lemma: Unified frequency Prof. Qiu is a Fellow of IEEE and serves as the Chair of the IEEE Indus-
domain inequalities with design applications,” IEEE Trans. Automatic trial Electronics Society Harbin Chapter, China. He is an Associate Editor of
Control, vol. 50, no. 1, pp. 41–59, Jan. 2005. IEEE Transactions on Fuzzy Systems, IEEE Transactions on Cybernetics, and
[36] A. Chibani, M. Chadli, P. Shi, and N. B. Braiek, “Fuzzy fault detection IEEE Transactions on Industrial Informatics.
filter design for T-S fuzzy systems in the finite-frequency domain,”
IEEE Trans. Fuzzy Systems, vol. 25, no. 5, pp. 1051–1061, Oct. 2017.
Wenqiang Ji received the Ph.D. degree in control
[37] A. Chibani, M. Chadli, S. X. Ding, and N. B. Braiek, “Design of robust
science and engineering from the Harbin Institute of
fuzzy fault detection filter for polynomial fuzzy systems with new finite
Technology (HIT). He is currently a Professor in the
frequency specifications,” Automatica, vol. 93, pp. 42–54, Jul. 2018.
School of Artificial Intelligence, Hebei University of
[38] W. Li, Z. Xie, J. Zhao, P. K. Wong, H. Wang, and X. Wang, “Static- Technology. His main research interests include
output-feedback based robust fuzzy wheelbase preview control for modeling and control of fuzzy systems, nonlinear
uncertain active suspensions with time delay and finite frequency systems, sampled-data control and filtering design,
constraint,” IEEE/CAA J. Autom. Sinica, vol. 8, no. 3, pp. 664–678, Mar. robust control, and sliding mode control.
2021.
[39] R. Yang, L. Xie, and C. Zhang, “Generalized two-dimensional Kalman-