Professional Documents
Culture Documents
www.elsevier.com/locate/neucom
PII: S0925-2312(16)30394-0
DOI: http://dx.doi.org/10.1016/j.neucom.2016.05.042
Reference: NEUCOM17059
To appear in: Neurocomputing
Received date: 18 December 2015
Revised date: 29 April 2016
Accepted date: 13 May 2016
Cite this article as: Xiaoli Xu, Huaicheng Yan, Hao Zhang and Fuwen Yang, H∞
filtering for T-S fuzzy networked systems with stochastic multiple delays and
sensor faults, Neurocomputing, http://dx.doi.org/10.1016/j.neucom.2016.05.042
This is a PDF file of an unedited manuscript that has been accepted for
publication. As a service to our customers we are providing this early version of
the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting galley proof before it is published in its final citable form.
Please note that during the production process errors may be discovered which
could affect the content, and all legal disclaimers that apply to the journal pertain.
H ∞ filtering for T-S fuzzy networked systems with
stochastic multiple delays and sensor faults
Abstract
This paper deals with the problem of H∞ filter design for a class of nonlinear net-
worked systems based on T-S fuzzy model with multiple stochastic time-varying
delays, and sensor faults and packet dropouts are considered simultaneously. A
sequence of stochastic and independent variables, which obey the Bernoulli dis-
tribution, are introduced to depict stochastic time-varying delays. The possible
of sensor failure can be described by unrelated random variables taking val-
ues on a interval, and the packet dropouts are described as a set of Bernoulli
distributed white noises. The approach of piecewise quadratic Lyapunov func-
tion is applied to reduce the conservatism. The filter parameters are obtained
by solving a set of linear matrix inequalities. Finally, a simulation example is
provided to illustrate the effectiveness of the proposed filter design approach.
Keywords: H ∞ filtering, multiple time-varying delays, sensor faults, packet
dropouts, piecewise quadratic Lyapunov function
1. INTRODUCTION
It is generally known that the issue of state/signal estimation has been dis-
cussed in the areas of control and signal processing. One of the most famous
2
induced delays have been investigated by many researchers [11, 13, 19–24, 26,
27, 31–34, 36]. In [19], H∞ controller for discrete-time T-S fuzzy systems with
time-varying state delays has been investigated. A fuzzy controller has been
designed for nonlinear impulsive fuzzy systems with time-delay in [20]. Since
network delays are usually time-varying and stochastic, recently, the delays
have been modeled in various probabilistic ways [21–23]. On the other hand,
packet dropouts have attracted much attention because it may result in the bad
performance, and even instability of the system[23–25]. Robust H∞ filtering
for a class of nonlinear networked systems with randomly occurring distributed
delays, missing measurements and sensor saturation has been discussed in [23],
where the occurrence probability of the packet dropout phenomenon obeys an
individual and certain probabilistic distribution taking values on 0 and 1. In [24],
Dong et al. consider robust H∞ fuzzy output feedback control with multiple
probabilistic delays and multiple packet dropouts, where the packet dropout
phenomenon occurs randomly.
Besides, sensor faults always occur in the practical control systems, which
may affect the performance of systems. Therefore, there is a practical interest
to consider the sensor faults. Up to now, a great deal of literatures have been
reported on the sensor faults [34–39]. In [37], the control problem of a class
of T-S fuzzy systems with stochastic sensor faults has been studied. The fault
statistics of each sensor is individually quantified and stochastic sensor faults
and non-ideal network quality of services are coupled in a unified framework.
To the best of the authors’ knowledge, the problem of networked H ∞ filtering
for T-S fuzzy systems with stochastic sensor faults, packet dropouts and multi-
ple stochastic time-varying delays being considered simultaneously has not been
fully investigated, which motivates us to study on this problem. The main con-
tributions of this paper can be concluded as follows: (i) Networked H ∞ filtering
with multiple stochastic time-varying communication delays, stochastic sensor
faults and packet dropout phenomena are simultaneously considered in the T-S
fuzzy systems framework; (ii) An approach of piecewise quadratic Lyapunov
functional is adopted to reduce the conservatism of the results.
3
By concluding above discussion, in this paper, our aim is to provide the T-S
fuzzy-model-based piecewise H ∞ filter design for networked control systems,
which include multiple stochastic time-varying communication delays, sensor
faults and successive data missing phenomenon. Both the sensor faults and
packet dropouts in the measurement equation are considered. Moreover, packet
dropouts are described by a Bernoulli random process.
The rest of the paper is organized as follows. System descriptions and prob-
lem formulations are presented in section 2. In section 3, fuzzy filter is de-
signed with piecewise quadratic Lyapunov function. A simulation is conducted
to demonstrate that the performance of the system can be guaranteed with the
proposed approaches in section 4 and the conclusion is provided in section 5.
Notations. The notations though out the paper are fairly standard. The
superscript “T ” stands for matrix transpose; Rn denotes the n-dimensional Eu-
clidean space; Rm×n is the set of all real m×n matrices; and I n represents n×n
identity matrix; and 0m×n represents m × n zero matrix, respectively. P > 0
means that P is a real symmetric and positive definite. We use an asterisk (∗)
to represent a term that is induced by symmetry in symmetric matrices, and
diag{· · · } stands for a block-diagonal matrix, respectively. E{x} and E{x|y}
stand for the expectation of x and the expectation of x conditional on y, and
λmax (A) and λmin (A) denote the largest and smallest eigenvalue of the square
matrix A, respectively. L2 [0, ∞) denotes the space of square-integrable vector
functions over [0, ∞).
In this section, we use a T-S fuzzy model to represent the nonlinear physical
plant. The measured information received from the plant is transmitted via
the shared communication channel, where the sensor faults and the randomly
occurring data missing phenomena happen. In what follows, physical plant will
be modeled.
4
A. T-S Fuzzy Model of Nonlinear Physical Plant
Consider a discrete T-S fuzzy model with r fuzzy rules, and the i − th fuzzy
rule is as follows Ri : IF ξ1 (k) is W i1 , and · · · , ξ g (k) is W ig THEN
⎧
⎪ q
⎪
⎪ x(k + 1) = Ai x(k) + Adi αm (k)x(k − τm (k)) + Bi w(k)
⎪
⎪
⎨ m=1
where Ri (i ∈ = 1, 2, · · · , r) denotes the i−th fuzzy inference rule, r is the
number of fuzzy implications. ξ(k) = (ξ1 (k), (ξ2 (k), . . . , (ξg (k)) ∈ Rg is the
premise variable vector and assumed to be measurable. Wij (j = 1, 2, · · · , g)
are the fuzzy set, x(k) ∈ Rn is the state vector, y(k) ∈ Rm is the system
measurement output, and z(k) ∈ Rp is the signal to be estimated, w(k) ∈ Rq
is the external disturbance input vector belonging to L2 [0, ∞), and τ m (k)(m =
1, 2, · · · , q) are randomly time-varying communication delays. (Ai , Adi , B i , C i ,
Li ) stands for the ith local model of the fuzzy system.
The stochastic variables αm (k)(m = 1, 2, · · · , q) in (1) are Bernoulli dis-
tributed white noise sequences and satisfy
5
where
g
Wij (ξ(k))
j=1
hi (ξ(k)) = (3)
r g
W (ξ(k))
i=1 j=1
where ⎡ ⎤ ⎡ ⎤
A(h) Ad (h) B(h) Ai Adi Bi
⎢ ⎥ r ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢C(h) 0 0 ⎥ = h i ⎢C 0 0⎥
⎣ ⎦ i=1 ⎣ i ⎦
L(h) 0 0 Li 0 0
In order to design the filter based on piecewise Lyapunov functions, we
partition the premise variable space Ω ⊆ Rð into n̄ polyhedral regions by the
boundaries as in [15]
where θ is the set of the face indices of the polyhedral hull satisfying ∂Ωσ =
θ
θ (∂Ωσ ).
6
By partitioning the variable space Ω, in each region Ωs , we have
hi (ξ(k)) = 1
i∈φ(s)
where ⎡ ⎤ ⎡ ⎤
As (h) Ads (h) Bs (h) Ai Adi Bi
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ Cs (h) 0 0 ⎥= hi ⎢ C i 0 0 ⎥
⎣ ⎦ ⎣ ⎦
i∈φ(s)
Ls (h) 0 0 Li 0 0
B. Sensor Faults
7
C. Communication Channel with Packet Dropouts
where v(k) ∈ Rs is the measurement noise, and ỹ(k) is the measurement output
obtained from the plant with packet dropouts via sensor. β(k) is white noise
sequences obeying random binary distribution with
⎧
⎪
⎨P rob{β(k) = 1} = β̄
⎪
⎩P rob{β(k) = 0} = 1 − β̄
D. Filtering-Error System
8
where
⎡ ⎤ ⎡ ⎤
As (h) 0 0 α̃m (k)Ads (h) 0 0
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
Ās (h) = ⎢ 0 Af s (1 − β̄)Bf s ⎥ , Ãdsm = ⎢ 0 0 0⎥
⎣ ⎦ ⎣ ⎦
0 0 1 − β̄ 0 0 0
⎡ ⎤ ⎡ ⎤
ᾱm Ads (h) 0 0 B (h) 0
⎢ ⎥ ⎢ s ⎥
⎢ ⎥ ⎢ ⎥
Ādsm = ⎢ 0 0 0⎥ , B̄s (h) = ⎢ 0 Bf s D ⎥
⎣ ⎦ ⎣ ⎦
0 0 0 0 D
⎡ ⎤
0 ⎡ ⎤
⎢ ⎥
⎢ ⎥ w(k)
B̄1s = ⎢Bf s ⎥ , w̄(k) = ⎣ ⎦ , e3 = [ 0 0 I ]
⎣ ⎦ v(k)
I
C̄s (h) = Cs (h) 0 0 , L̄s (h) = Ls (h) −Lf s 0
and α̃m (k) = αm (k) − ᾱm . It is easy to obtain that E {α̃m (k)} = 0 and
E {α̃m (k)α̃m (k)} = ᾱm (1 − ᾱm ), where m = 1, 2, · · · , q.
Definition 2.2. The filtering error system in (14) is said to be stochastically
9
3. MAIN RESULTS
In this section, the problems of H∞ performance and filter design for the
filtering-error system (14) will be investigated.
Theorem 1. Consider the system in (1), and assume the filter parameters
Af i , Bf i , Lf i in (14) and H∞ attenuation level γ > 0 are given. The filtering-
error system in (14) is stochastically stable in the mean square if there exist
matrices PsT = Ps > 0(s ∈ Υ) satisfying the following linear matrix inequality
(LMI) ⎡ ⎤
Γ11 + L̄si ∗ ∗
⎢ ⎥
⎢ ⎥
Γ=⎢ Γ21 Γ22 ∗ ⎥<0 (15)
⎣ ⎦
Γ31 Γ32 B̄iT Pl B̄i − γ 2 I
where
Ξ1 = (Āsi + β̄ B̄1s Λ̄C̄i )T Pl (Āsi + β̄ B̄1s Λ̄C̄i ), Ξ2 = 2μ21 C̄iT Λ̄T B̄1s T
Pl B̄1s Λ̄C̄i
m
Ξ3 = 2μ21 eT3 B̄1s
T
Pl B̄1s e3 , Ξ4 = (μ21 + β̄ 2 ) δi2 C̄iT ElT B̄1s
T
Pl B̄1s El C̄i
l=1
q
Γ11 = Ξ1 + Ξ2 + Ξ3 + Ξ4 − Ps + (τ2 − τ1 + 1)Qm
m=1
Γ21 = ẐiT Pl (Āsi + β̄ B̄1s Λ̄C̄i ), Ãm = ᾱm (1 − ᾱm )ÂTdi Pl Âdi
Γ22 = diag −Q1 + Ã1 , −Q2 + Ã2 , · · · , −Qq + Ãq + ẐiT Pl Ẑi
Γ31 = B̄iT Pl (Āsi + β̄ B̄1s Λ̄C̄i ), Γ32 = B̄iT Pl Ẑi
Ẑi = Ādi1 Ādi2 · · · Ādiq , μ21 = β̄(1 − β̄)
Proof. Considering the system (14), we choose following Lyapunov function
where
V1 (η(k)) = η T (k)Ps η(k), s ∈ Υ
q
k−1
V2 (η(k)) = η T (i)Qm η(i)
m=1 i=k−τm (k)
q −τ1
k−1
V3 (η(k)) = η T (i)Qm η(i)
m=1 n=−τ2 +1 i=k+n
Define
V (η(k)) = E {V (η(k + 1)) − V (η(k))}
10
When w̄(k) = 0, by taking mathematic expectation and difference along the
trajectory of system (14), we have
V1 (η(k)) = E η T (k + 1)Pl η(k + 1) − η T (k)Ps η(k)
+β̄ B̄1s (Λ − Λ̄)C̄s (h)η(k) + (β(k) − β̄)B̄1s Λ̄C̄s (h)η(k) + β̄ B̄1s Λ̄C̄s (h)η(k)
q
+ Ādsm (h) + Ãdsm (h) η(k − τm (k))]T Pl [Ās (h)η(k)
m=1
+(β̄ − β(k))B̄1s e3 η(k) + (β(k) − β̄)B̄1s (Λ − Λ̄)C̄s (h)η(k)
Consider
⎧⎡ ⎤T ⎡ ⎤⎫
⎪
⎪ α̃m (k)Ads (h) 0 0 ⎪ ⎪
⎪
⎨⎢ m
α̃ (k)Ads (h) 0 0
⎥ ⎢ ⎥⎪⎬
⎢ ⎥ ⎢ ⎥
E ÃTdsm (h)Pl Ãdsm (h) = E ⎢ 0 0 0⎥ P l ⎢ 0 0 0 ⎥
⎪
⎪ ⎣ ⎦ ⎣ ⎦⎪⎪
⎪
⎩ ⎪
0 0 0 0 0 0 ⎭
⎧ ⎡ ⎤T ⎡ ⎤⎫
⎪
⎪ Ads (h) 0 0 ⎪ ⎪
⎪
⎨
A (h) 0
⎢ ds
0
⎥ ⎢ ⎥⎪⎬
2 ⎢ ⎥ ⎢ ⎥
= E α̃m (k) ⎢ 0 0 0 ⎥ P l ⎢ 0 0 0 ⎥
⎪
⎪ ⎣ ⎦ ⎣ ⎦⎪⎪
⎪
⎩ ⎪
0 0 0 0 0 0 ⎭
= ᾱm (1 − ᾱm )ÂTds (h)Pl Âds (h) (17)
with ⎡ ⎤
Ads (h) 0 0
⎢ ⎥
⎢ ⎥
Âds (h) = ⎢ 0 0 0⎥
⎣ ⎦
0 0 0
11
By combining (11),(16) and (17), we have
q
V1 (η(k)) = E{[Ās (h)η(k) + β̄ B̄1s Λ̄C̄s (h)η(k) + Ādsm (h)η(k − τm (k))]T
m=1
q
Pl [Ās (h)η(k) + β̄ B̄1s Λ̄C̄s (h)η(k) + Ādsm (h)η(k − τm (k))]
m=1
q
T
+ ᾱm (1 − ᾱm )η (k − τm (k)) ÂTds (h)Pl Âds (h)η(k − τm (k))
m=1
m
+μ21 η T (k)eT3 B̄1s
T
Pl B̄1s e3 η(k) + β̄ 2
δl2 η T (k)C̄sT (h)ElT B̄1s
T
Pl B̄1s
l=1
m
El C̄s (h)η(k) + μ21 δl2 η T (k)C̄sT (h)ElT B̄1s
T
Pl B̄1s El C̄s (h)η(k)
l=1
+μ21 η T (k)C̄sT (h)Λ̄T B̄1s
T
Pl B̄1s Λ̄C̄s (h)η(k)
12
Next, by taking the difference of V2 (η(k)) and V3 (η(k)), we obtain
q
V2 (η(k)) = [η T (k)Qm η(k) − η T (k − τm (k))Qm η(k − τm (k))
m=1
1
k−τ
+ η T (i)Qm η(i)] (21)
i=k−τ2 +1
q 1
k−τ
V3 (η(k)) = [(τ2 − τ1 )η T (k)Qm η(k) − η T (i)Qm η(i)] (22)
m=1 i=k−τ2 +1
Letting
T
ς(k) = η T (k) η T (k − τ1 (k)) ··· η T (k − τq (k))
≤ ς T (k)Σ(h)ς(k) (23)
where
⎡ ⎤
Σ11 ∗
Σ(h) = ⎣ ⎦
Σ21 Σ22
with
Σ111 = (Ās (h) + β̄ B̄1s Λ̄C̄s (h))T Pl (Ās (h) + β̄ B̄1s Λ̄C̄s (h))
m
Σ112 = (μ21 + β̄ 2 ) δl2 η T (k)C̄sT (h)ElT B̄1s
T
Pl B̄1s El C̄s (h)η(k)
l=1
Σ113 = 2μ21 η T (k)eT3 B̄1s
T
Pl B̄1s e3 η(k)
Σ114 = 2μ21 η T (k)C̄sT (h)Λ̄T B̄1s
T
Pl B̄1s Λ̄C̄s (h)η(k)
q
Σ11 = Σ111 + Σ112 + Σ113 + Σ114 − Ps + (τ2 − τ1 + 1)Qm
m=1
Σ21 = ẐsT Pl (Ās (h) + β̄ B̄1s Λ̄C̄s (h))
Σ22 = diag −Q1 + Ã1 , −Q2 + Ã2 , · · · , −Qq + Ãq + ẐsT Pl Ẑs
According to theorem 1, via Schur complement, (15) implies Σ(h) < 0,
which means V (η(k)) < 0. Therefore, we have E {V (η(k + 1)) − V (η(k))} <
−λmin (−Σ(h))ς T (k)ς(k).
13
Summing up both sides, from k = 0, 1, · · · , N we have
E |ς(k)|2 ≤ λ−1
min (−Σ(h))(V (0) − E {V (N + 1)}) (24)
where ⎡ ⎤
Σ11 + L̄Ts (h)L̄s (h) ∗ ∗
⎢ ⎥
⎢ ⎥
Σ̄ = ⎢ Σ21 Σ22 ∗ ⎥
⎣ ⎦
Σ31 B̄sT (h)Pl Ẑs Σ33
with
Σ31 = B̄sT (h)Pl (Ās (h) + β̄ B̄1s Λ̄C̄s (h)), Σ33 = B̄sT (h)Pl B̄s (h) − γ 2 I
According to (15) in theorem 1, for all non-zero w̄(k), we have
∞
" #
V (η(k)) + eT (k)e(k) − γ 2 w̄T (k)w̄(k) ≤ 0 (29)
k=0
14
which means that
∞
∞
||e(k)||2 ≤ γ 2 ||w̄(k)||2 (30)
k=0 k=0
where
q
Φ = Pl − sym(Gs ), Φ̃ = Im ⊗ Φ, Π11 = −Ps + (τ2 − τ1 + 1)Qm
m=1
Π22 = diag −Q1 + Ã1 , −Q2 + Ã2 , · · · , −Qq + Ãq
⎡ & ⎤
δ1 μ21 + β̄ 2 GTs B̄1s E1 C̄i
⎢ ⎥
⎢ .. ⎥
Π8si = ⎢ . ⎥
⎣ & ⎦
δm μ21 + β̄ 2 GTs B̄1s Em C̄i
Proof. It follows from (33) that Pl − Gs − GTs < 0, and we know Gs is a
non-singular matrix. Note that (Pl − Gs ) Pl−1 (Pl − Gs ) ≥ 0, which implies
T
15
Together with (33), which yields
⎡ ⎤
Π11 ∗ ∗ ∗ ∗ ∗ ∗ ∗
⎢ ⎥
⎢ ⎥
⎢ 0 Π22 ∗ ∗ ∗ ∗ ∗ ∗⎥
⎢ ⎥
⎢ ⎥
⎢ 0 0 −γ 2 I ∗ ∗ ∗ ∗ ∗⎥
⎢ ⎥
⎢ ⎥
⎢ L̄si 0 0 −I ∗ ∗ ∗ ∗⎥
⎢ ⎥<0
⎢ T$ % ⎥ (33)
⎢Gs Āsi + β̄ B̄1s Λ̄C̄i T T
Gs Ẑs Gs B̄i 0 Φ1 ∗ ∗ ∗⎥
⎢ √ ⎥
⎢ ⎥
⎢ 2μ1 GTs B̄1s e3 0 0 0 0 Φ1 ∗ ∗⎥
⎢ √ ⎥
⎢ ⎥
⎢ T
2μ1 Gs B̄1s Λ̄C̄i 0 0 0 0 0 Φ1 ∗⎥
⎣ ⎦
Π8si 0 0 0 0 0 0 Φ̃1
16
G1s ,Ss ,Ws ,Āf i ,B̄f i ,Lf i satisfying
⎡ ⎤
Π ∗ ∗ ∗ ∗ ∗ ∗ ∗
⎢ 11 ⎥
⎢ ⎥
⎢ 0 Π22 ∗ ∗ ∗ ∗ ∗ ∗⎥
⎢ ⎥
⎢ ⎥
⎢ 0 0 −γ 2 I ∗ ∗ ∗ ∗ ∗⎥
⎢ ⎥
⎢ ⎥
⎢ L̄si 0 0 −I ∗ ∗ ∗ ∗⎥
⎢ ⎥<0 (35)
⎢ ⎥
⎢Π3si Π4si Π5si 0 Ω ∗ ∗ ∗⎥
⎢ ⎥
⎢ ⎥
⎢Π6si 0 0 0 0 Ω ∗ ∗⎥
⎢ ⎥
⎢ ⎥
⎢Π7si 0 0 0 0 0 Ω ∗⎥
⎣ ⎦
Π̂8si 0 0 0 0 0 0 Ω̃
where ⎡ ⎤
sym(G1s ) S s + Ws G3s + GT6s
⎢ ⎥
⎢ ⎥
Ω = Pl − ⎢ ∗ sym(Ws ) YsT ⎥, Ω̃ = Im ⊗ Ω
⎣ ⎦
∗ ∗ sym(G8s )
⎡ ⎤
Π31si Āf s (1 − β̄)B̄f s + (1 − β̄)GT6s
⎢ ⎥
⎢ ⎥
Π3si = ⎢Π32si Āf s (1 − β̄)B̄f s + (1 − β̄)YsT ⎥
⎣ ⎦
Π33si 0 (1 − β̄)G8s T
Π31si = GT1s Ai + β̄ B̄f s Λ̄Ci + β̄GT6s Λ̄Ci , Π32si = SsT Ai + β̄ B̄f s Λ̄Ci + β̄YsT Λ̄Ci
Π33si = GT3s⎡Ai + β̄GT8s Λ̄Ci ⎤
ᾱ1 GT1s Adi 0 0 ᾱ2 GT1s Adi 0 0 ··· ᾱq GT1s Adi 0 0
⎢ ⎥
⎢ ⎥
Π4si = ⎢ ᾱ1 SsT Adi 0 0 ᾱ2 GT1s Adi 0 0 ··· ᾱq SsT Adi 0 0⎥
⎣ ⎦
ᾱ1 GT3s Adi 0 0 ᾱ2 GT3s Adi 0 0 · · · ᾱq GT3s Adi 0 0
⎡ ⎤ ⎡ √ ⎤
GT1s Bi B̄f s D + GT6s D 0 0 2μ1 (B̄f s + μ1 GT6s )
⎢ ⎥ ⎢ √ ⎥
⎢ T ⎥ ⎢ ⎥
Π5si = ⎢ Ss Bi B̄f s D + YsT D ⎥ , Π6si = ⎢0 0 2μ1 (B̄f s + μ1 YsT ) ⎥
⎣ ⎦ ⎣ √ ⎦
GT3s Bi GT8s D 0 0 2μ1 GT8s
⎡√ ⎤
2μ (B̄ Λ̄C + GT6s Λ̄Ci ) 0 0
⎢√ 1 fs i ⎥
⎢ ⎥
Π7si = ⎢ 2μ1 (B̄f s Λ̄Ci + YsT Λ̄Ci ) 0 0⎥
⎣ √ ⎦
2μ1 GT8s Λ̄Ci 0 0
⎡ & ⎤
δ μ2 + β̄ 2 (B̄f s E1 Ci + GT6s E1 Ci ) 0 0
⎢ 1& 1 ⎥
⎢ ⎥
Π8si = ⎢ δ1 μ21 + β̄ 2 (B̄f s E1 Ci + YsT E1 Ci ) 0 0⎥
⎣ & ⎦
δ1 μ21 + β̄ 2 (GT8s E1 Ci ) 0 0
17
Furthermore, the parameters of an admissible piecewise H∞ filter are given
by
Af i = Ws−1 Āf i
Bf i = Ws−1 B̄f i
Proof. In order to simplify the filter synthesis procedure, we introduce the
slack matrix variable Gs as
⎡ ⎤
G1s G2s G3s
⎢ ⎥
⎢ ⎥
Gs = ⎢G4s G5s 0 ⎥ (36)
⎣ ⎦
G6s G7s G8s
and the following matrix variables Ws = GT4s G−1 −1
5s G4s , Ss = G2s G5s G4s ,Ys =
G7s G−1
5s G4s .
⎡ ⎤T ⎡ ⎤ ⎡ ⎤
In 0 0 In 0 0 G1s + GT1s S s + Ws G3s + GT6s
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ T ⎢ ⎥ ⎢ ⎥
⎢0 G−1
4s G3s 0 ⎥ (Gs +Gs ) ⎢ 0 G−1
4s G3s 0⎥=⎢ ∗ Ws + WsT YsT ⎥
⎣ ⎦ ⎣ ⎦ ⎣ ⎦
0 0 Im 0 0 Im ∗ ∗ G8s + GT8s
(37)
Thus, one can represent (36) with the following form without losing gener-
ality
⎡ ⎤
G1s Ss G3s
⎢ ⎥
⎢ ⎥
Gs = ⎢WsT WsT 0 ⎥ (38)
⎣ ⎦
G6s Ys G8s
Define Āf i = Ws Af i , B̄f i = Ws Bf i , and substitute (38) into (31), then (35)
can be obtained. The proof is completed.
4. ILLUSTRATIVE EXAMPLE
18
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
1 0.045 0.14 −0.02 −0.11
A1 = ⎣ ⎦, Ad1 = ⎣ ⎦, B1 = ⎣ ⎦
−0.8558 0.7894 −0.2 0.094 −0.412
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
1 0.025 0.14 −0.02 −0.5
A2 = ⎣ ⎦, Ad2 = ⎣ ⎦, B2 = ⎣ ⎦
−0.6315 0.8018 −0.2 0.094 −0.184
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
1 0.035 0.14 −0.02 −0.3
A3 = ⎣ ⎦, ⎦ , B3 = ⎣
Ad3 = ⎣ ⎦
−0.7558 0.6894 −0.2 0.094 −0.105
L1 = 0.55 −1 , L2 = 0.44 −1 , L3 = 0.62 −1
C1 = 0.9949 −0.8 , C2 = 0.8270 −0.8 , C3 = 0.6366 −0.8
In this example, we choose the following parameters as D = 0.1, q = 2,
ᾱ1 = 0.1, ᾱ2 = 0.1, β̄ = 0.9, τ1 = 2, τ2 = 3. In the following, two possible cases
are considered, which can prove the effectiveness of the proposed filter with and
without sensor faults.
Case 1: When the sensors are in good state, that is Λ = I. According
to theorem 3, we can obtain the performance index γ = 17.3627 and filter
parameters⎡as follows ⎤ ⎡ ⎤
0.6830 0.0834 −0.0608
Af 1 = ⎣ ⎦ , Bf 1 = ⎣ ⎦ , Lf 1 = 0.2474 1.0401
0.1263 0.6080 0.2966
⎡ ⎤ ⎡ ⎤
1.0745 0.1274 −0.0986
Af 2 = ⎣ ⎦ , Bf 2 = ⎣ ⎦ , Lf 2 = −0.4265 1.0078
−0.5004 0.5819 0.3036
⎡ ⎤ ⎡ ⎤
0.9190 0.0742 −0.1004
Af 3 = ⎣ ⎦ , Bf 3 = ⎣ ⎦ , Lf 3 = −0.5368 0.9666
−0.4906 0.5757 0.3164
Besides, we suppose w(k) = exp(−0.2k)sink, v(k) = sin(10k + 1)/(3k + 1).
T T
With the initial condition xT1 (0) xT2 (0) = 0 0 , we obtain the following
simulation results from Fig. 1 to Fig. 4. Membership function is shown in
Fig. 1. In Fig. 2, output measurements with and without packet dropouts are
depicted. The simulation results of the filter error and estimation under zero
initial condition are depicted in Fig. 3 and Fig. 4, respectively.
19
Fig.1. The membership function.
0.4
y(k)
0.3
0.2
0.1
−0.1
−0.2
−0.3
−0.4
0 20 40 60 80 100
Time in samples
20
0.4
e(k)
0.3
0.2
0.1
−0.1
−0.2
−0.3
−0.4
0 20 40 60 80 100
Time in samples
0.4
z(k)
zf(k)
0.3
0.2
0.1
−0.1
−0.2
−0.3
−0.4
0 20 40 60 80 100
Time in samples
Case 2: The sensor faults occurs and the parameters for the sensor failure
are Λ = 0.8 and δ = 0.05. The other parameters are the same as case 1. Besides,
we assume the fault occurs at the 50th time instant, and the fault function is
choosed as follows
⎧
⎪
⎨0.1 ∗ sin(k) 50 ≤ k ≤ 70
f (k) =
⎪
⎩0 else
21
The response of measurement output with sensor fault is depicted in Fig.
5. Fig. 6 and Fig. 7 show the result of the filter error and estimation signal
with sensor fault, respectively. It can be seen from the picture that the designed
filter can satisfy the system performance in the presence of sensor fault.
0.4
y(k)
0.3
0.2
0.1
−0.1
−0.2
−0.3
−0.4
0 50 100 150 200
Time in samples
0.4
e(k)
0.3
0.2
0.1
−0.1
−0.2
−0.3
−0.4
0 50 100 150 200
Time in samples
22
0.4
z(k)
zf(k)
0.3
0.2
0.1
−0.1
−0.2
−0.3
−0.4
0 50 100 150 200
Time in samples
5. CONCLUSION
In this paper, the H∞ filtering problem for a class of T-S fuzzy system
with multiple stochastic time-varying delay, sensor faults and successive packet
dropouts has been investigated. The packet dropouts occur randomly and obey
the Bernulli distribution taking on values of 0 and 1. An approach of piece-
wise quadratic Lyapunov function has been adopted to lessen the conservatism,
and it has been proved that the filtering error system is stochastically stable in
mean square and H∞ performance is guaranteed. An example has been given
to demonstrate the effectiveness of the proposed approach. Further research
directions would include the investigation on the distributed H ∞ filtering and
control for nonlinear systems with quantization and event-triggered communi-
cation strategy.
Acknowledgements
23
References
24
[10] T. Takagi, M. Sugeno, Fuzzy identification of systems and its application
to modeling and control, IEEE Transations on Systems, Man, and Cyber-
netics, 1985, 15(1): 116–132.
[15] H. Wang, G. Yang, Piecewise controller design for affine fuzzy systems via
dilated linear matrix inequality characterizations, ISA Transactions, 2012,
51(12): 771-777.
[16] Z. Xia, J. Li, Delay-dependent H∞ control for T-S fuzzy systems based on
a switching fuzzy model and piecewise Lyapunov function, Acta Automatic
Sinica, 2009, 35(9): 1235-1239.
[18] M. Gao, L. Sheng, Y. R. Liu, Robust H∞ control for T-S fuzzy systems
subject to missing measurements with uncertain missing probabilities, Neu-
rocomputing, 2016, 193(12): 235-241.
25
[19] H. Zhang, C. Dang, Piecewise H∞ controller design of uncertain discrete-
time fuzzy systems with time delays, IEEE Transactions on Fuzzy Systems,
2008, 16(6): 1649-1654.
[20] H. Zhang, H. Yan, T. Liu, Q. Chen, Fuzzy controller design for nonlin-
ear impulsive fuzzy systems with time-delay, IEEE Transactions on Fuzzy
Systems, 2011, 19(5): 844-856.
[26] H. Zhang, H. Yan, F. Yang, Q. Chen, Quantized control design for impul-
sive fuzzy networked systems, IEEE Transactions on Fuzzy Systems, 2011,
19(6): 1153-1162.
[27] H. Li, C. Wu, S. Yin, H. K. Lam, Observer-based fuzzy control for nonlinear
networked systems under unmeasurable premise variables, IEEE Transac-
tions on Fuzzy Systems, DOI: 10.1109/TFUZZ.2015.2505331.
26
[28] W. Guan, F. C. Liu, Non-fragile fuzzy dissipative static output feedback
control for Markovian jump systems subject to actuator saturation. Neu-
rocomputing, 2016, 193(6): 123-132.
[31] H. Li, Y. Gao, P. Shi, H. K. Lam, Observer-based fault detection for non-
linear systems with sensor fault and limited communication capacity, IEEE
Transactions on Automatic Control, DOI: 10.1109/TAC.2015.2503566.
[35] H. Li, Z. Chen, L. Wu, H. K. Lam and H. Du, Event-triggered fault detec-
tion of nonlinear networked systems, IEEE Transactions on Cybernetics,
DOI:10.1109/TCYB.2016.2536750.
27
[37] C. Peng, M. R. Fei, E. G. Tian, Networked control for a class of T-S fuzzy
systems with stochastic sensor faults, Fuzzy Sets and Systems, 2013, 212(8):
62-77.
[38] A. C. Li, S. J. Yi, X. Wang, New reliable H∞ filter design for networked
control systems with external disturbances and randomly occurring sensor
faults, Neurocomputing, 2016, 185(3): 21-27.
28