This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2812754, IEEE
Transactions on Industrial Informatics

Fault-tolerant control of time-delay Markov jump

systems with Itô stochastic process and output
disturbance based on sliding mode observer
Hongyan Yang, Yuchen Jiang and Shen Yin

Abstract—This work focuses on the fault-tolerant control with human operators, and other areas, considerable attention
(FTC) problem of Markov jump systems (MJS) with Itô stochas- has been devoted to the stochastic systems governed by Itô
tic process and output disturbances. Such a problem widely exists stochastic differential equations [21], [22], [23]. In fact, all
in practical systems such as mobile manipulator systems. Since
MJS can suitably describe mobile manipulator systems, in this the phenomena mentioned above may occur simultaneously
work, a new approach based on the MJS model is proposed. in real plants, and therefore researches are demanded for the
Firstly, a proportional-derivative sliding mode observer and an systems with hybrid characteristics.
observer-based controller are designed and synthesized. Two new In previous studies, sliding mode observer (SMO) design
theorems are derived to ensure the close-loop stochastic stability has been proposed for fault-free control systems [24], [25], [9],
and the reachability of the sliding mode surface. Compared
with the existing works, the system model is more general, [10], [26], [27], [28], [29]. SMO holds the superiorities of fast
which could describe a larger variety of plants or processes. response, excellent transient performance, and insensitiveness
The controller design procedure is simplified by solving the to perturbations and disturbances. In more recent studies, SMO
sliding mode parameters and the controller gain simultaneously was extended to be utilized for FTC purposes. For instance, in
with only one linear matrix inequality problem. In addition, [8], actuator faults and sensor faults were considered for time-
the augmented fault vector can be reconstructed by employing
a descriptor sliding mode observer. Simulations are provided invariant systems. The authors in [7] coped with the Markov
to demonstrate the validity of the derived theorems and the jump system (MJS) with time-delay and nonlinearities for the
effectiveness of the proposed algorithm. FTC of sensor faults. By constructing a SMO, the sensor fault
Index Terms—Sliding mode observer, Fault-tolerant control, signals are online estimated and compensated. It is worth to
Fault diagnosis, Output disturbances. mention that, in [7], the observer design and controller design
are in separate steps and require solving two linear matrix
I. I NTRODUCTION
inequalities, which can be further simplified.
Owing to the increasing demand for stability and perfor- In this paper, FTC against sensor faults of a time-delay MJS
mance of the control systems, fault-tolerant control (FTC) has with Itô stochastic and output disturbances is investigated. Us-
become a popular research domain over the past decades. ing a proportional-derivative (PD) sliding mode observer, the
Fruitful outcomes have been reported [1], [2], [3], [4], [5], sufficient conditions of the closed-loop stochastic stability and
[6]. From the perspective of design technique, FTC can be the sliding mode surface reachability are given. In addition, a
categorized into passive approaches (including simultaneous fault estimation (FE) and FTC system design and synthesis
stabilization of generalized system [7], [8], robust FTC [9], approach is presented.
[10], etc.), and active approaches (controller reconstruction The contributions of this work are as follows. First, the
[11], adaptive FTC [12], intelligent FTC [13], etc.). Specif- system model adopted in this paper includes output distur-
ically, passive FTC considers system perturbations, sensor bance and stochastic noise terms, and introduces time-delay
faults and actuator faults during the design phase, and avoids to the state variables. As a result, the proposed approach
the effort to design additional explicit fault detection and can characterize more features in practical applications than
isolation (FDI) systems, since FDI is still an open topic for the existing approaches (such as those in [7]). Second, the
complicated systems [14], [15]. observer coefficients and the controller gain are solved in
The major difficulty of passive FTC system design is one linear matrix inequality (LMI) problem, instead of two
to derive the sufficient conditions for closed-loop system as in most existing works (such as [7]). This reduces the
stabilization, especially when various dynamic characteristics design complexity and the computational load. Third, the
cannot be ignored. For instance, time-varying parameters often reachability of the sliding mode surface under the closed-
exist in the domain of manufacturing, biology and economics, loop configuration is guaranteed, and the faults can be directly
and time-delay is a remarkable feature of process control sys- identified with the proposed sliding mode observer.
tems and network control systems [16], [17], [18], [19], [20]. Notations. In this work, E{·} represents the mathematical
On another research forefront, since stochastic phenomenon expectation. || · || denotes the Euclidean norm of a vector.
typically exhibits in mechanical systems, economics, systems Rm×n denotes an m × n dimensional matrix with elements of
real numbers. Given a symmetric matrix A, A > 0(< 0) rep-
Hongyan Yang, Yuchen Jiang and Shen Yin are with Harbin Insti-
tute of Technology, China. Email of corresponding author Shen Yin: resents that A is a positive-definite (negative-definite) matrix.
[email protected] The matrix elements denoted by the symbol “∗” refer to the

1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2812754, IEEE
Transactions on Industrial Informatics

transposed elements in the diagonal positions.

II. P ROBLEM FORMULATION
Consider the following MJS with time delays, sensor faults,
Itô stochastic process as well as output disturbances defined
in the fixed probability space (Ω, F, P)


 dx(t) = [A(rt )x(t) + B(rt )u(t) + Ah (rt )x(t − h)]dt
+ [Bw (rt )x(t) + Bh (rt )x(t − h)]dw(t)

(1)

 x(t) = φ(t), t ∈ [−h, 0], r(0) = r0 ,
ys (t) = C(rt )x(t) + Fs (rt )fs (t) + Dd (rt )d(t)


where x(t) ∈ Rn represents the state variable, ys (t) ∈ Rp

represents the measurement output and u(t) ∈ Rm denotes Fig. 1. The architecture of the FTC system
the input. fs (t) ∈ Rs and d(t) ∈ Rd represent the sensor
faults and the bounded disturbances, respectively. {rt , t ≥ 0}
which takes values in S , {1, 2, · · · , N } denotes a Markov III. M AIN RESULTS
process with continuous trajectories on (Ω, F, P) in the right- A. Augmented system and sliding mode observer formulation
half plane. The transition probability with the generator matrix
Π , [πij ] is in the following form In this section, we firstly define some variables and aug-
mented matrices to synthesize a FTC strategy for plant (1),
Pij = Pr{rt+∆ = j|rt = i}  
x(t)
f (t) + f˙s (t)
  
1 + πii ∆ + o(∆) if i = j x̄(t) , Fsi fs (t) , f¯(t) , s (4)
= (2) d(t)
πij ∆ + o(∆) if i 6= j Ddi d(t)
P    
where i, j ∈ S, πij > 0, i 6= j and πii = − j6=i πij < 0; In 0 0 Ai 0 0
∆ > 0 and lim∆→0 o(∆)/∆ = 0. A(rt ) ∈ Rn×n , B(rt ) ∈ Ē ,  0 Ip 0 Āi ,  0 −Ip 0  (5)
Rn×m , C(rt ) ∈ Rp×n , Fs (rt ) ∈ Rp×s , Dd (rt ) ∈ Rp×d refer 0 0 0
 
0

0 −Ip

to the system matrices. For notational simplicity, the matrices 0n×s 0n×d Bi
A(rt ), B(rt ), C(rt ), Fs (rt ), C(rt ), Dd (rt ) will be represented N̄i ,  Fsi 0p×d  , B̄i , 0p×m  (6)
by Ai , Bi , Ci , Fsi , Ddi where rt = i, i ∈ S. 0p×s Ddi 0p×m
The following assumptions and preliminaries are required  
C̄i , Ci Ip Ip (7)
in this work.
(A1) (Ai , Ci ) is observable.
   
Ahi 0 0 Bwi 0 0
(A2) The matrices Bi , Fsi and Ddi are of full column rank, Āhi ,  0 0 0 B̄wi ,  0 0 0 (8)
and the matrix Ci is a full row rank matrix. 0 0 0 0 0 0
 
(A3) The output disturbances and sensor faults are bounded. Bhi 0n×p 0n×p
||d(t)|| ≤ rd , ||fs (t)|| ≤ rs1 , ||f˙s (t)|| ≤ rs2 , where B̄hi , 0p×n 0p×p 0p×p  (9)
rd > 0, rs1 > 0 and rs2 > 0 represent known constants. 0p×n 0p×p 0p×p
Definition 1.[8] The plant (1) is said to be stochastic stable, Then, the augmented plant is constructed as follows
if for arbitrary initial conditions x0 ∈ Rn , r0 ∈ S and u(t) ≡ 
0, it holds that  Ēdx̄(t) = [Āi x̄(t) + B̄i u(t) + Āhi x̄(t − h)
Z ∞ + N̄i f¯(t)]dt + [B̄wi x(t) + B̄hi x(t − h)]dw (10)
||x(t)||2 |x0 , r0 } < ∞. ys (t) = C̄i x̄(t).

E{ (3)
0
Since plant (10) is singular, the matrices C̄i and Ē satisfy
Definition 2.[30] Let C2 (Rn × S; R+ ) represent the family Ē
of nonnegative functions V (x(t), i) on Rn ×S which are twice the following property: rank = n + 2p = n̄.
C̄i
differentiable continuously in x(t). For V ∈ C2 (Rn × S; R+ ), We define two matrices as follows:
define an infinitesimal operator LV (x(t), i) in the following T
L̄Di = 0p×n 0p×p WiT , S̄i , (Ē + L̄Di C̄i ), (11)

form:
1
lim∆→0+ ∆ [E{V (x(t + ∆), rt+∆ )|x(t), rt = i} − V (x(t), i)]. where W = diag{l1 , l2 , . . . , lp } and li > 0 for i = 1, 2, . . . , p.
In this paper, an observer-based controller is utilized for By the proof of Lemma 1 in [8], we obtain S̄i and S̄i−1 .
FTC. To achieve effective FTC for time-delay MJS with sensor Now, the descriptor SMO is developed for plant (10)
faults and output disturbances, the design of a SMO and 
ˆ(t) + B̄i u(t) + L̄pi ys (t)
S̄i dz̄(t) = [(Āi − L̄pi C̄i )x̄
an observer-based controller is necessary. Figure 1 presents 

 ˆ(t − h)
+ (Āhi − L̄hi C̄i )x̄
an illustrative architecture of the whole FTC system to be (12)
+ L̄ y
hi s (t − h) + L̄si us (t)]dt
designed. As shown in Figure 1, there are mainly three blocks 
 −1
 ˆ = z̄(t) + S̄i L̄Di ys (t)
x̄
represented in different colors according to their functionali-
ties, namely, the system model block with blue background, where z̄(t) , [zxT (t), zsT (t), zdT (t)]T denotes the interme-
the sliding mode observer block with orange background, and ˆ , [x̂T (t), fˆsT (t), dˆT (t)]T denotes the
diate variable and x̄
the controller block with green background.

1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2812754, IEEE
Transactions on Industrial Informatics

estimation of x̄(t); L̄Di ∈ Rn̄×p , L̄pi , L̄hi ∈ Rn̄×p and overall closed-loop plant is formulated in the following form
L̄si ∈ Rn̄×(a+s+d) represent the observer gains which will be 
dx(t) = [Ai x(t) + Bi u(t) + Ahi x(t − h)]dt
designed in subsection III-C. Based upon the above discussion,


+ [Bwi x(t) + Bhi x(t − h)]dw(t)


S̄i , (Ē + L̄Di C̄i ) is non-singular by choosing an appropriate

dē(t) = S̄i−1 [(Āi − L̄pi C̄i )ē(t)


L̄Di . (19)
+ (Āhi − L̄hi C̄i )ē(t − h)
It is noticed that Ddi and Fsi denote full column rank

+ L̄si us (t) − N̄i f¯(t)]dt



matrices. Thus (Ddi T
Ddi )−1 and (Fsi T
Fsi )−1 exist. Then,

− S̄i−1 [B̄wi x̄(t) + B̄hi x̄(t − h)]dw(t)


we know that the real estimation of d(t) and fs (t) are
T ˆ T ˆ
(DdiT
Ddi )−1 Ddi d(t) and (FsiT
Fsi )−1 Fsi fs (t). C. Stability analysis of the overall closed-loop system
For further analysis, the following lemmas will be adopted.
The input manipulation signal is
Lemma 1.[31] Given a pair (Ã, C̃) with à ∈ Rn×n , C̃ ∈
p×n ˆ(t),
R , the following statements are equivalent:(1) If (Ã, C̃) is u(t) = Ki x̂(t) = K̄i x̄ (20)
observable, then ÃT P̃ + P̃ Ã = −C̃ T C̃ has a unique solution.
(2) Ã is stable. with K̄i = [Ki , 0m×p , 0m×d ]. The gain matrix Ki will be
Lemma 2.[8] For each i ∈ S, under (A1), a gain matrix designed such that Ai + Bi Ki is Hurwitz. By employing (20)
L̄pi exists, such that the matrix S̄i−1 (Āi − L̄pi C̄i ) is Hurwitz. to the original plant, we have
dx(t) = [(Ai + Bi Ki )x(t) + Bi K̄i ē(t) + Ahi x(t − h)]dt
B. Derivation of the error dynamics
+[Bwi x(t) + Bhi x(t − h)]dw (21)
In this subsection, the error plant will be constructed. Firstly,
from (12), the following equation can be reached Then we hold the following closed-loop plant as follows

ˆ(t) = [(Āi − L̄pi C̄i )x̄
S̄i dx̄ ˆ(t) + (Āhi − L̄hi C̄i )x̄
ˆ(t − h) 
 dx(t) = [(Ai + Bi Ki )x(t) + Bi K̄i ē(t)
+ Ahi x(t − h)]dt


+B̄i u(t) + L̄pi ys (t) + L̄hi ys (t − h)


+ [Bwi x(t) + Bhi x(t − h)]dw(t)



+L̄si us (t)]dt + L̄Di dys (t). (13) dē(t) = S̄i−1 [(Āi − L̄pi C̄i )ē(t) (22)
+ (Āhi − L̄hi C̄i )ē(t − h)



Then, by adding L̄Di C̄i dx̄(t) to both sides of (10), we have 


 + L̄si us (t) − N̄i f¯(t)]dt
− S̄i−1 [B̄wi x̄(t) + B̄hi x̄(t − h)]dw(t)


S̄i dx̄(t) = [Āi x̄(t) + B̄i u(t) + Āhi x̄(t − h) + N̄i f¯(t)]dt
+[B̄wi x(t) + B̄hi x(t − h)]dw where Fi = [Bi Ki , 0n×s , 0n×d ].
˙
+L̄Di C̄i x̄(t). (14) Now, a theorem is proposed to establish the stability con-
dition of the plant (22).
Furthermore, by defining Theorem 1. Applying us (t) (18) to the error plant (16),
the plant (22) is stochastic stable, if there exist matrices Ki ∈
ˆ(t) − x̄(t) = [eTx (t), eTs (t), eTd (t)]T ,
ē(t) , x̄ (15) Rm×n and positive definite matrices P̄i , Q2 ∈ Rn̄×n̄ , R̄i , Q1 ∈
Rn×n , Ȳi , Ȳhi ∈ Rn̄×p and Hi ∈ R(s+d)×p for each i ∈ S,
and subtracting (14) from (13), the following error plant can such that the following (23) and (24) hold
be obtained
N̄iT S̄i−T P̄i = Hi C̄i , (23)
˙
S̄i ē(t) = [(Āi − L̄pi C̄i )ē(t) + (Āhi − L̄hi C̄i )ē(t − h)
+L̄si us (t) − N̄i f¯(t)]dt − [B̄wi x(t)
 
Γ11i Γ12i Γ13i 0
+B̄hi x(t − h)]dw, (16)  ∗ Γ22i 0 0 
Ψi = 
 ∗
 < 0, (24)
∗ Γ33i Γ34i 
where eTx (t) , x̂(t) − x(t). ∗ ∗ ∗ −Q2
In order to design us (t), the sliding mode surface is firstly
where
defined as
s(t, i) = N̄iT S̄i−T P̄i ē(t) (17) Γ11i = R̄i (Ai + Bi Ki ) + (Ai + Bi Ki )T R̄i +
N
for each i ∈ S with P̄i satisfying N̄iT S̄i−T P̄i = Hi C̄i , P̄i > 0,
X
T
πij R̄j + Bwi Ri Bwi + T̄iT S̄i−T P̄i S̄i−1 T̄i
where Hi ∈ Rp will be determined later. j=1
Furthermore, us (t) is constructed as follows Γ12i T
= R̄i Ahi + Bwi R̄i Bhi + T̄iT S̄i−T P̄i S̄i−1 T̄hi
T −T
us (t) = −(rs1 + rs2 + rd + )sgn(s(t, i)) Γ22i = −Q1 + Bhi T
R̄i Bhi + T̄hi S̄ P̄i S̄i−1 T̄hi
 i
N
X Γ13i = Ri Bi Ki 0n×s 0n×d
− 0.5 πij (N̄iT S̄i−T P̄j S̄i N̄i )−1 s(t, i)sgn(δi ) (18) Γ33i = P̄i S̄i−1 Āi − Ȳi C̄i + ĀTi S̄i−T P̄i − C̄iT ȲiT
j=1
X N

where rPs1 , rs2 and rd have been defined in (A3). δi = + πij P̄j + Q2
N
s(t, i)T j=1 πij (N̄iT S̄i−T P̄j S̄i N̄i )−1 s(t, i) and  > 0 de- j=1

notes a parameter which will be designed later. Then, the Γ34i = P̄i S̄i−1 Āi − Ȳhi C̄i

1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2812754, IEEE
Transactions on Industrial Informatics

 T 
T̄i = B̄wi = Bwi 0n×p 0n×p we can solve this problem by finding the minimum βi in the
constraint below
 T 
T̄hi = B̄hi = Bhi 0n×p 0n×p .
−βi In̄ (N̄iT S̄i−T P̄i − Hi C̄i )T
 
Furthermore, the gains L̄pi and L̄si are obtained as follows < 0. (33)
∗ −Is+d
L̄pi = S̄i P̄i−1 Ȳi (25)
The details of dealing with the linear equality condition can
L̄hi = S̄i P̄i−1 Ȳhi (26) be found in [8].
L̄si = S̄i P̄i−1 C̄i−T HiT = N̄i (27)
D. Reachability of the sliding mode surface
Proof. Consider the error dynamics (22). We choose Lya- The validity of the SMO under the aforementioned con-
punov function as follows figurations is analyzed in this part. A theorem given below
V (t) = Vx (t) + Ve (t), (28) proves the reachability and convergence of the sliding mode
surface in finite time. It should be noted that the Brownian
h
with Vx (t) = xT (t)R̄i x(t) + t−h xT (t)Q1 x(t), and Ve (t) =
R
motion term is eliminated from s(t, i) by means of a matrix
h
ēT (t)P̄i ē(t) + t−h ēT (t)Q2 ē(t). By Definition 2, the follow- parameters design.
R

ing infinitesimal operators along (22) hold Theorem 2. The sliding mode input us (t) guarantees the
sliding motion to be driven to the sliding mode surfaces
LVx (t) = xT (t)[R̄i (Ai + Bi Ki ) + (Ai + Bi Ki )T R̄i s(t, i) = 0 in finite time, if there exist parameter matrix Hi
+Q1 ]x(t) + 2xT (t)R̄i Ahi x(t − h) and matrices P̄i , R̄i which are positive definite for each i ∈ S
+2xT (t)R̄i B̄i K̄i ē(t) + [Bwi x(t) such that formulas (23) and (24) hold.
Proof. Inspired by [32], the solution ē(t) of (16) is given
+Bhi x(t − h)]T R̄i [Bwi x(t) + Bhi x(t − h)]
N
as
X Z t
+xT (t)( πij R̄j )x(t)
ē(t) = S̄i−1 [(Āi − L̄pi C̄i )ē(τ ) + (Āhi − L̄hi C̄i )
j=1
0
−xT (t − h)Q1 x(t − h), (29) × ē(τ − h) + L̄si us (τ ) − N̄i f¯(τ )]dτ
Z t

LVe (t) = ēT (t)[P̄i S̄i−1 (Āi − L̄pi C̄i ) + (Āi − S̄i−1 [B̄wi x̄(τ ) + B̄hi x̄(τ − h)]dw(τ ). (34)
0
N
Substituting (34) into sliding mode surface function s(t, i), we
X
−L̄pi C̄i )T S̄i−T P̄i + πij P̄j + Q2 ]ē(t)
j=1
obtain
Z t
+2ē T
(t)P̄i S̄i−1 [(Āi
− L̄hi C̄i )ē(t − h)
s(t, i) = N̄iT S̄i−T P̄i S̄i−1 [(Āi − L̄pi C̄i )ē(τ )
+L̄si us (t) − N̄i f¯(t)] + [B̄wi x̄(t) + B̄hi x̄(t − h)]T 0

×S̄i−T P̄i S̄i−1 [B̄wi x̄(t) + B̄hi x̄(t − h)] +(Āhi − L̄hi C̄i )ē(τ − h) + L̄si us (τ )
Z t
−ēT (t − h)Q2 ē(t − h). −N̄i f¯(τ )]dτ − Hi C̄i S̄i−1 B̄wi x̄(τ )dw(τ )
0
Recall that L̄si = (P̄i S̄i−1 )−1 C̄iT HiT and N̄iT S̄i−T P̄i = Z t
Hi C̄i , one gets −Hi C̄i S̄i−1 B̄hi x̄(τ − h)dw(τ ). (35)
0
2ēT (t)P̄i S̄i−1 [L̄si us (t) − N̄i f¯(t)] It is observed that
= 2ēT (t)C̄iT HiT − 2ēT (t)P̄i S̄i−1 N̄i f¯(t)
C̄i S̄i−1 B̄wi = 0, C̄i S̄i−1 B̄hi = 0. (36)
≤ −2sT (t, i)us (t) + 2ēT (t)||C̄iT HiT ||||f¯(t)||
≤ −2||s(t, i)||. (30) Taking (36) into consideration, (35) can be rewritten as
Z t
From (25), we have T −T
s(t, i) = N̄i S̄i P̄i S̄i −1
[(Āi − L̄pi C̄i )ē(τ )
0
Ȳi = P̄i S̄i−1 L̄pi , Ȳhi = P̄i S̄i−1 L̄hi , L̄si = N̄i . (31)
+(Āhi − L̄hi C̄i )ē(τ − h) + L̄si us (τ )
Substituting (30) and (31) into (28) yields −N̄i f¯(τ )]dτ. (37)
LV (t) = LVx (t) + LVe (t) ≤ z(t)T Ψi z(t), (32) Then, we have
T
where z(t) = xT (t) xT (t − h) ēT (t) ēT (t − h) . We

ṡ(t, i) = N̄iT S̄i−T P̄i S̄i−1 [(Āi − L̄pi C̄i )ē(t)
have noticed that Ψi < 0 from (24). Then, LV (t) < 0 can +(Āhi − L̄hi C̄i )ē(t − h) + L̄si us (t)
be derived. Therefore, we conclude that the plant (22) is
stochastic stable. The proof is completed.  −N̄i f¯(t)]. (38)

Construct the Lyapunov function Vs (t) in the following form

In addition, it is worth to point out that the condition in (23)
can be converted into a minimization problem. That means Vs (t) = 0.5sT (t, i)(N̄iT S̄i−T P̄i S̄i−1 N̄i )−1 s(t, i). (39)

1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2812754, IEEE
Transactions on Industrial Informatics

For notational simplicity, define Gi = N̄iT S̄i−T and consider +N̄i us (t) − N̄i f¯(t)]dt
(38), we hold −S̄i−1 [B̄wi x̄(t) + B̄hi x̄(t − h)]dw(t). (47)
LVs (t, i) = Define a new matrix B̄si = S̄i−1 N̄i ∈ R(n+2p)×(s+d) , where
sT (t, i)(Gi P̄i GTi )−1 Gi P̄i S̄i−1 × [(Āi − L̄pi C̄i )ē(t) N̄i is full column rank. B̂si is an orthogonal matrix of B̄si ,
T T −1
+(Āhi − L̄hi C̄i )ē(t − h) + L̄si us (t) − N̄i f¯(t) that is, B̂si B̄si = B̂si S̄i N̄ i = 0. We define a coordinate
N B̂si
X transformation matrix Ti = ∈ R(n+2p) .
+0.5sT (t, i) × πij (N̄iT S̄i−T P̄i S̄i−1 N̄i )−1 s(t, i)]. (40) T
B̄si P̄i
j=1 Then, it can be easily verified that

Recall that L̄si , N̄i S̄i−T Pi = Hi C̄i , and it can be derived that
   
−1 B̂si −1 0
Ti S̄i N̄i = S̄ N̄i = T . (48)
T
B̄si P̄i i B̄si P̄i B̄si
sT (t, i)(Gi P̄i GTi )−1 Gi P̄i S̄i−1 [L̄si us (t) − N̄i f¯(t)]
= sT (t, i)(Gi P̄i GTi )−1 Gi P̄i GTi us (t) Consider the transformed state z̃(t) = Ti ē(t), the regular
form of the error plant can be formulated as follows:
−sT (t, i)(Gi P̄i GTi )−1 Gi P̄i GTi f¯(t)
 
< −||s(t, i)|| − 0.5sT (t, i) dz̃(t) = Âi z̃(t) + Âhi z̃(t − h) +
0
(us (t) − f¯(t))
T
X N B̄si P̄i B̄si
× πij (N̄iT S̄i−T P̄i S̄i−1 N̄ )−1
i s(t, i). (41) −S̄i−1 [B̄wi x̄(t) + B̄hi x̄(t − h)]dw(t), (49)
j=1
where Âi = S̄i−1 (Āi − L̄pi C̄i ), Âhi = S̄i−1 (Āhi − L̄hi C̄i ).
Substituting (41) into (40), one has
By multiplying N̄iT S̄i−1 P̄i on both sides of (49), we have
LVs (t, i) < −||s(t, i)|| + ||s(t, i)||||(Gi P̄i GTi )−1 Gi P̄i S̄i−1  
0
×(Āi − L̄pi C̄i )||||ē(t)|| + ||s(t, i)||||(Gi P̄i GTi )−1 dz̃(t) = Ãi z̃(t) + Ãhi z̃(t − h) + N̄iT S̄i−1 P̄i T
B̄si P̄i B̄si
×Gi P̄i S̄i−1 (Āi − L̄pi C̄i )||||ē(t − h)|| ¯ T −1 −1
(us (t) − f (t)) − N̄ S̄ P̄i S̄ [B̄wi x̄(t)
i i i
We define +B̄hi x̄(t − h)]dw(t), (50)
δi = ||(Gi P̄i GTi )−1 Gi P̄i S̄i−1 (Āi − L̄pi C̄i )||. (42) where Ãi = N̄iT S̄i−1 P̄i S̄i−1 (Āi − L̄pi C̄i ), Ãhi =
N̄iT S̄i−1 P̄i S̄i−1 (Āhi − L̄hi C̄i ). Since N̄iT S̄i−1 P̄i = Hi C̄i ,
ρi = ||(Gi P̄i GTi )−1 Gi P̄i S̄i−1 (Āhi − L̄hi C̄i )||. (43) C̄i S̄i−1 B̄wi = 0 and C̄i S̄i−1 B̄hi = 0, the regular form of the
error plant is in the following form:
From (42), we have  
0
LVs (t, i) < −||s(t, i)||( − δi ||ē(t)|| − ρi ||ē(t − h)||). (44) dz̃(t) = Ãi z̃(t) + Ãhi z̃(t − h) + N̄iT S̄i−1 P̄i T
B̄si P̄i B̄si
Then, we define the following domain for each i ∈ S, ¯
×(us (t) − f (t)). (51)
s  
\ Ã11i Ã12i
Ω, Ωi (δi , ρi ), (45) In system (51), we decompose Ãi = , Ãhi =
i=1   Ã21i Ã22i
Ãh11i Ãh12i
, and system (51) is decomposed into the
Ωi (δi , ρi ) , { − δi ||ē(t)|| − ρi ||ē(t − h)|| > 0}. (46) Ãh21i Ãh22i
following version:
From (44) and (45),it is noticed that LVs (t, i) < 0 is in 
the domain Ω. Furthermore, we have proved the stochastic 
 z̃1 (t) = Ã11i z1 (t) + Ã12i z2 (t)
+Ãh11i z1 (t − h) + Ãh12i z2 (t − h)

stabilization of (22) in Theorem 1. This implies that the


trajectories of ē(t) will enter Ω in finite time and remains z̃2 (t) = Ã21i z1 (t) + Ã22i z2 (t) (52)
− −

there. Until now, we complete the proof. 


 + Ã z
h21i 1 (t h) + Ã z
h22i 2 (t h)
+N̄iT S̄i−1 P̄i B̄si
T
P̄i B̄si (us (t) − f¯(t)).

Remark 1. In classical sliding mode control scheme, the
sliding mode stability is first proved to help the system
stability. Since the stabilization of plant (19) has been proved
T
It’s obvious that z̃2 (t) = B̄si P̄i ē(t) = N̄iT S̄i−T P̄i ē(t) =
in Theorem 1, someone will wonder that whether the Theorem s(t, i), that means
2 is necessary. Actually, the main reason that we employ ṡ(t) = z̃˙2 (t) = Ã21i z̃1 (t) + Ã22i s(t, i) + Ãh21i z̃1 (t − h)
Theorem 2 in our work is to obtain the estimation of the fault
vector f¯(t). In the following part, we will give the details of +Ãh22i s(t − h, i) + N̄iT S̄i−1 P̄i B̄si T
P̄i B̄si (us (t)
estimating the fault vector f¯(t). −f¯(t)). (53)
Recall the error system as shown in (22):
Since on the sliding surface, we have s(t) = 0, ṡ(t) = 0,
Since L̄si = N̄i , we have
and z̃1 (t) → 0. Then, we conclude that us (t) ≈ f¯(t).
dē(t) = S̄i−1 [(Āi − L̄pi C̄i )ē(t) This is the main reason that we employ Theorem 2 in our
+(Āhi − L̄hi C̄i )ē(t − h) work, which is is employed to obtain the estimation of the
fault vector f¯(t).

1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2812754, IEEE
Transactions on Industrial Informatics

 T
E. The complete system synthesis algorithm Dd1 = Dd2 = 0.1 0.2
 
A complete controller design and the FTC system synthesis −0.6 0.6
algorithm are provided below. Step 1, Steps {2, 4, 5} and Π=
0.8 −0.8
Step 3 focus on the system model (blue) block, the SMO
(orange) block, and the controller (green) block in Figure 1,
respectively.

Algorithm 1. SMO based FTC system synthesis

Step 1. Establish the augmented system in the form of (10).
Calculate the coefficient matrices of the augmented system,
including matrices Ē, Āi , B̄i , Āhi , N̄i , B̄wi , B̄hi and C̄i ,
according to equations (4) to (9), respectively.
Step 2. Refer to equation (11), select the observer gain matrix
L̄Di such that S̄i is not singular.
Step 3. Controller design.
Solve for the closed-loop control gain matrix K̄i via an
equivalent linear matrix inequality problem given in Theo-
Fig. 2. Trajectories of s(t) when us (t) = −1.0010sgn(s(t))
rem 1, with the aid of Matlab LMI Toolbox.
Step 4. Construct the SMO in the form of (12).
(i) Solve for observer related coefficient matrices by solving
the equivalent linear matrix inequality problem, which is
described by equations (24) and (33). To be more specific,
the solutions are the matrices P̄i , Ȳi , Ȳhi and Hi .
(ii) Calculate the sliding mode observer coefficient matrices
L̄pi , L̄hi and L̄si according to equations (25), (26) and (27),
respectively.
(iii) Calculate the sliding mode surface function s(t) and the
sliding mode input us (t) according to equations (17) and
(18).
Step 5. Inspect the residual signals ē(t).
If ē(t) converges to zeros, the FTC system design is Fig. 3. Trajectories of s(t) when us (t) = −2.0000sgn(s(t))
completed. Otherwise, go to Step 4 and tune the adjustable
parameter in equation (18) to suppress the sliding mode
chattering.

IV. S IMULATION STUDY

In this section, an example is provided to testify the validity
of the above derivations and to demonstrate the algorithm with
detailed implementation steps and intermediate results. After
that, the performance of the FTC system is further illustrated
and discussed.
A. Simulation setup
Recall the time-delay MJS with sensor faults and output
disturbances (1). The model parameters are given as follows
    Fig. 4. Sensor fault estimation performance
−1 −0.1 −5 0
A1 = , A2 =
0 −2 −0.4 −4 B. Results illustration and discussion
   
B1 = −1 0 , B2 = −1 0 By selecting L̄Di
   
0.4 0 3 0 T
Ah1 = , Ah2 =

0 0.1 0 0 0 0 1 0
0 2 L̄D1 = L̄D2 = (54)
0 0 0 0 0 1
  
−0.1 −0.1 −0.5 0
Bw1 = , Bw2 =
0  −0.2  −0.4
 −0.4 the corresponding S1 and S2 are non-singular matrices, which
0.4 0 3 0
Bh1 = , Bh2 = satisfie the necessary condition in Algorithm 1 - Step 2.
0 0.1  0 2 

0.1 −0.2

0.5 0.2 In Step 3 and Step 4 (i), the problem is solved with the
C1 = , C2 = Matlab LMI Toolbox, and the controller gain Ki are
0.1 0.5 0.1 −0.1
 T    
Fs1 = Fs2 = 0.1 0 K1 = −0.5000 0.1000 , K2 = 3.5153 0.5824 (55)

1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2812754, IEEE
Transactions on Industrial Informatics

those of (A2 + B2 K2 ) are −8.5663 and −3.9490, all of which

are in the left-hand side of the complex plane.
By following Step 4 (ii) and Step 4 (iii), we obtain
L̄p1 ,L̄p2 ,L̄h1 ,L̄h2 ,L̄s1 ,and L̄s1 . Then, the sliding surface can
be constructed.
Until now, the design phase of the FTC system is completed,
and the simulation can be carried out. The simulation time is
selected as t ∈ [0, T ], where T = 30, with a time step of
10−3 . The time-delay is h = 0.5. The number of discretized
Brownian paths is 20. The initial condition is chosen as x(0) =
[1, −1]T , x̂(0) = [0, 0]T , fs (0) = 0.201, fˆs (0) = 0, d(0) =
ˆ = 0. The sensor fault signals follow
0.101, d(0)
Fig. 5. Output disturbance estimation performance

x 10
−4 The residual signals fs (t) = 0.1 sin(10t) + 0.2 cos(10t) (56)
20
ē1 (t)
ē2 (t) Figures 2 and 3 show the trajectories of s(t). Both tra-
15
jectories achieve satisfying convergence and present classical
10 chattering phenomenon in the SMC. It is interesting to see
that by tuning the coefficient of us (from −1.0010 to −2), the
5 chattering can be suppressed. The amplitudes of chattering are
2 × 10−4 and 1 × 10−4 , respectively.
0
The true and the estimated sensor fault and output distur-
−5
bance are shown in Figures 4 and 5. With tiny residuals of
0 5 10 15 20 25 30 fs and d, these two plots demonstrate the effectiveness of the
Time (Sec.)
sliding mode observer.
Fig. 6. Residual signals of the state variables Finally, the trajectories of the state variables are shown in
Figure 7 while the estimation error is shown in Figure 6.
Trajectories of the state variables All the state variables converge to the stable point rapidly
1.5
x1 (t) in an exponential manner. Although the residual signals still
x2 (t)
1 suffer from the chattering phenomenon, the overall closed-
loop system stability is not affected. Furthermore, similar
0.5 simulations are carried out for plenty of times, and in most
times the linear matrix inequality solutions are available.
0
Figure 8 shows the FTC results along 20 individual paths
−0.5
and the average trajectory, which verifies the applicability and
stochastic stability of the proposed approach.
−1
0 5 10 15 20 25 30
Time (Sec.) V. C ONCLUSIONS

Fig. 7. Trajectories of the state variables FTC of time-delay MJS with Itô stochastic and output
disturbances has been investigated in this paper. As a passive
FTC scheme, the augmented system is firstly derived where
sensor faults and output disturbances are treated as state
variables. After that, the error system and the closed-loop
system are considered simultaneously. Theorem 1 provides a
sufficient condition to stabilize the closed-loop system, which
could be used for SMO and controller synthesis. The designed
observer inherits the advantages of both the PD observer
and the sliding mode observer. Specifically, the sliding mode
input aims to eliminate the sensor fault while the PD terms
provide more design freedom to improve the performance.
Theorem 2 proves the reachability and convergence of the
sliding mode surface by removing the Brownian motion term.
In the meantime, the fault vector f¯(t) has been reconstructed.
Fig. 8. Stochastic stability of the algorithm Finally, an applicable algorithm is summarized and tested over
a numerical example. The simulation results show the validity
of the newly derived theorems and the effectiveness of the
and the eigenvalues of (A1 + B1 K1 ) are −1 and −0.5, and proposed algorithm.

1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2812754, IEEE
Transactions on Industrial Informatics

R EFERENCES [22] S. P. Bhat, “Controllability of nonlinear time-varying systems: appli-

cation to spacecraft attitude control using magnetic actuation,” IEEE
[1] Z. Gao, X. Liu, and M. Z. Q. Chen, “Unknown input observer-based Transaction on Automatic Control, vol. 50, no. 11, pp. 1725–1735, 2005.
robust fault estimation for systems corrupted by partially decoupled [23] M. A. Khanesar, O. Kaynak, S. Yin, and H. Gao, “Adaptive indirect
disturbances,” IEEE Transactions on Industrial Electronics, vol. 63, fuzzy sliding mode controller for networked control systems subject to
no. 4, pp. 2537–2547, April 2016. time-varying network-induced time delay,” IEEE Transactions on Fuzzy
[2] Z. Gao, C. Cecati, and S. X. Ding, “A survey of fault diagnosis and Systems, vol. 23, no. 1, pp. 205–214, 2015.
fault-tolerant techniques-part i: Fault diagnosis with model-based and [24] X. Yu and Z. Man, “Multi-input uncertain linear systems with terminal
signal-based approaches,” IEEE Transactions on Industrial Electronics, sliding-mode control,” Automatica, vol. 34, no. 3, pp. 389–392, 1998.
vol. 62, no. 6, pp. 3757–3767, June 2015. [25] L. Wu, P. Shi, and H. Gao, “State estimation and sliding-mode control
[3] Z. Ge, S. Zhong, and Y. Zhang, “Semisupervised kernel learning for of markovian jump singular systems,” IEEE Transactions on Automatic
F DA model and its application for fault classification in industrial Control, vol. 55, no. 5, pp. 1213–1219, May 2010.
processes,” IEEE Transactions on Industrial Informatics, vol. 12, no. 4, [26] J. Davila, L. Fridman, and A. Levant, “Second-order sliding-mode
pp. 1403–1411, Aug 2016. observer for mechanical systems,” IEEE Transactions on Automatic
[4] S. Yin, H. Luo, and S. X. Ding, “Real-time implementation of fault- Control, vol. 50, no. 11, pp. 1785–1789, 2005.
tolerant control systems with performance optimization,” IEEE Trans- [27] J. A. Moreno and M. Osorio, “A lyapunov approach to second-order
actions on Industrial Electronics, vol. 61, no. 5, pp. 2402–2411, 2014. sliding mode controllers and observers,” IEEE Conference on Decision
[5] P. S. Flannery and G. Venkataraman, “A fault tolerant doubly fed and Control, pp. 2856–2861, 2008.
induction generator wind turbine using a parallel grid side rectifier and [28] P. Li, X. Yu, and B. Xiao, “Adaptive quasi-optimal higher order
series grid side converter,” IEEE Transactions on Power Electronics, sliding mode control without gain overestimation,” IEEE Transactions
vol. 23, no. 3, pp. 1126–1135, 2008. on Industrial Informatics, vol. PP, no. 99, pp. 1–1, 2017.
[6] S. Sundaram and C. N. Hadjicostis, “Distributed function calculation [29] K. Liao and Y. Xu, “A robust load frequency control scheme for power
via linear iterative strategies in the presence of malicious agents,” IEEE systems based on second-order sliding mode and extended disturbance
Transactions on Automatic Control, vol. 56, no. 7, pp. 1495–1508, 2011. observer,” IEEE Transactions on Industrial Informatics, vol. PP, no. 99,
[7] M. Liu, P. Shi, L. Zhang, and X. Zhao, “Fault-tolerant control for pp. 1–1, 2017.
nonlinear markovian jump systems via proportional and derivarive slding [30] A. V. Skorokhod, Asymptotic Methods in the Theory of Stochastic
mode observer technique,” IEEE Transactions on Circuits and Systems, Differential Equations. American Mathematical Society, 1989.
vol. 58, no. 11, pp. 2755–2764, 2011. [31] C. Chen, Linear system theory and design. New York: Oxford, 1999.
[8] M. Liu and P. Shi, “Sensor fault estimation and tolerant control for ito [32] V. Kolmanovskii and A. Myshkis, Applied Theory of Functional Differ-
stochastic systems witha descriptor sliding mode approach,” Automatica, ential Equations. Kluwer: Dordrecht, The Netherlands, 1992.
vol. 49, pp. 1242–1250, 2013.
[9] Y. Niu, D. W. C. Ho, and J. Lam, “Robust integral sliding mode control
for uncertain stochastic systems with time-varying delay,” Automatica,
vol. 41, pp. 873–880, 2005.
[10] Y. Niu and D. W. C. Ho, “Robust observer design for ito stochastic time-
Hongyan Yang (S’16) is currently working towards
delay systems via sliding mode control,” Systems and Control Letters,
the Ph.D. degree at Intelligent Control and Systems
vol. 55, pp. 781–793, 2006.
Institute, Harbin Institute of Technology, Harbin,
[11] H. N. Wu and M. Z. Bai, “Active fault-tolerant fuzzy control design
China. Her current research interests include adap-
of nonlinear model tracking with application to chaotic systems,” IET
tive fuzzy control, adaptive neural control, sliding
Control Theorey and Applications, vol. 3, no. 6, pp. 642–653, 2009.
mode control and nonlinear systems.
[12] B. Jiang, Z. Gao, P. Shi, and Y. Xu, “Adaptive fault-tolerant tracking
control of near-space vehicle using takagi-sugeno fuzzy models,” IEEE
Transactions on Fuzzy Systems, vol. 18, no. 5, pp. 1000–1007, 2010.
[13] S. Yin, H. Yang, H. Gao, J. Qiu, and O. Kaynak, “An adaptive nn-
based approach for fault-tolerant control of nonlinear time-varying delay
systems with unmodeled dynamics,” IEEE Transactions on Neural
Networks and Learning Systems, vol. 99, no. 1, pp. 1–12, 2016.
[14] Z. Gao, C. Cecati, and S. X. Ding, “A survey of fault diagnosis and
fault-tolerant techniques–part i: Fault diagnosis with model-based and
signal-based approaches,” IEEE Transactions on Industrial Electronics, Yuchen Jiang (S’16) received his B.E. degree with
vol. 62, no. 6, pp. 3757–3767, 2015. first-class honors in automation from Harbin Insti-
[15] L. D. Lago, O. Ferrante, R. Passerone, and A. Ferrari, “Dependability tute of Technology, Harbin, China, in 2016. He is
assessment of soa-based cps with contracts and model-based fault currently a Ph.D. student at Intelligent Control and
injection,” IEEE Transactions on Industrial Informatics, vol. 14, no. 1, Systems Institute, Harbin Institute of Technology,
pp. 360–369, 2018. China. He was in an exchange program in UNSW,
[16] M. Basin and J. Rodriguez-Gonzalez, “A closed-form optimal control for Australia during the period 2014/07 to 2014/12. His
linear systems with equal state and input delays,” Automatica, vol. 41, research interests include data-driven fault diagnosis,
no. 5, pp. 915–920, 2005. process monitoring and artificial intelligence. He is
[17] M. Basin, P. Shi, D. Calderon-Alvarez, and J. Wang, “Central subop- also working on online learning based optimization
timal H∞ filter design for linear time-varying systems with state or and controller design.
measurement delay,” Circuits Systems Signal Processing, vol. 28, no. 2,
pp. 305–330, 2009.
[18] H. Dong, Z. Wang, S. X. Ding, and H. Gao, “On H∞ estimation of
randomly occurring faults for a class of nonlinear time-varying systems
with fading channels,” IEEE Transactions on Automatic Control, vol. 61, Shen Yin (M’12-SM’15) received the B.E. degree
no. 2, pp. 479–484, Feb 2016. in automation from Harbin Institute of Technology,
[19] B. Shen, Z. Wang, and X. Liu, “Bounded H∞ synchronization and state Harbin, China, in 2004, the M.Sc. degree in control
estimation for discrete time-varying stochastic complex networks over a and information system an d the Ph.D. degree in
finite horizon,” IEEE Transactions on Neural Networks, vol. 22, no. 1, electrical engineering and information technology
pp. 145–157, Jan 2011. from University of Duisburg-Essen, Duisburg, Ger-
[20] M. Mohiuddin, W. Saab, S. Bliudze, and J. Y. L. Boudec, “Axo: many. He is currently a Professor with the Research
Detection and recovery for delay and crash faults in real-time control Center of Intelligent Control and Systems, Harbin
systems,” IEEE Transactions on Industrial Informatics, vol. PP, no. 99, Institute of Technology. His research interests are
pp. 1–1, 2017. model-based and data-driven fault diagnosis and
[21] H. Rotstein and S. Raz, “Gabor transform of time-varying systems: exact prognosis in process control and their applications,
representation and approximation,” IEEE Transactions on Automatic fault-tolerant control, and big data focused on industrial electronics applica-
Control, vol. 44, no. 4, pp. 729–741, 1999. tions.

1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.