652 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 20, NO.

4, AUGUST 2012

Fault-Tolerant Control for T–S Fuzzy Systems

With Application to Near-Space Hypersonic
Vehicle With Actuator Faults
Qikun Shen, Bin Jiang, Senior Member, IEEE, and Vincent Cocquempot

Abstract—This paper addresses the problem of fault-tolerant the past two decades, as can be attested to by the abundant lit-
control for Takagi–Sugeno (T–S) fuzzy systems with actuator faults. erature [3], [4], [12]–[23]. Fault-tolerant control (FTC) aims to
First, a general actuator fault model is proposed, which inte- preserve the functionalities of a faulty system with acceptable
grates time-varying bias faults and time-varying gain faults. Then,
sliding-mode observers (SMOs) are designed to provide a bank performances. FTC can be achieved in two ways: passive and
of residuals for fault detection and isolation. Based on Lyapunov active. The former uses feedback control laws that are robust
stability theory, a novel fault-diagnostic algorithm is proposed to with respect to possible system faults [12], [13]. On the other
estimate the actuator fault, which removes the classical assumption hand, the latter uses an FDI module and accommodation tech-
that the time derivative of the output errors should be known as niques. The problem to design a robust fault-detection system
in some existing work. Further, a novel fault-estimation observer
is designed. Utilizing the estimated actuator fault, an accommoda- for uncertain Takagi–Sugeno (T–S) fuzzy models was studied
tion scheme is proposed to compensate for the effect of the fault. In in [14]. In [15], the problem of fault detection for fuzzy systems
addition, a sufficient condition for the existence of SMOs is derived with intermittent measurements was investigated. In [16], using
according to Lyapunov stability theory. Finally, simulation results the obtained fault information, an additive controller was de-
of a near-space hypersonic vehicle are presented to demonstrate signed to compensate for the fault. Less conservative criteria for
the efficiency of the proposed approach.
fault accommodation of time-varying delay systems were pro-
Index Terms—Fault diagnosis and estimation, near-space hy- posed using adaptive fault-diagnosis observers in [17]. In [18],
personic vehicle (NSHV), sliding-mode observer (SMO), Takagi– the problem of robust fault detection for a class of nonlinear
Sugeno (T–S) fuzzy model.
time-delay systems was investigated. In [19], the robust fault-
detection problem was investigated for a class of discrete-time
I. INTRODUCTION networked systems with unknown input and multiple state de-
lays. In [20], the reliable H∞ filtering problem is considered
ODERN control systems, such as the near-space hyper-
M sonic vehicle (NSHV) that is considered in this paper,
have become increasingly complex and involve an increasing
against actuator failures for a class of uncertain discrete-time
systems with norm-bounded uncertainties and time-varying de-
lays. In [21], robust static output feedback control for linear
number of actuators and sensors. These physical components discrete-time systems with time-varying uncertainties was in-
may become faulty, which can cause deterioration in system per- vestigated. In [22], a general active FTC framework was pro-
formance and lead to instability that can further produce catas- posed for nonlinear systems with actuator faults. However, this
trophic accidents. To improve system reliability and guarantee control framework works under the assumption that system
system stability in all situations, fault detection and isolation states are bounded in all cases. Furthermore, the fault-estimation
(FDI) and fault accommodation methods have become attrac- algorithm was not given. In [23], the problem to adaptively
tive topics which received have considerable attention during compensate actuator uncertainties was addressed in a feedback-
based framework. It is valuable to point out that most results
concerning actuator faults reported in the literature only con-
sidered bias faults. Gain faults did not attract enough attention,
Manuscript received June 26, 2011; revised September 20, 2011; accepted
December 5, 2011. Date of publication December 22, 2011; date of current
which motivates this paper. In addition, in some existing work,
version August 1, 2012. This work was supported in part by the National Natu- estimation error limt→∞ ex (t) = ex (∞) was considered as an
ral Science Foundation of China under Grant 91116018, Grant 61174046, and indicator, by which the faulty system can be distinguished from
Grant 61175111.
Q. Shen is with the College of Automation Engineering, Nanjing University
the normal system. That is to say, if ex (∞) = 0, then the sys-
of Aeronautics and Astronautics, Nanjing, 210016, China, and with the College tem is healthy; if ex (∞) = 0, then the system is faulty. However,
of Information Engineering, Yangzhou University, Yangzhou, 225009, China ex (∞) is not available in practice, and ex (∞) = 0 cannot prac-
(e-mail: qkshen@[Link]).
B. Jiang is with the College of Automation Engineering, Nanjing University
tically be considered as a fault indicator. Another motivation of
of Aeronautics and Astronautics, Nanjing 210016, China (e-mail: binjiang@ this study is, thus, to provide a fault indicator with an associated
[Link]). decision algorithm which is efficient in practical application.
V. Cocquempot is with the LAGIS FRE CNRS, 3303, Lille 1 Univer-
sity: Sciences and Technologies, 59655 Villeneuve d’Ascq, France (e-mail:
The concept of NSHV was first proposed by the United
[Link]@[Link]). States Air Force in a military exercise called “Schrieffer” in
Color versions of one or more of the figures in this paper are available online 2005. NSHV is a class of vehicle flying in near space which
at [Link]
Digital Object Identifier 10.1109/TFUZZ.2011.2181181
offers a promising and new, lower cost technology for future

1063-6706/$31.00 © 2012 IEEE

SHEN et al.: FAULT-TOLERANT CONTROL FOR T–S FUZZY SYSTEMS WITH APPLICATION TO NEAR-SPACE HYPERSONIC VEHICLE 653

spacecraft. It can advance space transportation as well as prompt given, which include fault detection, isolation, estimation, and
global strike capabilities. Such a complex technological system FTC scheme. The NSHV application is presented in Section IV.
attracts considerable interests from the control research commu- The T–S fuzzy model is employed to approximate the nonlin-
nity and aeronautical engineering in the past couple of decades, ear NSHV attitude dynamics, and simulation results of NSHV
and significant results were reported [1], [2], [5]–[8], [29]–[32]. are presented to demonstrate the effectiveness of the proposed
For such a high technological system, with great economical technique. Finally, Section V draws the conclusion.
and societal issues, it is of course essential to maintain high re- Notations: In this paper, R, Rn , and Rn ×m denote, respec-
liability against possible faults [9]–[11]. One of the difficulties tively, the real numbers, the real n-vectors, and the real n × m
to deal with FTC for NSHV is that the dynamics are complex matrices. We define the norm of a vector x ∈ Rn as ||x||2 =

nonlinear, multivariable, and strongly coupled ones. To solve x21 + x2 + · · · + x2 and the norm of a matrix A ∈ Rn ×m as
2 n
the difficulties, a T–S fuzzy system was used to describe the ||A||2 = λm ax (AT A) = σm ax (A), λm in (A) and λm ax (A) are
NSHV attitude dynamics [29]. During the past two decades, the the smallest and largest eigenvalues of a matrix A, and σm ax (A)
stability analysis for T–S fuzzy systems has attracted increas- is the maximum singular value. The absolute value is denoted
ing attention [33]–[41]. These studies combine the flexibility by | · |.
of fuzzy logic theory and rigorous mathematical theory of lin-
ear/nonlinear systems into a unified framework. The important II. TAKAGI–SUGENO FUZZY MODEL WITH ACTUATOR FAULTS
advantage of a T–S fuzzy system is its universal approximation AND FAULT-TOLERANT CONTROL PROBLEM FORMULATION
of any smooth nonlinear function by a “blending” of some local
Consider the following T–S fuzzy model that is composed of
linear models, which greatly facilitates the analysis and synthe-
a set of fuzzy implications, where each implication is expressed
sis of the complex nonlinear system. Many stability criteria of
by a linear state-space model. The ith rule of this T–S fuzzy
T–S fuzzy systems have been expressed in terms of linear ma-
model is of the following form.
trix inequalities (LMIs) via various stability analysis methods
Plant Rule i: IF z1 (t) is Mi1 and · · · zq (t) is Miq , THEN
(see [24]–[28] and [30]–[32] and references therein). In [11],

the authors studied the problem of fault-tolerant tracking con- ẋ(t) = Ai x(t) + Bi u(t)
trol for near-space-vehicle attitude dynamics with bias actuator (1)
y(t) = Ci x(t)
fault, where the bias fault was assumed to be unknown constant.
However, in practical application, the fault may be time varying, where i = 1, . . . , r, r is the number of the IF–THEN rules,
which motivates this paper. Mij , j = 1, . . . , q is the fuzzy set, z(t) = [z1 (t), . . . , zq (t)]T
In this paper, we investigate the problem of FTC for T–S are the premise variables that are supposed to be known,
fuzzy systems with actuator time-varying faults, with the objec- x(t) = [x1 (t), . . . , xn (t)]T ∈ Rn , u(t) ∈ Rm , Ai ∈ Rn ×n , and
tive to provide an efficient solution to control NSHV in faulty Bi ∈ Rn ×m .
situations. Compared with some existing work, there are four The overall fuzzy system is inferred as follows:
main contributions that are worth to be emphasized. ⎧ r
⎪
⎨ ẋ(t) = hi (z(t))(Ai x(t) + Bi u(t))
1) The actuator fault model that is presented in this paper in- i=1
 (2)
tegrates not only time-varying gain faults but time-varying ⎪
⎩ y(t) =
r
hi (z(t))Ci x(t)
bias faults as well, which means that a wide class of faults i=1
can be handled. The theoretic developments and results of
where hi (z(t)) is defined as
this paper are, thus, valuable in a wide field of practical
n
applications. j =1 Mij [z(t)]
2) An adaptive fault-estimation algorithm is proposed, where hi (z(t)) = r n , i = 1, 2, . . . , r (3)
i=1 j =1 Mij [z(t)]
the common assumption that the derivative of the output
errors with respect to time should be known is removed where Mij [z(t)] is the grade of membership of zj (t) in Mij .
(see [9] for instance) and the parameter drift phenomenon It is assumed in this paper that nj=1 Mij [z(t)] ≥ 0 for all t.

is prevented even in the presence of bounded disturbances. Therefore, we have ri=1 hi (z(t)) = 1, 0 ≤ hi (z(t)) ≤ 1, for
3) Compared with some results [9]–[11], [17], a decision all t.
threshold for FDI is defined and applied on an online In this paper, the state feedback control strategy is chosen
computable fault indicator and not on an asymptotic value as a parallel distributed compensation (PDC), which can be
of a criterion which is not easily available to use. The described as follows.
decision algorithm is, thus, more practical than in other Control Rule i: IF z1 (t) is Mi1 and · · · zq (t) is Miq , THEN
work, like [9]–[11], [17].
ui (t) = Ki x(t) (4)
4) Differing from the classical fault estimation in [9]–[11],
[17], the proposed fault-estimation observer is designed to where Ki is the controller gain matrix which is to be determined
online estimate not only bias faults but gain faults as well. later.
The rest of this paper is organized as follows. In Section II, The overall fuzzy controller is given as follows:
the T–S fuzzy model is first briefly recalled. Actuator faults are r
integrated in such a model, and the FTC objective is formu- u(t) = hi (z(t))Ki x(t). (5)
lated. In Section III, the main technical results of this paper are i=1
654 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 20, NO. 4, AUGUST 2012

The control objective under normal conditions is to design a is used to design a proper control input u(t) such that the system
proper state feedback controller u(t) such that the system (2) is (2) is still maintained stable under faulty case.
stable. Remark 1: In the literature, many papers consider actuator
However, in practical application, actuators may become faults. However, most of them only considered bias faults. Gain
faulty. Bias faults and gain faults are two kinds of actuator faults have not attracted enough attention. In [11], a class of bias
faults that are commonly occurring in practice. An actuator bias fault was studied, where the fault was assumed to be an unknown
fault can be described as constant. However, in practical application, the fault may be time
varying. Equation (10) is a deterministic but uncertain actuator
ufi (t) = ui (t) + fi (t), i = 1, . . . , m, (6)
model which represents a class of practical actuator faults such
where fi (t) denotes a bounded signal, and an actuator gain fault as actuator gain variations and measurement errors. In fact, the
can be described as fault model in [9]–[11] can be described by (10). If ρi (t) = 0,
model (10) then becomes the bias fault model that is considered
ufi (t) = (1 − ρi (t))ui (t), i = 1, . . . , m, (7) in [9] and [10]. If ρi (t) is an unknown constant and fi (t) = 0,
where 0 ≤ ρi (t) ≤ 1, which is supposed to be unknown, de- then model (10) denotes the constant bias faults model as in [11].
notes the remaining control rate. Therefore, the two previously Hence, the proposed actuator fault model (10) is more general
mentioned kinds of actuator faults can be uniformly described and has wider practical use than the classical ones.
as
ufi (t) = (1 − ρi (t))ui (t) + fi (t). (8) III. FAULT DIAGNOSIS AND ACCOMMODATION

Furthermore, a more general fault model can be given as In this section, the main technical results of this paper are
given. We will first formulate the fault-diagnosis and accommo-
pi
dation problem of the aforementioned T–S fuzzy system. We
ufi (t) = (1 − ρi (t))ui (t) + gi.j fi,j (t) (9) will then design a bank of sliding-mode observers (SMOs) to
j =1
generate residuals, investigate the FDI algorithm that is based
where fi,j (t), i = 1, . . . , m, j = 1, . . . , pi , denotes a on the SMOs, and propose an FTC scheme to tolerate the fault
bounded signal, pi is a known positive constant, and gi,j using estimated fault information.
denotes an unknown constant. With no restriction, let us sup-
pose p1 = p2 = · · · = pm = p, with p being a known positive
A. Preliminary Description
constant. Consider the following notation: ai,j (t) = gi.j fi,j (t).
Then, (9) can be rewritten as follows: Consider the T–S fuzzy faulty system that is described in (2).
p We assume that only actuator faults occur, and no sensor fault
ufi (t) = (1 − ρi (t))ui (t) + ai,j (t). (10) is involved. For simplicity, we consider the case that only one
j =1 single actuator is faulty at one time. The actuator fault-diagnosis
problem is formulated as follows: With the available output y,
Denote we propose an observer-based scheme to identify the faulty
Γ(t) = diag(ρ1 (t), . . . , ρm (t)) (11) actuator, and then estimate the fault.
p
To solve the problem, we will design a bank of SMOs with
desired actuator fault-detection and fault-estimation properties.
F (t) = [f1 , f2 , . . . , fm ]T , fi = ai,j (t). (12)
Thus, the following assumptions are made in this paper.
j =1
Assumption 1: Matrix Bi is of full column rank, and the pair
Then, we have (Ai , Ci ) is observable.
Assumption 2: There exist known positive constants ρ̄i , ρ̄¯i ,
uf (t) = (I − Γ(t))(u(t) + F (t)), t ≥ tj (13)
ρ̄1 , ρ̄2 such that |ρi (t)| ≤ ρ̄i , |ρ̇i (t)| ≤ ρ̄¯i , ρ̄1 = max{ρ̄1 ,
where the failure time instant tj is unknown, and I denotes the ρ̄2 , . . . , ρ̄m }, ρ̄2 = max{ρ̄¯1 , ρ̄¯2 , . . . , ρ̄¯m }, i = 1, . . . , m.
identity matrix with appropriate dimensions. In this paper, both Assumption 3: There exist known positive constants
bias and gain faults are handled by considering the general fault ¯i,j such that |ai,j (t)| ≤ āi,j , |ȧi,j (t)| ≤ ā
ā1 , ā2 , āi,j , ā ¯i,j , ā1 =
model (13). max{ā1,1 , . . . , āi,p , . . . , ām ,1 , . . . , ām ,p }, ā2 =max{ā ¯1,1 , . . . ,
Notice that, in the following, just for the sake of notational ¯i,p , . . . , ā
ā ¯m ,p }, i = 1, . . . , m, j = 1, . . . , p.
¯m ,1 , . . . , ā
simplicity, we will use hi , ρi , and ai,j to denote hi (z(t)), ρi (t), Our actuator fault-diagnosis and accommodation scheme
and ai,j (t), respectively. consists of FDI and FTC. We first design the fault-diagnosis ob-
Now, the control objective is redefined as follows. An active server utilizing SMOs to detect, isolate, and estimate the fault,
FTC approach is proposed to make the system (2) stable under and then propose an FTC method to compensate the fault.
normal and faulty conditions. Under normal condition (no fault),
a state feedback control input u(t) is designed such that the
B. Fault Detection
system (2) is stable. Meanwhile, the FDI algorithm is working.
As soon as an actuator fault is detected and isolated, the fault- In order to detect the actuator faults, we design a fuzzy state-
estimation algorithm is activated. The obtained fault estimation space observer for the system (8), which is described as follows.
SHEN et al.: FAULT-TOLERANT CONTROL FOR T–S FUZZY SYSTEMS WITH APPLICATION TO NEAR-SPACE HYPERSONIC VEHICLE 655

Observer Rule i: IF z1 (t) is Mi1 and · · · zq (t) is Miq , THEN ≤ −hi (z(t))[λm in (Qi )/λm ax (P )]V (t) = −κV (t)

˙
x̂(t) = Ai x̂(t) + Bi u(t) + Li (y(t) − ŷ(t)) (20)
(14)
ŷ(t) = Ci x̂(t) where
 
where Li , i = 1, . . . r, is the observer gain for the ith observer λm in (Q1 ) λm in (Q2 ) λm in (Qr )
rule. κ = min , ,..., ∈ R.
λm ax (P ) λm ax (P ) λm ax (P )
The overall fuzzy system is inferred as follows:
⎧ Hence
r
⎪ ˙
⎨ x̂(t) hi (z(t))(Ai x̂(t) + Bi u(t) + Li (y(t) − ŷ(t))
= V1 (t) ≤ e−κt V (0). (21)
i=1
⎪ r
⎩ ŷ(t) = hi (z(t))Ci x̂(t). Furthermore, we have
i=1
(15) λm in (P )||ex (t)||2 ≤ e−κt λm ax (P )||ex (0)||2 . (22)
Denote
Therefore, the norm of the error vector satisfies
ex = x(t) − x̂(t), ey = y(t) − ŷ(t). (16) 
e−κt λm ax (P )
||ex (t)|| ≤ ||ex (0)||
Then, the error dynamics are described by λm in (P )
⎧ 
r 
⎪
⎨ ėx = hi (z(t))(Ai − Li Ci )ex (t)) = λm ax (P )/λm in (P )||ex (0)||e−κt/2 . (23)
i=1
 (17)
⎪
⎩ ey =
r
Furthermore, the detection residual can be defined as
hi (z(t))Ci ex (t).
i=1
J = ||y(t) − ŷ(t)||. (24)
Lemma 1: The estimation error ex converges asymptotically
to zero if there exist matrices P = P T > 0 and Qi > 0 with From (23), it can be seen that the following inequality holds
appropriate dimensions such that the following LMI is satisfied: in the healthy case:
r 
P (Ai − Li Ci ) + (Ai − Li Ci )T P ≤ −Qi ∀i = 1, 2, . . . , r. J≤ hi (z(t)) λm ax (P )/λm in (P )||Ci ||||ex (0)||e−κt/2 .
(18) i=1
Proof: Consider the following Lyapunov function: (25)
Then, the fault detection can be performed using the following
V1 = eTx (t)P ex (t). mechanism:

Differentiating V1 with respect to time t, one has J ≤ Td , no fault occurred
(26)
r J > Td , fault has occurred
V̇1 (t) = hi (z(t)) eTx (t)(P (Ai − Li C) where threshold Td is defined as follows:
i=1
r 
+ (Ai − Li C)T P )ex (t) Td = hi (z(t)) λm ax (P )/λm in (P )||Ci ||||ex (0)||e−κt/2 .
r i=1

≤− hi (z(t)) eTx (t)Qi ex (t) Remark 2: It is easy to find from (20) that if no actuator
i=1 fault occurs, we have limt→∞ ex = 0. If there is an actuator
≤ 0. (19) fault, then limt→∞ ex = 0. Therefore, in some existing work,
the fault detection is carried out as
Because V1 (t) ∈ L∞ is a monotonic and nonincreasing bounded 
function, V1 (+∞)exists. Hence, we have V1 (0)− V1 (+∞)≥ lim ex = 0, no fault occurred
+∞ r
t→∞ (27)
− 0 lim ex = 0, fault has occurred
i=1 hi (z(t))[ex (t)Qi ex (t)], i.e., ex (t) ∈ L2 . And
T
t→∞
since ex (t), ėx (t) ∈ L∞ , using the Lyapunov stability the-
and the aforementioned observer that is given by (15) was re-
ory, we obtain limt→∞ ex (t) = 0. Furthermore, we have
ferred to as the fault-detection observer for the system described
limt→∞ ey (t) = 0. The proof is complete. 
by (2). However, it is valuable to point out that ex (∞) is not
From Lemma 1, we have
available in practice; thus, ex (∞) = 0 cannot be considered as
r
an indicator of fault occurrence. That is to say, the fault de-
V̇1 (t) ≤ − hi (z(t)) eTx (t)Qi ex (t) tection (27) does not work in practical applications. Therefore,
i=1
mechanism (26) is more efficient for fault detection in practical
r
cases.
≤− hi (z(t)) λm in (Qi )eTx (t)ex (t)
i=1
C. Fault Isolation
r
≤− hi (z(t))[λm in (Qi )/λm ax (P )eTx (t)P ex (t)] Since the system has m actuators and it is assumed that only
i=1 one single fault occurs at one time, we have m possible faulty
656 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 20, NO. 4, AUGUST 2012

cases in total. When the sth (1 ≤ s ≤ m) actuator is faulty, the r 

faulty model can be described as = hi (z(t)) (Ai − Li Ci )eis (t)
⎧ i=1

⎪  r 
r
⎪
⎪ x˙s (t) = hi (z(t))Ai xs (t) + hi (z(t))Bi u(t)
⎪
⎪ + bi,s (−ρs (t)us (t) − μs ρ̄s |us (t)|)
⎪
⎨
i=1 i=1
 r p
− hi (z(t))bi,s [ρs (t)us (t) − as,j (t)] (28) p 
⎪
⎪ i=1 j =1
⎪
⎪  + (as,j (t) − μs ās,j ) . (31)
⎪
⎪
r
⎩ y(t) = hi (z(t))Ci x(t) j =1
i=1
For s = l, we have
where Bi = [bi,1 , bi,2 , . . . , bi,m ], bi,l ∈ Rn ×1 , 1 ≤ l ≤ m.
r
ρs (t), as,j (t), j = 1, 2, . . . , p denote the time profiles of the sth
actuator fault, which are described by (10); us (t) is the desired ėxs (t) = hi (z(t))(Ai − Li Ci )eis (t)
i=1
controller when the sth actuator is healthy.
Inspired by the SMOs in [10], we are ready to present one of r

the results of this paper. It is assumed that fuzzy-observer and + hi (z(t))[ − bi,l ρl (t)ul (t) − bi,s μs ρ̄s |us (t)|)
i=1
fuzzy-control systems have the same premise variables z(t);
p
then the following fuzzy observers are proposed to isolate the 
actuator fault. + (bi,l al,j (t) − bi,s μs ās,j ]. (32)
Isolation Observer Rule i: IF z1 (t) is Mi1 and · · · zq (t) is j =1

Miq , THEN The stability of the state error dynamics is guaranteed by the
⎧ following theorem.
⎪
⎪ x̂˙ is (t) = Ai x̂is (t) + Li (y(t) − ŷis (t)) + Bi u(t)
⎪
⎨   Theorem 1: Under Assumptions 1–3, if there exist a common
p
symmetric positive-definite matrix P and matrices Li , Fi , Qi >
+bi,s μs ρ̄s |us (t)| + ās,j (29)
⎪
⎪ 0, i =1, 2, . . . , r, with appropriate dimensions, such that the
⎪
⎩
j =1
ŷis (t) = Cis x̂is (t) following conditions hold:

where x̂is (t), ŷis (t) are the sth fuzzy observer’s state and output, (Ai − Li Ci )T P + P (Ai − Li Ci ) ≤ −Qi (33)
respectively. Li is the observer’s gain matrix for the ith observer. P Bi = (Fi Ci )T (34)
The global fuzzy observer is represented as
⎧ r 
r
then when the lth actuator is faulty, for s = l, limt→∞ exs = 0,
⎪
⎪ x̂˙ s (t) = hi (z(t))Ai x̂is (t) + hi (z(t))Li (y(t) and for s = l, limt→∞ exs = 0.
⎪
⎪
⎪
⎪ i=1 i=1 Proof:
⎪
⎪ r
⎪
⎪ −ŷis (t)) + 1) For s = l, according to (31), we have
⎪
⎪
hi (z(t))Bi u(t)
⎪
⎪ i=1  
⎨ r p r
+ hi (z(t))bi,s μs ρ¯s |us (t)| + ās,j ėxs (t) = hi (z(t))(Ai − Li Ci )eis (t)
⎪
⎪
⎪
⎪ 
i=1 i=1
i=1
⎪
⎪
r

⎪
⎪ ŷs (t) = hi (z(t))Ci x̂s (t) r
⎪
⎪  r
⎪
⎪
i=1
  + hi (z(t))bi,s (−μs ρ̄s |us (t)| − ρs (t)us (t))
⎪
⎪
r
⎩ μs = − hi (z(t))Fis ey s (t) || hi (z(t))Fis ey s (t)|| i=1
i=1 i=1 p p 
(30)
where Fis ∈ R is the sth row of Fi ∈ R
1×n m ×n
, which will − μs ās,j + as,j (t) .
be defined later; Li ∈ Rn ×n is chosen such that Ai − Li Ci j =1 j =1

is Hurwitz; exs (t) = xs (t) − x̂s (t) and ey s (t) = y(t) − ŷs (t) Define the following Lyapunov function:
are, respectively, the state error and the output error between the
plant and the sth SMO observer. V2 (t) = eTxs (t)P exs (t). (35)
For s = l, the error dynamics is obtained from (28) and (30):
Differentiating V2 with respect to time t, and using (33), one
r r has
ėxs (t) = hi (z(t))Ai eis (t) − hi (z(t))Li (y(t) − ŷis (t))
i=1 i=1 V̇2 (t) = ėTxs (t)P exs (t) + eTxs (t)P ėxs (t)
r  r
+ hi (z(t))bi,s (−ρs (t)us (t) − μs ρ̄s |us (t)|) ≤ −eTxs (t)Qi exs (t) + 2eTxs (t)P hi (z(t))bi,s ·
i=1 i=1
p   p p 
+ (as,j (t) − μs ās,j ) (−μs ρ̄s |us (t)|−ρs (t)us (t))− μs ās,j + as,j (t) .
j =1 j =1 j =1
SHEN et al.: FAULT-TOLERANT CONTROL FOR T–S FUZZY SYSTEMS WITH APPLICATION TO NEAR-SPACE HYPERSONIC VEHICLE 657

 
From μs = − ri=1 hi (z(t))Fis ey s (t)/|| ri=1 hi (z(t))Fis Hence, the isolation law for actuator fault can be designed as
ey s (t)|| and (34), one has
r 
Js (t) ≤ TI , l = s ⇒ the lth actuator is faulty
2eTxs (t)P hi (z(t))bi,s (−μs ρ̄s |us (t)| − ρs (t)us (t)) ≤ 0 (41)
Js (t) > TI , l = s
i=1

r  p p 
where the threshold TI is defined as follows:
2eTxs (t)P hi (z(t))bi,s − μs ās,j + as,j (t) ≤ 0.
i=1 j =1 j =1

Hence
r 
TI = hi (z(t)) λm ax (P )/λm in (P )||ey s (0)||e−κt/2 .
V̇2 (t) ≤ −eTxs (t)Qi exs (t) ≤ 0. (36) i=1

Because V2 (t) ∈ L∞ is a monotonic and nonincreasing bounded  

function, V2 (+∞) exists. Hence, we have V2 (0)− V2 (+∞) ≥ Note that μs = − ri=1 hi (z(t))Fis ey s (t)/|| ri=1 hi (z(t))
+∞
− 0 eTxs (t)Qi exs (t), i.e., exs (t) ∈ L2 . Since exs (t) and Fis ey s (t)|| in (30), which denominator contains ey s (t). Just as
ėxs (t) ∈ L∞ , using the Lyapunov stability theory, we have pointed out in [10], the chattering phenomenon occurs when
limt→∞ exs (t) = 0. Thus, we have limt→∞ ey s (t) = 0. ey s (t) → 0 in practice. Inspired by [33], in order to reduce this
2) For s = l, it follows from (28) and (30) that chattering in practical applications, we modify SMOs (30) by
r the introduction of a positive constant δ as follows:
ėxs (t) = hi (z(t))(Ai − Li Ci )eis (t)
i=1 ⎧ r r
 ⎪
⎪ x̂˙ s (t) = hi (z(t))Ai x̂s (t) − hi Li (ŷs (t) − y(t))
r ⎪
⎪
⎪
⎪ i=1 i=1
+ hi (z(t)) (−bi,l ρl (t)ul (t) − bi,s μs ρ̄s |us (t)|) ⎪
⎪ r
⎪
⎪ + hi (z(t))Bi u(t)
i=1 ⎪
⎪
 ⎪
⎪ i=1  
p ⎪
⎪ r p
⎪
⎪ − hi (z(t))μs ρ̄s |us (t)| +
+ (bi,l al,j (t) − bi,s μs ās,j ) . ⎨ ās,j
i=1 j =1
j =1
⎪ r
⎪
⎪ ŷs (t) = hi (z(t))Cis x̂s (t)
Because the matrix Bi is of full-column rank (see Assump- ⎪
⎪
⎪
⎪ i=1 
tion 1), we know that bis and bil are linearly independent. ⎪
⎪ r
⎪
⎪ μ = − h (z(t))F e (t)
Therefore ⎪
⎪ s i is y s
⎪
⎪  i=1 
r ⎪
⎪ 
r
⎪
⎩ ||
lim hi (z(t))[(−bi,l ρl (t)ul (t) − bi,s μs ρ̄s |us (t)|) hi (z(t))Fis ey s (t)|| + δ
t→∞ i=1
i=1 (42)
p where δ > 0 ∈ R is a constant, s = 1, 2, . . . , m. Obviously,
+ (bi,l al,j (t) − bi,s μs ās,j ) = 0. (37) the denominator of μs will converge asymptotically to δ when
j =1 ey s (t) → 0, which reduces this chattering phenomenon.
Thus, we have limt→∞ exs (t) = 0 and limt→∞ ey s (t) = 0. From the aforementioned analysis, it is easy to find that a
From 1) and 2), we obtain the conclusions. This ends the suitable threshold δ must be selected such that Js (s = l) tends to
proof. 
be very small when the lth actuator is faulty, while other residuals
Now, we denote the residuals between the real system and Js (s = l) are not equal to zero on any small time intervals. Thus,
SMOs as follows: the modified SMOs can not only decrease the chattering problem
in practice, but can also realize fault diagnosis successfully.
Js (t) = ey s (t) = ŷs (t) − y(t) , 1 ≤ s ≤ m. (38)
According to Theorem 1, when the lth actuator is faulty, i.e.,
s = l, the residual Js (t) must tend to zero, while for any s = l, D. Fault Estimation
basically, Js (t) does not equal zero. Furthermore, from Lemma After fault isolation, we can estimate the fault. Assume that
1, we have that if l = s, then the sth (1 ≤ s ≤ m) actuator is faulty; the faulty system can be
r  described as
Js (t) ≤ hi (z(t)) λm ax (P )/λm in (P )||ey s (0)||e−κt/2
i=1 ⎧ r  r
(39) ⎪
⎪
⎪
ẋ(t) = hi (z(t))Ai x(t) + hi (z(t))Bi u(t)
and if l = s, then ⎪
⎪ i=1 i=1
⎨ r p
r  − hi (z(t))bi,s [ρs us (t) − as,j (t)] (43)
Js (t) > hi (z(t)) λm ax (P )/λm in (P )||ey s (0)||e−κt/2 . ⎪
⎪ i=1 j =1
⎪
⎪
⎩ y(t) =  hi (z(t))Ci x(t).
⎪ r
i=1
(40)
i=1
658 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 20, NO. 4, AUGUST 2012

To estimate the fault, an observer is presented as follows: ˙ = P {2η2 Fi,s ey }

âi,j
⎧ r r ⎧
⎪
⎪ ˙
x̂(t) = hi (z(t))Ai x̂(t) + hi (z(t))Bi u(t) ⎨0, âi,j > a¯1 and 2η2 Fi,s ey > 0
⎪
⎪
⎪
⎪ i=1  i=1  = or âi,j < −ā1 and 2η2 Fi,s ey < 0 (49)
⎪
⎪ r p ⎩
⎪
⎪ − hi (z(t))bi,s ρˆs us (t) − 2η2 Fi,s ey , otherwise, i = 1, 2, . . . , m, j = 1, 2, . . . , p
⎨ âs,j
i=1 j =1
(44) where Fis ∈ R1×n is the sth row of Fi ∈ Rm ×n , η1 > 0, η2 > 0
⎪
⎪ r
⎪
⎪ + h (z(t))L (y(t) − ŷ(t))
⎪
⎪
i i denote the adaptive rates, then the error system (45) is asymp-
⎪
⎪
i=1
⎪
⎪ r totically stable. Moreover, ex (t), ρ̃s , and ãs,j are semiglob-
⎩ ŷ(t) = hi (z(t))Ci x(t) ally uniformly ultimately bounded, converging asymptotically
i=1 
where ρ̂s , âs,j are the estimated values of ρs (t), as,j (t) at time
to a small
√ neighborhood of zero, √ namely, |ex | ≤ α/λm in (P ),
|ρ̃i | ≤ 2η1 α, and |g̃i,j | ≤ 2η2 α, where
t.
r  p 
Remark 3: Many results about the observer design were re- 2ρ̄1 (2ρ̄1 + ρ̄2 ) 2ā1 (2ā1 + ā2 )
ported in the literature. For faulty systems with only bias fault μ0 = hl (z(t)) +
η1 j =1
η2
fa described as follows: l
  
ẋ(t) = Ax(t) + B(u(t) + fa ) λm in (Q1 ) λm in (Qr )
λ0 = min ,..., ,1
ŷ(t) = Cx(t) λm ax (P ) λm ax (P )

an observer is classically designed in the following form: and α =μ0 /λ0 + V (0).
˙ Proof: Define the following smooth function:
x̂(t) = Ax̂(t) + B(u(t) + fˆa ) + L(y(t) − ŷ(t))
ŷ(t) = C x̂(t). V = V1 + V2 + V3 (50)

Let ex (t) = x(t) − x̂(t); then the error dynamics is described where
by
V1 = eTx (t)P ex (t) (51)
ėx (t) = (A − LC)ex (t) + B(fa − fˆa ) r  
1 2
V2 = hi (z(t)) ρ̃ (t) (52)
where fˆa denotes the estimation of fa . However, in this paper, 2η1 s
i=1
actuator bias faults and gain faults are both considered; the above
r p  
observer does not work. The novel observer (44) is proposed in 1 2
order to estimate the two kinds of fault. V3 = hi (z(t)) as,j (t) . (53)
i=1 j =1
2η2
Using (43) and (44), the error dynamics is obtained:
Differentiating V, Vi , i = 1, 2, 3, with respect to time t, leads
r to
ėx (t) = hi (z(t))[(Ai − Li Ci )ex (t))]
i=1 V̇ = V̇1 + V̇2 + V̇3 (54)
r  p 
where
− hi (z(t))bi,s ρ̃s us − ãs,j (45)
r
i=1 j =1
V̇1 = hi (z(t)) eTx (t)(P (Ai − Li Ci )
where ex (t) = x(t) − x̂(t), ρ̃s = ρs (t) − ρ̂s , ãs,j = as,j (t) − i=1
âs,j .
+ (Ai − Li Ci )T P )ex (t)
Now, an adaptive fault-diagnostic algorithm is proposed to
estimate the actuator fault. The stability of the error dynamics r  p 
is guaranteed by the following theorem. − hi (z(t)) 2eTx (t)P bi,s ρ̃s us − 2eTx (t)P bi,s ãs,j
Theorem 2: Under Assumptions 1–3, if there exists a common i=1 j =1

symmetric positive-definite matrix P , and real matrices Li and (55)

Qi > 0, i =1, 2, . . . , r, with appropriate dimensions, such that
the following conditions hold:
r   r  
1 ˙ 1
V˙2 = hi (z(t)) ρ̃s ρ̃s = hi (z(t)) ρ˜s (ρ˙s − ρˆ˙s
P (Ai + Li Ci ) + (Ai + Li Ci )T P < −Qi (46) η1 η1
i=1 i=1
T
P Bi = (Fi Ci ) (47) r r
1 1
= hi (z(t)) ρ˜s ρ˙s − hi (z(t)) ρ˜s ρˆ˙s (56)
ρ̂˙i = P {−2η1 Fi,s ey us } η1 η1
⎧ i=1 i=1
⎨ 0, ρ̂i = ρ̄1 and − 2η1 Fi,s ey > 0 r p
ãs,j ã˙ s,j
= or ρ̂i = −ρ̄1 and − 2η1 Fi,s ey < 0 (48) V̇3 = hi (z(t))
⎩ η2
−2η1 Fi,s ey us , otherwise, i = 1, 2, . . . , m i=1 j =1
SHEN et al.: FAULT-TOLERANT CONTROL FOR T–S FUZZY SYSTEMS WITH APPLICATION TO NEAR-SPACE HYPERSONIC VEHICLE 659

r p p 
ãs,j (ȧs,j − â˙ s,j ) 2ρ̄1 (2ρ̄1 + ρ̄2 ) 2ā1 (2ā1 + ā2 )
= hi (z(t)) + +
i=1 j =1
η2 η1 j =1
η2
r p r p r  p
ãs,j ȧs,j ãs,j â˙ s,j ρ̃2i ã2i,j
= hi (z(t)) − hi (z(t)) . ≤ hl (z(t)) − ex Qi ex −
T
−
i=1 j =1
η2 i=1 j =1
η2 η1 j =1 η2
l

(57) p 
2ρ̄1 (2ρ̄1 + ρ̄2 ) 2ā1 (2ā1 + ā2 )
+ +
Substitution of (55)–(57) into (54) yields η1 j =1
η2
r r ⎡ ⎤
1 r 2 p
ã2
V̇ = − hi (z(t))eTx Qi ex + hi (z(t)) ρ̃s ρ˙s ρ̃
hl (z(t)) ⎣−eTx Qi ex − i − + μ⎦
i,j
η1 ≤
i=1 i=1 η1 j =1 η2
l=1
r p
1 ⎡ ⎤
+ hi (z(t)) ãs,j ȧs,j r
ρ̃2i
p
ã2i,j
i=1 j =1
η2 ≤ hl (z(t)) ⎣−λm in (Qi )eTx ex − − + μ⎦
2η1 2η2
r   l=1 j =1
1 ⎡ ⎤
− hi (z(t))ρ̃s 2eTx P bi,s us + ρˆ˙s r p
η1 λ (Q ) ρ̃2 ã2
hl (z(t)) ⎣− + μ⎦
i=1 m in i i,j
≤ eT P ex − i −
r p  λm ax (P ) x 2η1 j =1 2η2
1 l=1
+ hi (z(t))ãs,j 2eTx P bi,s − â˙ s,j ). (58)
i=1 j =1
η2 ≤ −λ0 V (t) + μ0 (60)
Substitution of (48) and (49) into (58) yields where
r r
1 p
V̇ = − hi (z(t))eTx Qi ex + hi (z(t)) ρ̃s ρ˙s 2ρ̄1 (2ρ̄1 + ρ̄2 ) 2ā1 (2ā1 + ā2 )
η1 μ= +
i=1 i=1 η1 j =1
η2
r p
1 r  p 
+ hi (z(t)) ãs,j ȧs,j . (59) 2ρ̄1 (2ρ̄1 + ρ̄2 ) 2ā1 (2ā1 + ā2 )
η2 μ0 = hl (z(t)) +
i=1 j =1 η1 j =1
η2
l
Since we have  
λm in (Q1 ) λm in (Q2 ) λm in (Qr )
ρ̃i ρ̇i ρ̃2 ρ̃i (ρ̃i + ρ̇i ) ρ̃2 (ρi − ρ̂i )(ρi − ρ̂i + ρ̇i ) λ0 = min , ,..., ,1 .
=− i + =− i + λm ax (P ) λm ax (P ) λm ax (P )
η1 η1 η1 η1 η1
ρ̃2i (|ρi | + |ρ̂i |)(|ρi | + |ρ̂i | + |ρ̇i |) Then, one has d
dt (V (t)eλ0 t ) ≤ eλ0 t μ0 . Furthermore
≤− +
η1 η1  
μ0 μ0 −λ0 t μ0
p p
ã2i,j
p
ã2i,j
p 0 ≤ V (t) ≤ + V (0) − e ≤ + V (0).
ãi,j ȧi,j ãi,j ȧi,j λ0 λ0 λ0
=− + +
j =1
η2 j =1
η2 j =1
η2 j =1
η2 
Let α = λ μ0
+ V (0); one has |ex | ≤ λ α (P ) , |ρ̃i | ≤
p
ã2i,j
p
(|ai,j | + |âi,j |)(|ai,j | + |âi,j | + |ȧi,j |) √ 0
√ m in

≤− + 2η1 α, and |ãi,j | ≤ 2η2 α. This ends the proof.

j =1
η2 j =1
η2 
Remark 4: If there exist two known constants fm in , fm ax such
and |ρ̂i (t)| ≤ ρ̄1 and |âi,j (t)| ≤ ā1 , which can be guaranteed by that fm in ≤ |f (t)| ≤ fm ax , then the fault f (t) can be approxi-
using the adaptive laws (48) and (49), and Assumptions 2 and 3 mated by the following form:
(i.e., |ρi (t)| ≤ ρ̄1 , |ρ̇i (t)| ≤ ρ̄2 , |ai,j (t)| ≤ ā1 , and |ȧi (t)| ≤ ā2 )
are satisfied, one has 1
f (t) = (fm ax − fm in )(1 − tanh ζ) + fm in (61)
ρ̃i ρ̇i ρ̃2 2ρ̄1 (2ρ̄1 + ρ̄2 ) 2
≤− i +
η1 η1 η1 where ζ is an unknown constant. Thus, the fault f (t) is estimated
p
ãi,j ȧi,j ã2i,j
p p
2ā1 (2ā1 + ā2 ) through the estimation of ζ̂, namely
≤− + .
η2 η 2 η2 1
j =1 j =1 j =1
fˆ(t) = (fm ax − fm in )(1 − tanh ζ̂) + fm in . (62)
2
Hence, from (59), one has
r  p This method prevents the phenomenon of parameter drift in the
ρ̃2i ã2i,j presence of bounded disturbances because of | tanh ςˆ| < 1, and
V̇ ≤ hl (z(t)) − ex Qi ex −
T
−
η1 j =1 η2 ensures fm in ≤ |fˆ(t)|≤ fm ax .
l=1
660 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 20, NO. 4, AUGUST 2012

E. Fault Accommodation then the system (2) is asymptotically stable under the feedback
FTC (65), and all signals that are involved in the closed-loop sys-
After that the fault information is obtained, we will consider
the FTC problem of the system (2), and design an FTC law to tem are semiglobally uniformly ultimately bounded, converging
asymptotically
 to a small √ neighborhood of √ zero, namely, |e| ≤
recover the control system’s dynamics performance when an ac-
α/λm in (P ), |ρ̃i | ≤ 2η1 α, and |ãi,j | ≤ 2η2 α, where α =
tuator fault occurs. First, we consider the fuzzy-control problem
for the following nominal system without actuator faults: V (0) + μ0 /λ0 , λ0 = min{ λλm i n (Q 1)
, . . . , λλm i n (Q r)
, 1}, μ0 =
m a x (P ) m a x (P )
⎧ r p 2ā 1 (2ā 1 + ā 2 )
r
l hl (z(t))[
2 ρ̄ 1 (2 ρ̄ 1 + ρ̄ 2 )
+ j =1 ].
⎪
⎨ ẋ(t) = hi (z(t))(Ai x(t) + Bi u(t)) η1 η2
i=1
Proof: Similar to the proof of Theorem 2, it is easy to ob-
⎪ 
r tain the conclusions of Theorem 3. The detailed proof is, thus,
⎩ y(t) = hi (z(t))Ci x(t).
i=1
omitted here. 

The PDC technique offers a procedure to design a fuzzy-

IV. SIMULATION RESULTS
control law from a given T–S fuzzy model. In the PDC design,
each control rule is designed from the corresponding rule of T–S A. Near-Space Hypersonic Vehicle Modeling and Analysis
fuzzy model. The designed fuzzy controller has the same fuzzy
Considering the longitudinal flight mode of NSHV, a math-
sets as the considered fuzzy system.
ematical model for a generic NSHV which is developed at the
Control Rule i: IF z1 (t) is Mi1 and · · · zq (t) is Miq , THEN
NASA Langley Research Center is presented in [8]. The longitu-
ui (t) = Ki x(t) dinal dynamics of NSHV can be described by a set of differential
equations involving its velocity V , flight-path angle γ, altitude
and the overall fuzzy controller is given as follows: h, angle of attack α, and pitch rate q as
r
T cos α − D u sin γ
u(t) = hi (z(t))Ki x(t) (63) V̇ = − (70)
m r2
i=1
L + T sin α (μ − V r2 ) cos γ
where the controller gain matrix Ki is determined by solving γ̇ = + (71)
mV V r2
the following LMI:
ḣ = V sin γ (72)
P (Ai + Bi Ki ) + (Ai + Bi Ki )T P < −Qi (64)
α̇ = q − γ̇ (73)
where P = P T > 0 and Qi > 0 are the matrices with appropri-
My y
ate dimensions. q̇ = (74)
On the basis of the estimated actuator fault, the fault-tolerant Iy y
controller is constructed as where L = q̄SCL , D = q̄SCD , T = q̄SCT , r = h + Re ,
 N p i 
us − j =1 âi,j My y = q̄Sc̄[CM (α) + CM (δe ) + CM (q)], CL = 0.6203α,
us = (65) CD = 0.6450α2 + 0.0043378α + 0.003772, CM (δe ) =
(1 − ρ̂s )
ce (δe − α), CM (q) = (c̄/2V )q(−6.796α2 + 0.3015α −
where uN s is the sth normal control input, and ρ̂s , âi,j are the
0.2289), CM (α) = −0.035· α2 + 0.036617(1 + ΔCM α )α +
estimations of ρs , ai,j , which are used to compensate for the 5.3261e − 06, and
gain fault and bias fault. 
0.02576δT , when δT < 1
Theorem 3: Consider system (2) under Assumptions 1–3. If CT =
0.0224 + 0.00336δT , when δT > 1.
there exist a common symmetric positive-definite matrix P , and
real matrices Li and Qi > 0, i = 1, 2, . . . , r, with appropriate The aircraft mass m, the gravitational constant μ, the mo-
dimensions, such that the following conditions hold: ments of inertia Iy y , and the pitch moment coefficients are the
parameters. The aerodynamic coefficients and inertia data are
P (Ai − Li Ci ) + (Ai − Li Ci )T P < −Qi (66) coupled with state variables and control inputs. The control
P Bi = (Fi Ci )T (67) input vector is u(t) = [δe , δT ]T , where δe is the elevator detec-
tion, and δT is the throttle setting. The longitudinal model of
ρ̂˙i = P {−2η1 Fi,s ey us } the NSHV that is described by (70)–(74) can be written in the
⎧
⎨ 0, ρ̂i = ρ̄1 and − 2η1 Fi,s ey > 0
⎪ following affine nonlinear form:

= or ρ̂i = −ρ̄1 and − 2η1 Fi,s ey < 0 (68) ẋ(t) = f (x) + g(x)u(t)
⎪
⎩ (75)
−2η1 Fi,s ey us , otherwise, i = 1, . . . , m y(t) = Cx(t)
˙ = P {2η2 Fi,s ey }
âi,j where x(t) = [V, γ, h, α, q]T ∈ Rn denotes the state vector,
⎧ u(t) = [δe , δT ]T ∈ Rm denotes the control input vector, and
⎪
⎨ 0, âi,j > ā1 and 2η2 Fi,s ey > 0 y(t) is the output vector.
= or âi,j < −ā1 and 2η2 Fi,s ey < 0 (69) In this section, some simulation results are presented to
⎪
⎩
2η2 Fi,s ey , otherwise, i = 1, . . . , m, j = 1, . . . , p demonstrate the effectiveness of the proposed techniques. For
SHEN et al.: FAULT-TOLERANT CONTROL FOR T–S FUZZY SYSTEMS WITH APPLICATION TO NEAR-SPACE HYPERSONIC VEHICLE 661

the purpose of this study, the aerodynamic coefficients are sim-

plified around the cruising flight mode. The nominal flight ⎧
of the NSHV is at a trimmed cruise conditions: Mach = 15, ⎨ y2 (t), t<5
uf2 (t) = 
p
V = 15060 ft/s, and h = 110 000 ft/s. ⎩ (1 − ρ2 (t))(y2 (t) + g2,j f2,j (t)), t ≥ 5
j =1
If each state variable is selected as a premise variable, then the
number of fuzzy rules will become too large. However, from the where ρ2 (t) = 0.4 sin(πt), p = 1, g2,1 = 0.4, f2,1 (t) = cos(t).
property of NSHV, we know that the angle of attack α is a key In order to compare with the results in [9] and [11], we
variable that affects the nonlinear characteristic of NSHV, and consider the following cases.
the velocity V has constraint relationship to the altitude h and the Case 2 (Bias Fault) [9]:
pitch angle θ = α + γ. Similar to [10], we select x̄ = [V, θ, q]T
as a new state vector. As a result, we denote z1 = V , z2 = α + γ, uf1 (t) = u1 (t)
z3 = q, and select z1 , z2 , and z3 as premise variables for the T–S
fuzzy system model. Hence, it can not only reduce the number
of fuzzy rules but can well approximate the nonlinear system ⎧
⎨ 0, t < 4s
and characterize the NSHV model as well [10]. Furthermore, uf2 (t) = u2 (t) + f2,1 (t), f2,1 (t) = 5, t ≥ 4s
we assume ⎩
5 + 2(t − 7), t ≥ 7s
z1 ∈ (6000 16000) m/s, z2 ∈ (−0.5 0.5) rad/s where ρ2 (t) = 0, p = 1, g2,1 = 1.
z3 ∈ (−0.5 0.5) rad/s. Case 3 (Gain Fault) [11]:

Suppose that each premise variable has two associated fuzzy uf1 (t) = u1 (t)
sets:

{z1 = 6000, 16000}, {z2 = −0.5, 0.5}, {z3 = −0.5, 0.5}. 

0, t < 2s
uf2 (t) = (1 − ρ2 (t))u2 (t), ρ2 (t) =
0.4, t ≥ 2s
The corresponding fuzzy membership functions are defined
as where ρ2 (t) = 0, p = 0, and g2,1 = 0.
Remark 5: If each state variable of the NSHV model is
Mz 1 =6000 = exp[−(z1 /ς1 )2 ], Mz 1 =16000 = 1 − Mz 1 =6000
selected as premise variable, then the number of fuzzy rules
1 becomes too large, which leads to the increasing amount of
Mz 2 =−0.5 =
1 + exp[((z2 )2 − σ)/ς2 ] computing and, thus, affects the setting time of the closed-
Mz 2 =0.5 = 1 − Mz 2 =−0.5 loop system. In order to reduce the number of fuzzy rules,
   taking into account the main characteristics of the NSHV, we
z3 select x̄ = [V, θ, q]T as premise variables where θ = α + γ. As
Mz 3 =−0.5 = exp − − σ̄ , Mz 3 =−0.5 = 1−Mz 3 =−0.5
ς3 pointed out in [10], it can not only reduce the number of fuzzy
rules, but it provides also a good approximation of the nonlin-
where the unknown parameters σ, σ̄, ς1 , ς2 , ς3 should be se- ear system. As a result, it can achieve satisfactory accuracy and
lected to symmetrically cover the space of the input variables. dynamic performance of the proposed FTC.
We choose eight working points of NSHV, which are [z1 , z2 ,
z3 ]T =: [6000, − 0.5, 0.5]T , [6000, 0.5 , 0.5]T , [6000, − 0.5,
−0.5]T ,[6000,0.5 , − 0.5]T , [−0.5, 0.5, 5000]T , [16000, 0.5, B. Simulation Results
0.5], [16000,0.5, −0.5], [16000, 0.5, − 0.5]T , respectively. The By using the MATLAB toolbox to solve the matrices inequal-
parameters of the membership are selected as σ = 0.15, σ̄ = 4, ities (18), one can obtain the fault-diagnostic observer gains Li .
ς1 = 3200, ς2 = 0.05, ς3 = 0.4. Then, eight plant rules and cor- By solving (64) and (67), one can obtain the positive-definite
responding control rules can be obtained. We give the first rule symmetric matrix P and the nominal controller gains Ki . Be-
as an example, and the other rules have the similar form. cause of the space limitation, only the common matrix P, and
Rule 1: IF z1 is about 6000 m/s, z2 is about −0.5 rad/s, and z3 the matrices Q1 ,L1 , K1 of the first working point of the NSHV
is about −0.5 rad /s, THEN are given here. Therefore, one can design the fault-tolerant con-
troller (65):
x̄˙ (t) = A1 x̄(t) + B1 u(t), y(t) = C x̄(t)
P = 1.0e + 005
where Ai and Bi , i = 1, 2, . . . , 8, can be easily obtained by the ⎡ ⎤
3.4852 −0.0000 0.0000 0.0000 0.0000
substitution of each of the eight operating points to f (x) and
⎢ −0.0000 3.4852 0.0000 0.0000 −0.0000 ⎥
g(x). In this study, we assume that only an actuator is faulty at ⎢ ⎥
∗ ⎢ 0.0000 0.0000 3.4852 0.0000 0.0000 ⎥
one time. We consider the following case. ⎣ ⎦
0.0000 0.0000 0.0000 3.4852 −0.0000
Case 1:
0.0000 −0.0000 0.0000 −0.0000 3.4852
uf1 (t) = u1 (t) Q1 = 1.0e + 005
662 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 20, NO. 4, AUGUST 2012

Fig. 1. Observer errors time responses: e1 , e2 , e3 (healthy case).

Fig. 3. Fault-detection residuals J1 , J2 with threshold T I = 0.1.

Fig. 2. Fault-detection residual J with threshold T d = 0.1.

⎡ ⎤
3.4852 −0.0000 −0.0000 0.0001 −0.0006
⎢ 0.0000 3.4852 −0.0000 −0.0000 0.0000 ⎥ Fig. 4. Time responses of the observer errors: e1 , e2 , e3 (no compensation
⎢ ⎥ for fault).
∗ ⎢ 0.0000 0.0000 3.4852 −0.0000 0.0000 ⎥
⎣ ⎦
0.0001 0.0000 0.0000 3.4852 0.0001
0.0006 −0.0000 −0.0000 −0.0001 3.4852
  lected as Td = TI = 0. However, there may exist noise and dis-
9.4165 44487.8491 0.8575 181.5760 1.6392
K1 = turbance in practical situation. In the simulations, a white noise,
5.6423 18484.9800 −0.5165 85.7563 0.7744
with zero mean and standard deviation which is equal to 0.1, is
L1 = 1.0e + 008∗ added on each output. As a result, the detection threshold Td and
⎡ ⎤
−0.0035 −0.1100 −0.0003 0.0354 0.0003 the isolation threshold TI can be selected as Td = 0.1, TI = 0.1
⎢−0.1100 −0.0035 6.9356 0.0000 0.0000 ⎥ according to the definition of detection residual and isolation
⎢ ⎥
⎢−0.0003 6.9356 −0.0035 −0.0000 −0.0000 ⎥ residuals. Fig. 2 shows that when an actuator fault occurs in
⎣ ⎦
0.0354 0.0000 0.0000 −0.0035 −0.7706 the system, an alarm is generated since the residual signal de-
0.0003 −0.0000 0.0000 −0.7706 −0.7755 viates significantly from zero. Meanwhile, the SMOs quickly
isolate the fault, as shown in Fig. 3. From Fig. 4, we can see that
The simulation results are presented in Figs. (1)–(9). From when an actuator fault occurs, with no fault compensation, the
Fig. 1, it is easy to see that, under normal operating condition, observation errors do not converge to zero. However, compen-
observation errors globally asymptotically converge to zero. In sating for the fault, the error system becomes stable, as shown in
this paper, it is assumed that the error system is stable before Fig. 5. From Figs. 6 and 7, we can clearly draw the conclusion
−κt/2
 occurrence, namely,ex (0) = 0, ēxs (0) = 0,||ex (0)||e
fault that both gain faults and bias faults can be estimated accurately
∗ λm ax (P )/λm in (P ) = 0. Hence, in the ideal situation, the and promptly. Compared with [9]–[11] and [17], because a clear
detection threshold Td and the isolation threshold TI can be se- definition of threshold for FDI is provided, it is easy to detect
SHEN et al.: FAULT-TOLERANT CONTROL FOR T–S FUZZY SYSTEMS WITH APPLICATION TO NEAR-SPACE HYPERSONIC VEHICLE 663

Fig. 5. Time responses of the observer errors: e1 , e2 , e3 (with compensation

Fig. 8. Fault and its estimation (Case 2).
for fault).

Fig. 6. Gain fault ρ2 (t) = 0.4 sin(πt) and its estimation ρ̂2 (t).
Fig. 9. Fault and its estimation (Case 3).

and isolate the faults. The fault-estimation observer that is pre-

sented in this paper has the following two properties. On the
one hand, differing from the classical fault-estimation schemes
in [9]–[11], in which only bias faults or gain faults can be es-
timated, it is designed to estimate the two types of faults. On
the other hand, from Figs. 8 and 9, it is obvious that it can esti-
mate the types of faults that are considered in [9] and [11], and
the fault-estimation algorithm has better performances. From
the aforementioned simulation results, it can be seen that, by
the proposed FDI observer, an actuator fault can be quickly de-
tected and isolated, and using the fault-estimation algorithm, the
fault can be estimated online, which can be used to compensate
for the fault and to ensure the stability of the closed-loop system
Fig. 7. Bias fault f2 (t) = 0.4 cos(t) and its estimation fˆ2 (t).
in spite of actuator fault.
Remark 6: From the simulation results, it can be seen that 1)
the proposed FDI/FTC scheme is effective because the fault can
664 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 20, NO. 4, AUGUST 2012

be detected, estimated, and accommodated quickly, and 2) the [18] L. Bai, Z. Tian, and S. Shi, “Robust fault detection for a class of nonlinear
performance of our algorithm is better than that presented in the time-delay systems,” J. Franklin Inst., vol. 344, no. 6, pp. 873–888, 2007.
[19] X. He, Z. D. Wang, and D. H. Zhou, “Robust fault detection for net-
literature. worked systems with communication delay and data missing,” Automat-
ica, vol. 45, pp. 2634–2639, 2009.
V. CONCLUSION [20] Y. H. Liu, Z. D. Wang, and W. Wang, “Reliable H ∞ filtering for dis-
crete time-delay systems with randomly occurred nonlinearities via delay-
In this paper, the problem of FTC for T–S fuzzy systems with partitioning method,” Signal Process., vol. 91, pp. 713–727, 2011.
[21] J. X. Dong and G. H. Yang, “Robust static output feedback control for lin-
actuator faults has been studied. We first design a bank of SMOs ear discrete-time systems with time-varying uncertainties,” Syst. Control
to detect and estimate the fault, and a sufficient condition for Lett., vol. 57, pp. 123–131, 2008.
the existence of SMOs is derived. Simulation results of NSHV [22] Y. Wang, D. Zhou, S. J. Qin, and H. Wang, “Active fault-tolerant control
for a class of nonlinear systems with actuator faults,” Int. J. Control,
show that the designed fault detection, isolation, and estimation Autom., Syst., vol. 6, no. 3, pp. 339–350, 2008.
algorithms and FTC scheme have good dynamic performances [23] S. Li and G. Tao, “Feedback based adaptive compensation of control
in the presence of actuator faults. system actuator uncertainties,” Automatica, vol. 45, pp. 393–404, 2009.
[24] S. S. Zhou, J. Lam, and W. X. Zheng, “Control design for fuzzy systems
based on relaxed nonquadratic stability and H ∞ performance conditions,”
IEEE Trans. Fuzzy Syst., vol. 15, no. 2, pp. 188–199, Apr. 2007.
[25] S. K. Nguang and P. Shi, “H∞ fuzzy output feedback control design for
REFERENCES nonlinear systems: An LMI approach,” IEEE Trans. Fuzzy Syst., vol. 11,
[1] Z. Gao, B. Jiang, R. Qi, Y. Xu, and Y. Cheng, “Fuzzy observer design for no. 3, pp. 331–340, Feb. 2003.
[26] H. G. Zhang, S. X. Lun, and D. R. Liu, “Fuzzy H∞ filter design for a
near space vehicle with application to sensor fault estimation,” ICIC Exp.
class of nonlinear discrete-time systems with multiple time delays,” IEEE
Lett, vol. 4, no. 1, pp. 177–182, 2010.
Trans. Fuzzy Syst., vol. 15, no. 3, pp. 453–469, Jun. 2007.
[2] Y. Cheng, B. Jiang, Y. Fu, and Z. Gao, “Robust observer based reli-
able control for satellite attitude control systems with sensor faults,” [27] H. J. Gao, Y. Zhao, and T. W. Chen, “H∞ fuzzy control of nonlinear
systems under unreliable communication links,” IEEE Trans. Fuzzy Syst.,
Int. J. Innovative Comput. Inf. Control, vol. 7, no. 7B, pp. 4149–4160,
vol. 17, no. 2, pp. 265–278, Apr. 2009.
2011.
[28] T. Takagi and M. Sugeno, “Fuzzy identification of systems and its appli-
[3] R. Youssef and P. Hui, “Piecewise sliding mode decoupling fault tolerant
control system,” ICIC Exp. Lett., vol. 4, no. 4, pp. 1215–1222, 2010. cations to modeling and control,” IEEE Trans. Syst., Man, Cybern. B,
Cybern., vol. SMCB-15, no. 1, pp. 116–132, Feb. 1985.
[4] M. Mahmoud, “Stabilizing controllers for a class of discrete time fault
[29] H. J. Xu, M. D. Mirmirani, and P. A Ioannou, “Adaptive sliding mode
tolerant control systems,” ICIC Exp. Lett., vol. 2, no. 3, pp. 213–218,
control design for a hypersonic flight vehicle,” J. Guidance, Control,
2008.
[5] Y. Shtessel and J. McDuffie, “Sliding mode control of the X-33 vehicle Dyn., vol. 27, no. 5, pp. 829–838, 2004.
[30] H. Dong, Z. Wang, W. C. H. Daniel, and H. Gao, “Robust H ∞ fuzzy
in launch and re-entry modes,” in Proc. Amer. Inst. Aeronaut. Astronaut.
output-feedback control with multiple probabilistic delays and multiple
Guidance, Navig., Control Conf., 1998, pp. 1352–1362.
missing measurements,” IEEE Trans. Fuzzy Syst., vol. 18, no. 4, pp. 712–
[6] Y. J. Xu, “Nonlinear robust stochastic control for unmanned aerial ve-
725, Aug. 2010.
hicles,” AIAA J. Guid., Control, Dyn., vol. 32, no. 4, pp. 1308–1319,
[31] H. Dong, Z. Wang, and H. Gao, “H∞ fuzzy control for systems with
2009.
repeated scalar nonlinearities and random packet losses,” IEEE Trans.
[7] H. J. Xu, M. D. Mirmirani, and P. A. Ioannou, “Adaptive sliding mode
control design for a hypersonic flight vehicle,” AIAA J. Guid., Control, Fuzzy Syst., vol. 17, no. 2, pp. 440–450, Apr. 2009.
[32] J. Zhang, P. Shi, and Y. Xia, “Robust adaptive sliding-mode control for
Dyn., vol. 27, no. 5, pp. 829–838, 2004.
fuzzy systems with mismatched uncertainties,” IEEE Trans. Fuzzy Syst.,
[8] W. R. van Soest, Q. P. Chu, and J. A. Mulder, “Combined feedback
vol. 18, no. 4, pp. 700–711, Aug. 2010.
linearization and constrained model predictive control for entry flight,”
AIAA J. Guid., Control, Dyn., vol. 29, no. 2, pp. 427–434, 2006. [33] K. Zhang, B. Jiang, and M. Staroswiecki, “Dynamic output feedback-
fault tolerant controller design for Takagi–Sugeno fuzzy systems with
[9] Z. Gao, B. Jiang, P. Shi, and Y. Xu, “Fault accommodation for near
actuator faults,” IEEE Trans. Fuzzy Syst., vol. 18, no. 1, pp. 194–201, Feb.
space vehicle attitude dynamics via T–S fuzzy models,” Int. J. Innovative
2010.
Comput., Inf. Control, vol. 6, no. 11, pp. 4843–4856, 2010.
[10] Y. Xu, B. Jiang, G. Tao, and Z. Gao, “Fault accommodation for near space [34] J. Dong and G. H. Yang, “Control synthesis of TS fuzzy systems based on
a new control scheme,” IEEE Trans. Fuzzy Syst., vol. 19, no. 2, pp. 323–
hypersonic vehicle with actuator fault,” Int. J. Innovative Comput., Inf.
338, Apr. 2011.
Control, vol. 7, no. 5, pp. 1054–1063, 2011.
[35] H. Zhang and X. Xie, “Relaxed stability conditions for continuous-time
[11] B. Jiang, Z. Gao, P. Shi, and Y. Xu, “Adaptive fault-tolerant tracking
control of near-space vehicle using Takagi–Sugeno fuzzy models,” IEEE TS fuzzy-control systems via augmented multi-indexed matrix approach,”
IEEE Trans. Fuzzy Syst., vol. 19, no. 3, pp. 478–492, Jun. 2011.
Trans. Fuzzy Syst., vol. 18, no. 5, pp. 1000–1007, Oct. 2010.
[36] B. Ding, “Dynamic output feedback predictive control for nonlinear sys-
[12] Q. Zhao and J. Jiang, “Reliable state feedback control systems design
against actuator failures,” Automatica, vol. 34, no. 10, pp. 1267–1272, tems represented by a Takagi–Sugeno model,” IEEE Trans. Fuzzy Syst.,
vol. 19, no. 5, pp. 831–843, Oct. 2011.
1998.
[37] H. K. Lam and M. Narimani, “Quadratic-stability analysis of fuzzy-model-
[13] X. Tang, G. Tao, and S. M. Joshi, “Adaptive output feedback actuator
based control systems using staircase membership functions,” IEEE
failure compensation for a class of non-linear systems,” Int. J. Adaptive
Control Signal Process., vol. 19, no. 6, pp. 419–444, 2005. Trans. Fuzzy Syst., vol. 18, no. 1, pp. 125–137, Feb. 2010.
[38] H. K. Lam, “LMI-based stability analysis for fuzzy-model-based control
[14] S. K. Nguang, P. Shi, and S. Ding, “Fault detection for uncertain fuzzy
systems using artificial TS fuzzy model,” IEEE Trans. Fuzzy Syst., vol. 19,
systems: An LMI approach,” IEEE Trans. Fuzzy Syst., vol. 15, no. 6,
no. 3, pp. 505–513, Jun. 2011.
pp. 1251–1262, Dec. 2007.
[15] Y. Zhao, J. Lam, and H. Gao, “Fault detection for fuzzy systems with [39] C. Peng and T. C. Yang, “Communication-delay-distribution-dependent
networked control for a class of T–S fuzzy systems,” IEEE Trans. Fuzzy
intermittent measurements,” IEEE Trans. Fuzzy Syst., vol. 17, no. 2,
Syst., vol. 18, no. 2, pp. 326–335, Apr. 2010.
pp. 398–410, Apr. 2009.
[40] J. An, G. Wen, C. Lin, and R. Li, “New results on a delay-derivative-
[16] B. Jiang, M. Staroswiecki, and V. Cocquempot, “Fault accommodation
for nonlinear dynamic systems,” IEEE Trans. Autom. Control, vol. 51, dependent fuzzy H∞ filter design for T–S fuzzy systems,” IEEE Trans.
Fuzzy Syst., vol. 19, no. 4, pp. 770–779, Aug. 2011.
no. 9, pp. 1578–1583, Sep. 2006.
[41] D. H. Lee, J. B. Park, and Y. H. Joo, “A new fuzzy Lyapunov function
[17] B. Jiang, K. Zhang, and P. Shi, “Less conservative criteria for fault ac-
for relaxed stability condition of continuous-time Takagi–Sugeno fuzzy
commodation of time-varying delay systems using adaptive fault diagno-
systems,” IEEE Trans. Fuzzy Syst., vol. 19, no. 4, pp. 785–791, Aug.
sis observer,” Int. J. Adaptive Control Signal Process., vol. 24, no. 4,
2011.
pp. 322–334, Apr. 2010.
SHEN et al.: FAULT-TOLERANT CONTROL FOR T–S FUZZY SYSTEMS WITH APPLICATION TO NEAR-SPACE HYPERSONIC VEHICLE 665

Qikun Shen was born in 1971. He received the B.S. Vincent Cocquempot received the Ph.D. degree in
degree in computer science and applications from automatic control from the Lille University of Sci-
the Chinese University of Mining and Technology, ences and Technologies, Villeneuve d’Ascq, France,
Xuzhou, China, in 1996 and the M.S. degree in com- in 1993.
puter science and applications from Yangzhou Uni- He is currently a Full Professor of automatic con-
versity, Yangzhou, China, in 2007. He is currently trol and computer science with the Institut Universi-
pursuing the Ph.D. degree with the College of Au- taire de Technologies de Lille, France. He is a Re-
tomation Engineering, Nanjing University of Aero- searcher with the LAGIS FRE CNRS, 3303: Auto-
nautics and Astronautics, Nanjing, China. matic Control, Computer Science, and Signal Pro-
He is currently an Associate Professor with the cessing Laboratory, Lille 1 University, where he is
College of Information Engineering, Yangzhou Uni- the Head Fault-Tolerant Systems team. His research
versity, Yangzhou, China. His research interests include fault-tolerant control, interests include robust online fault detection and isolation for uncertain dynam-
adaptive control, fuzzy control, intelligent control, etc. ical nonlinear systems and fault-tolerant control for hybrid dynamical systems.

Bin Jiang (SM’05) was born in Jiangxi, China, in

1966. He received the Ph.D. degree in automatic con-
trol from Northeastern University, Shenyang, China,
in 1995.
He has been a Postdoctoral Fellow or a Research
Fellow in Singapore, France, and the United States.
He is a Professor and Vice Dean of the College
of Automation Engineering, Nanjing University of
Aeronautics and Astronautics, Nanjing, China. His
research interests include fault diagnosis and fault-
tolerant control and their applications.
Dr. Jiang currently serves as an Associate Editor or an Editorial Board Mem-
ber for a number of journals, such as the IEEE TRANSACTIONS ON CONTROL
SYSTEMS TECHNOLOGY, the International Journal of Systems Science, the Inter-
national Journal of Control, Automation and Systems, the International Journal
of Innovative Computing, Information and Control, the International Journal of
Applied Mathematics and Computer Science, Acta Automatica Sinica, and the
Journal of Astronautics. He is a member of International Federation of Auto-
matic Control Technical Committee on Fault Detection, Supervision, and Safety
of Technical Processes.