analysis, controller and filter design for switched systems
with time-delay. See e.g. [18–21].
Fault detection and isolation (FDI), on the other
hand, has been an active field of research over the past
decades because of the increasing demand for higher
performance, higher safety and reliability standards.
Fruitful theoretical results and increasing applications
in industrial practice, such as power plant, coal mills
and robotic systems can be found in [22–24]. Further-
more, with the rapid development of communication
networks, a great amount of efforts have recently
been devoted to the problems of an FDI for networked
control systems [25,26]. For example, the authors
in [25] have considered the fault detection problem
for a class of discrete-time networked systems with
multiple state delays and an unknown input. A new
measurement model has been proposed to account for
both the random measurement delays and the stochastic
data missing (package dropout) phenomenon. As has been
discussed before, the switched linear system is a powerful
method to model the NCSs. Therefore, it is not only
theoretically interesting but also practically important to
study the FDI for the switched linear systems with delay
or without delay. The pioneer of this topic has been
published in [27], where the switched Lyapunov func-
tional method is used for the FDI of discrete-time
switched systems with state delays, and sufficient condi-
tions for the solvability of the problem are established in
terms of LMIs. However, it should be pointed out that only
constant delays are considered. In addition, the obtained
results are delay-independent. To the best of the authors’
knowledge, the delay-dependent fault detection for
switched systems with time-varying delays has not yet
been fully investigated, and it is our intention in this
paper to shorten such a gap.
In this paper, the fault detection problem is investi-
gated for a class of discrete-time switched linear systems
with time-varying state delays. The fault detection
problem addressed is firstly converted into an auxiliary
H
N
filtering problem. Then, the Lyapunov functional
method and the average dwell time approach are
proposed for the analysis and synthesis of the considered
systems. A delay-dependent sufficient condition for the
existence of the desired fault detection filters is derived
and formulated in terms of LMIs. A numerical example is
finally given to show the effectiveness of the proposed
method.
Notations: the notation used throughout the paper is
fairly standard. We use W
T
, W
1
,
l
ðWÞ, TrðWÞ and 99W99
to denote, respectively, the transpose, the inverse, the
eigenvalues, the trace and the induced norm of any square
matrix W. We use W40 to denote a positive-definite
matrix W with
l
min
(W) and
l
max
(W) being the minimum
and maximum eigenvalues of W and I to denote the
identity matrix with an appropriate dimension. Let R
n
denote the n dimensional Euclidean space. R
m n
is the set
of all m n real matrices. The notation l
2
[0,N] refers to
the space of square summable infinite vector sequences
with the usual norm 99 99
2
. The symbol
n
will be used in
some matrix expressions to represent the symmetric
terms.
2. Problem formulation
Consider the following discrete-time switched linear
systems with time-varying state delays
xðkþ 1Þ¼A
s
ðkÞ
xðkÞþA
d
s
ðkÞ
xðkhðkÞÞþ E
s
ðkÞ
uðkÞþB
s
ðkÞ
dðkÞ
þG
s
ðkÞ
f ðkÞyðkÞ
¼ C
s
ðkÞ
xðkÞþC
d
s
ðkÞ
xðkhðkÞÞþ Q
s
ðkÞ
uðkÞþD
s
ðkÞ
dðkÞ
þJ
s
ðkÞ
f ðkÞxðkÞ
¼
j
ðkÞ, 8k 2½d
2
, 0ð1Þ
where xðkÞ2R
n
is the state, yðkÞ2R
r
is the measured
output, dðkÞ2R
p
, uðkÞ2R
s
and fðkÞ2R
q
are the unknown
input, control input, and fault, respectively, which belong
to l
2
[0,N].
j
(k) is a vector-valued initial function.
The time-varying delay h(k) is assumed to satisfy
d
1
r h(k)rd
2
.
s
(k):Z-M={1,2, y, m} is the switching
signal, m is a finite integer, and Z is the set of positive
integers. Meanwhile, for switching time sequence
k
0
o k
1
o k
2
o y of the switching signal
s
(k), the holding
time between [k
l
,k
l +1
] is called the dwell time of the
currently engaged subsystem, where l is a non-negative
integer. The matrices A
i
, A
di
, B
i
, C
i
, C
di
, D
i
, E
i
, G
i
, J
i
and Q
i
, i={1,2, y, m} are constant with appropriate
dimensions.
An FDI system consists of a residual generator and an
evaluation stage, including an evaluation function and a
threshold. To generate a residual signal, we consider the
following fault detection filter:
^
xðkþ 1Þ¼A
fi
^
xðkÞþB
fi
yðkÞ
rðkÞ¼C
fi
^
xðkÞþD
fi
yðkÞð2Þ
where
^
xðkÞ2R
n
is the filter’s state and rðkÞ2R
q
is the
residual signal. A
fi
, B
fi
, C
fi
and D
fi
are the filter gain
matrices to be determined.
For the purpose of fault detection, it may not be
necessary to estimate the fault f(k). Sometimes one is
more interested in the fault signal in a certain frequency
interval, which can be formulated as the weighted fault
^
f ðzÞ¼W
f
ðzÞf ðzÞ with W
f
being a given stable weighting
matrix [28]. A minimal realization of
^
f ðzÞ¼W
f
ðzÞf ðzÞ is
supposed to be [29]
xðkþ 1Þ¼A
w
xðkÞþB
w
f ðkÞ
^
f ðkÞ¼C
w
xðkÞþD
w
f ðkÞð3Þ
where
xðkÞ2Rn is the state of the weighted fault, f(k)is
the original fault, and
^
f ðkÞ2R
q
is the weighted fault. A
w
,
B
w
, C
w
and D
w
are known constant matrices. In this
paper, it is intended to make the error between the
residual signal r(k) and the weighted fault signal
^
f ðkÞ as
small as possible in an H
N
framework.
Since y(k)=C
i
x(k)+C
di
x(k h(k))+Q
i
u(k)+D
i
d(k)+J
i
f(k),
substituting y(k) into Eq. (2), and augmenting the system
Eqs. (1)–(3), we have
xðkþ1Þ
^
xðkþ1Þ
xðkþ1Þ
2
6
4
3
7
5
D. Zhang et al. / Signal Processing 91 (2011) 832–840 833