Mixed H
2
/H
∞
Control for a Class of Nonlinear Networked Control Systems 657
where x(k) ∈ R
n
is the state vector; u(k) ∈ R
m
is the con-
trol input;
ω
(k) ∈ R
p
is the disturbance signal which be-
longs to l
2
[0,∞); z
1
(k), z
2
(k) ∈ R
p
are the controlled out-
puts. The matrices A,B,N,E,C, D,G and H are real con-
stant matrices of appropriate dimensions.
It is supposed that the nonlinear vector function g (x (k))
satisfies the condition g(0) = 0 and the following condi-
tion for all (x,k) ∈ R
n
× Z:
∥ g(x(k)) ∥≤
β
∥ x(k) ∥, (2)
for some constant
β
> 0, which implies that
g
T
(x(k))g(x(k)) ≤
β
2
x
T
(k)x(k). (3)
Similar to [5], the stochastic packet dropouts process of
the NCSs described by (1) can be modeled as a homoge-
nous Markov chain, which should be much closer to real
cases.
Definition 1: Packet dropouts process of the nonlinear
NCSs is defined as
η
k
= {i
k+1
− i
k
,i
k
∈
ζ
}, (4)
where
ζ
= {i
1
,i
2
,i
3
,...} ⊂ {1, 2, 3,...}, representing time
point sequences of successful data transmission from the
sensor to actuator.
η
k
takes values in the finite set S =
{1,2,... ,
σ
};
σ
denotes the maximum packet dropouts
upper bound.
η
k
= 1 denotes no packet dropout;
η
k
= 2
implies 1 packet dropout between two consecutive suc-
cessful packets transmission, etc.
Definition 2: Given a probability space (Ω,F,P),
stochastic packet dropouts process is said to be a discrete-
time homogeneous Markov chain, which takes values in
the finite set S = {1,2,. . .,
σ
}. Let Λ = [
λ
r j
] ∈ R
σ
×
σ
be
the transition probability matrix, where
λ
r j
≜ Pr{
η
k+1
=
j |
η
k
= r}, 0 ≤
λ
r j
≤ 1,
∑
σ
j=1
λ
r j
= 1,r, j ∈ S.
On the basis of the above two definitions, the system (1)
with stochastic packet dropouts by time-varying sampling
periods
ˆ
T =
η
k
T (T denotes the fixed sampling period)
can be further transformed into
x(k +
η
k
) = A(
η
k
)x(k) +B(
η
k
)u(k) +N(
η
k
)g(x(k))
+ E(
η
k
)
ω
(k),
z
1
(k) = Cx(k) + Du(k),
z
2
(k) = Gx(k) + Hu(k).
(5)
Then, the mode-dependent controller is designed as fol-
lows:
u(k) = K(
η
k
)Πx(k −
τ
k
),K(
η
k
) ∈ R
m×n
,
0 ≤
τ
k
≤
τ
M
,
η
k
∈ S,
(6)
where
τ
k
stands for the network-induced delay from the
sensor to controller, and
τ
M
is the largest upper bound of
the delay. Π = diag{
ρ
1
,
ρ
2
,...,
ρ
n
},
ρ
i
stands for the fault
of each sensor,
ρ
i
(i = 1,2,... , n) are unrelated random
variables which are also unrelated with
ω
(k). It is sup-
posed that the probabilistic density functions F(
ρ
i
) (i =
1,2,... , n) of
ρ
i
take values in the interval [0, 1].
λ
i
and
µ
2
i
(i = 1,2,..., n) denote the mathematical expectation
and variance of
ρ
i
, respectively.
Remark 1: The majority of existing literature have
modeled the sensor fault as the Bernoulli random binary
distributed sequence, in which only two cases are con-
sidered; one is the complete fault, the other is the com-
pletely normal [16]. Meanwhile, the fault of each sensor
is supposed to be unrelated with another. Especially, [18]
has established a novel fault model, which has four cases:
outage, loss of effectiveness, stuck and normal. Though
the model of [18] is much more closer to real situation,
we suppose that the fault has three cases: complete fault
(
ρ
i
= 0), partly lost of effectiveness (0 <
ρ
i
< 1) and nor-
mal (
ρ
i
= 1) in this paper, which will be readily dealt with
in the later parts.
Consequently, the closed-loop system of the consid-
ered nonlinear NCSs with the effect of random network-
induced delays, stochastic packet dropouts and sensor
faults are formulated as follows
x(k +
η
k
) = A
i
x(k) + B
i
K
i
(Π −
¯
Π)x(k −
τ
k
)
+ B
i
K
i
¯
Πx(k −
τ
k
) + N
i
g(x(k)) +E
i
ω
(k),
z
1
(k) = Cx(k) + DK
i
(Π −
¯
Π)x(k −
τ
k
) + DK
i
¯
Πx(k −
τ
k
),
z
2
(k) = Gx(k) + HK
i
(Π −
¯
Π)x(k −
τ
k
) + HK
i
¯
Πx(k −
τ
k
),
x(k) =
ψ
(k), k = −
τ
M
,−
τ
M
+ 1, ...,0,
(7)
where
ψ
(k) is the given initial state of the system, and
¯
Π = E{Π} =
n
∑
j=1
α
j
ϒ
j
,
ϒ
j
= diag{0,...,0
j−1
,1,0,.. . ,0
n− j
},
A
i
= A
η
k
=i
, B
i
= B
η
k
=i
, N
i
= N
η
k
=i
,
E
i
= E
η
k
=i
, K
i
= K
η
k
=i
.
Definition 3 [
31]: The closed-loop system of (7) with
ω
(k) = 0 is said to be stochastically stable in mean square,
if for any initial condition, there exists a scalar c such that
E
∞
∑
k=0
∥x(k)∥
2
< c sup
−
τ
M
≤i≤0
E
∥x(i)∥
2
. (8)
The purpose of this paper is to design a mode-
dependent feedback controller in the form of (6), such
that for the effect of random network-induced delays,
stochastic packet dropouts and sensor faults, the result-
ing closed-loop system (7) is stochastically stable in mean