1382 Z. LIU ET AL.
2. System description and preliminaries
Consider the following n-dimensional NSS described by
⎧
⎪
⎪
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎪
⎪
⎩
d[x(t) − Dx(t − τ)] ={(A + A(t))x(t)
+(A
d
+ A
d
(t))x(t − d(t))
+B[u(t) + f (t, x(t ))]
+Gv (t )}dt + g(t, x(t))dω(t ),
y(t) = Cx(t),
x(θ ) = φ(θ),θ ∈ [−h, 0]
(3)
where x(t ) ∈ R
n
is the state vector, u(t) ∈ R
m
is the control
input, y(t) ∈ R
q
is the system output. τ>0isaconstantneutral-
term time-delay, d(t) is the time-varying delay, which satises
0 < d(t) d and
˙
d(t) ≤ μ<1, where d and μ are constants,
and h = max{τ, d}. ω(t) is a standard scalar Brownian motion
denedonacompletedprobabilityspace(, F , {F
t
}
t≥0
, P )
with a natural ltration {F
t
}
t≥0
, and satises E {dω(t )}=0,
E {dω
2
(t)}=dt. v(t) R
l
represents a set of exogenous distur-
bance which belongs to L
2
[0, ∞). φ(t) ∈ C
b
F
0
([−h, 0];R
n
) is
the initial condition. A, A
d
, B, C, D and G are known real matri-
ces, B is of full column rank, and the spectrum radius of the
matrix D,i.e.ρ(D), satises ρ(D) < 1. To facilitate the result,
the following preliminaries are introduced for system (3).
Assumption 2.1: The structural uncertainties A(t) and A
d
(t)
are norm bounded, i.e. [A(t)A
d
(t)] = EJ(t)[FF
d
],whereE,F
and F
d
are constant matrices, and J(t) is unknown matrix function
satisfying J
T
(t)J(t) Iforallt 0.
Assumption 2.2: (Li et al., 2016;Yaoetal.,2015) f(t, x) is
unknown nonlinearity which represents the lumped perturbation
of a physical plant through the control channel satisfying f(t, x)
αy(t),whereα>0 is an unknown constant.
Assumption 2.3: The diusion gain function g(t, x) may not be
exactly known but there exists a matrix M such that the inequality
holds: Tr{g
T
(t, x)g(t, x)} My(t)
2
.
Remark 2.1: Assumption 2.3 is reasonable to a certain degree.
Actually, the state-dependent stochastic noises g(t, x)maynot
be accessible but could be evaluated by Tr{g
T
(t, x)g(t, x)}
Nx(t)
2
,whereN isaconstantmatrix(seeHuang&Mao,2010;
Kao et al., 2014, 2015). With the relevance that y(t ) = Cx(t ),i.e.
x(t) = C
+
y(t), the Assumption 2.3 is easily introduced herein,
where C
+
denotes the Moore-Penrose inverse of C. Consider
that x(t) may be unmeasured, the output information is used to
facilitate the control design. As to C, its selection may be exible
bytheactualdesign,e.g.itcanbeoffullcolumnrank.
Based on the above conditions, one can verify that the
stochastic neutral system (3) with u(t) = 0hasauniquesolu-
tion according to Huang and Mao (2009)andMao(2007). In
fact, denote the following terms:
m(t) = (A + A(t))x(t ) + (A
d
+ A
d
(t))x(t − d(t))
+Bf(t, x) +Gv (t ), n(t) = g(t, x)
for all t 0. It is easy to observe that m(t)
2
K
m
x
t
2
, n(t)
2
K
n
x
t
2
,wherex
t
= {x(t + θ ): −h θ 0}, K
m
and K
n
are
positive and can be found upon the premise of each component
of m(t)andn(t), respectively. Thus, this implies m(t)andn(t)
satisfy the local Lipschitz condition and the linear growth con-
dition. At this point, there exists a unique continuous solution
expressed by x(t; φ) to the NSS (3), and the details can refer to
Theorem 3.1 of Mao (2007).
Denition 2.1 (Chen et al., 2010): System(3)issaidtobemean-
square exponentially stable if there exist scalars η>0, β>0such
that E {x(t)
2
}≤ηe
−βt
sup
−h≤θ≤0
E {φ(θ)
2
} for all admissi-
ble uncertainties.
Lemma 2.1 (Huang & Mao, 2010): For a pair of constant matri-
ces G ∈ R
p×p
and M ∈ R
p×q
,ifG 0,thenTr(M
T
GM)
λ
max
(G)Tr(M
T
M).
3. Design of the sliding-mode observer
In this section, we focus on the stability analysis and con-
troller synthesis of the closed-loop systems based on a new state
observer design. The research framework and basic contents are
as follows, respectively.
3.1. Non-fragile state observer and novel sliding-surface
design
First, the state observer technique is utilised to generate the
accurate state estimation of system (3). Here, the following non-
fragile observer is employed for the design
⎧
⎪
⎪
⎨
⎪
⎪
⎩
d[
ˆ
x(t) − D
ˆ
x(t − τ)] ={A
ˆ
x(t) + A
d
ˆ
x(t − d(t))
+(L + L(t))(y(t) −C
ˆ
x(t))}dt,
ˆ
y(t) = C
ˆ
x(t),
ˆ
x(θ ) =
ˆ
φ(θ),θ ∈ [−h, 0]
(4)
where
ˆ
x(t) ∈ R
n
represents the estimation of x(t),
ˆ
y(t) denotes
the output of the observer. L ∈ R
n×p
is the observer gain to be
designed later, and L(t) is an additive gain variation satisfying
L(t) δ,whereδ>0 is a constant, namely, the observer
may be aected by some perturbations (Kao et al., 2014, 2015).
Let the estimation error be e(t ) = x(t ) −
ˆ
x(t).Thus,bysub-
tracting (4) from (3), one can get the following estimation error
system:
⎧
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎩
d[e(t) − De(t − τ)] ={[A − LC − L(t)C]e(t)
+A
d
e(t − d(t )) + A(t)x(t )
+A
d
(t)x(t − d(t)) + B[u(t) + f (t, x(t ))]
+Gv (t )}dt + g(t, x(t))dω(t ),
y
e
(t) = Ce(t )
(5)
where y
e
(t) denotes the output of the error system.
Remark 3.1: In this work, a particular non-fragile observer is
proposed for the system. Dierent from previous designs of the
observer (see Elhsoumi et al., 2016;Kaoetal.,2014;Kaoetal.,
2015;Li&Li,2009;Lietal.,2016;Lin,Wang,Lee,He,&Chen,
2008;Liuetal.,2015; Qiao, Zhang, Zhu, & Zhang, 2009;Rahme
et al., 2015;Shietal.,2015;Wuetal.,2008;Yanetal.,2010;Yao
et al., 2015;Zhangetal.,2016 for details), it is worth noting that
the control input and/or its compensator are required to involve
into their observer designs; however, only an observer gain L is
to be determined here. Thus, this simplied the procedure of the
state observer design, which is the rst advantage of this paper.