978-1-5090-0022-7/15/$31.00 ©2015 IEEE 457
2015 8th International Conference on BioMedical Engineering and Informatics (BMEI 2015)
New Approach to Stability Analysis for Generalized
Neural Networks with Time-varying Delays
Saibing Qiu
†‡
, Xin-Ge Liu
†∗
, Mei-Lan Tang
†
, Feng-Xian Wang
†
and Yan-Jun Shu
†
†
School of Mathematics and Statistics, Central South University, Changsha 410083, China
‡
College of Mathematics and Computer Science, Hunan City University Yiyang, Hunan 413000, China
∗
Correspondence: liuxgjiayou@126.com
Abstract—In this paper, the problem of delay-dependent stabil-
ity for a class of generalized neural networks with time-varying
delays is investigated. By introducing two triple integral terms
and a new augmented term with more delay information to
the Lyapunov-Krasovskii functional(LKF), some less conserva-
tive delay-dependent stability criteria are derived in terms of
linear matrix inequalities (LMIs). Finally, the effectiveness of
the proposed criteria is shown via two numerical examples by
comparison the delay upper bounds with the existing ones.
Keywords-neural networks; asymptotical stability; Wirtinger in-
equality; LMIs
I. INTRODUCTION
In the past decades, the neural networks have been a
hot research topic and investigated extensively owing to
their successful application in various fields such as signal
processing, power systems, pattern recognition, associative
memories and other scientific areas (see[1-4]). In practice,
during the implementation of artificial neural networks, time
delays are frequently occur since the finite switching speed of
amplifiers. It is well known that delays may bring oscillation
even instability. Therefore, the neural networks with time-
varying delay have been widely studied in recent years, many
interesting results have been derived (see[5-8]). Generally
speaking, the stability criteria of neural networks can be
classified into delay-independent stability and delay-dependent
ones. The delay-dependent stability criteria have been paid
more attention, since they are generally less conservative than
delay-independent ones.
The main aim of delay-dependent stability criteria is to get
maximum delay bounds such that the designed networks are
asymptotically stable. In order to reduce the conservatism of
the delay-dependent stability criteria, some new techniques
were introduced to construct suitable Lyapunov-Krasovskii
functional(LKF) and estimate its time derivative. Such as,
triple integral terms [9], [10], delay-partitioning approach [11],
[12], [13], free-weighting matrices [13], [14], the augmented
vector [15], [16], [17], Jensen’s inequality [18], [19], [20],
Park’s inequality[21], and the reciprocally convex approach
[22]. Recently, a Wirtinger-based integral inequality was re-
ported in [23], which was applied to get some new stability
criteria for various time delay systems, and the advantages
were shown via the comparison of maximum delay bounds
for various systems. Very recently, Zeng et al.[24] introduced a
new free-matrix-based integral inequality, which encompassed
the Wirtinger-based inequality as a special case. The free
matrices provided freedom in reducing the conservativeness of
the inequality. But the introduction of free matrices increases
greatly computational burden. Therefore, there are rooms for
further improvement in stability analysis of neural networks
with time-varying delays, which motivated this research.
Inspired by the work mentioned above, in this paper, the
problem of delay-dependent stability is investigated for a
class of generalized neural networks with time-varying delays,
which includes static neural networks and local field networks
as their special cases. Some new delay-dependent stability
criteria are derived in terms LMIs. Finally, two numerical
examples are given to demonstrate the effectiveness of the
obtained criteria. The main contribution of this paper lies in
tree aspects.
(1) A new augmented term
t
t−h(t)
η
T
3
(t, s)Mη
3
(t, s)ds
taking account of delay information in a new way will be
introduced to the LKF, which has not been considered yet in
stability analysis of neural networks with time-varying delays
play an important role to reduce the conservatism.
(2) In the process of the handing some integral terms in the
derivative of the proposed LKF, in [25], the term
h
M
−h(t)
h(t)
is
approximated with
h
M
−h(t)
h
M
, where h(t) ∈ [0, h
M
]. In [26],
the terms
h
2
−h(t)
h(t)−h
1
and
h(t)−h
1
h
2
−h(t)
are approximated with
h
2
−h(t)
h
2
−h
1
and
h(t)−h
1
h
2
−h
1
respectively, where h(t) ∈ [h
1
, h
2
]. These may
be the cause of the conservatism. In this paper, by making
full use of Jensen’s inequality, Wirtinger-based inequality and
the reciprocally convex approach, all the integral terms in the
derivative of the proposed LKF are effectively handled without
introduce any conservative approximation.
(3) More information of neuron activation functions is uti-
lized in the proposed LKF. And Jensen’s inequality, Wirtinger-
based inequality technique are adopted to estimate the upper
bound of the derivative of LKF.
II. PROBLEM STATEMENT AND PRELIMINARIES
Consider the following generalized delayed neural network
˙x(t) = −Ax(t) + W
0
g(W
2
x(t)) + W
1
g(W
2
x(t − h(t))) (1)
where x(t) = [x
1
(t), x
2
(t), ..., x
n
(t)]
T
∈ R
n
is the neu-
ron state vector. A = diag{a
1
, a
2
, ..., a
n
} > 0, W
0
,
W
1
and W
2
are the connection weight matrices. g(.) =
[g
1
(.), g
2
(.), ..., g
n
(.)]
T
∈ R
n
denotes the neuron activation