Stability analysis of recurrent neural networks with interval
time-varying delay via free-matrix-based integral inequality
$
Wen-Juan Lin
a
, Yong He
a
, Chuan-Ke Zhang
a,b,
n
, Min Wu
a
, Meng-Di Ji
c
a
School of Automation, China University of Geosciences, Wuhan 430074, China
b
Department of Electrical Engineering & Electronics, University of Liverpool, Liverpool L69 3GJ, United Kingdom
c
School of Information Science and Engineering, Central South University, Changsha 410083, China
article info
Article history:
Received 11 December 2015
Received in revised form
3 March 2016
Accepted 12 April 2016
Communicated by Hongyi Li
Available online 13 May 2016
Keywords:
Recurrent neural networks
Interval time-varying delay
Stability
Augmented Lyapunov–Krasovskii functional
Free-matrix-based integral inequality
abstract
This paper is concerned with the stability analysis of recurrent neural networks with an interval time-
varying delay. A new Lyapunov–Krasovskii functional (LKF) containing some augmented double integral and
triple integral terms is constructed, in which the information of the activation function and the lower bound
of the delay are both fully considered. Then, a free-matrix-based integral inequality is employed to deal with
the derivative of the LKF such that an improved stability criterion is derived. Finally, two numerical
examples are provided to illustrate the effectiveness and the benefit of the proposed stability criterion.
& 2016 Elsevier B.V. All rights reserved.
1. Introduction
In recent years, recurrent neural networks (RNNs) have been
widely used in various fields due to their extensive applications
such as pattern recognition, signal processing, associative mem-
ories and other scientific areas [1,2]. And the key of these appli-
cations with RNNs is that the equilibrium points of the designed
network are stable. As a result, stability analysis of RNNs plays an
important role. On the other hand, as we all know, a time delay
which inevitably exists in many RNNs is usually a cause of oscil-
lation and instability. Therefore, the stability problem of the
delayed neural networks (DNNs) is of a great deal of importance in
both theory and practice, and also has attracted much attention
[4–22]. In the field of stability analysis, many results on this topic
can be classified into the delay-dependent one and delay-
independent one. Since the former considers more information
of the delay and is usually less conservative, much attention has
been put into employing some less conservative delay-dependent
stability conditions.
In delay-dependent stability analysis of system with interval
time-varying delay, we assume that there exists an upper bound of
the delay. When the delay is in the interval from the given lower
bound to the upper bound of delay, the delay system is asymptotic
stable. Based on Lyapunov theory, constructing a suitable Lyapu-
nov–Krasovskii functional (LKF) and estimating its derivative are
two key points to enhance the feasible regions of stability criteria
and reduce the conservatism. For the construction of the LKF, the
simple LKF was firstly employed in DNNs, and rich results are
gotten by that [4–6]. However, as we all know, the stability criteria
are conservative by using the simple LKF since the delay infor-
mation was not taken fully into account. To deal with this problem,
the augmented LKF method was proposed [3] and widely used to
the stability analysis problem of DNNs [7–21]. Furthermore, to
reduce the conservatism, using the delay-decomposition idea [7],
considering more information of the activation functions [8–10],
augmenting the double integral terms [10,11], and introducing the
triple integral terms [11] were employed to construct the LKF.
For estimating the derivative of the LKF, the free weighting
matrix (FWM) approach [12,13], its improved forms [14,15] and
the integral inequality method such as Jensen's inequality [13,15],
Wirtinger-based integral inequality [16,17] are the most popular
methods reported in the literature. The FWM method is once the
best one because of neither the model transformation nor the
cross-term bounding being required. However, with the proposal
of the convex combination approach, which successfully avoids
replacing the delay by its lower or upper bound directly, integral
Contents lists available at ScienceDirect
journal homepage: www.elsevier.com/locate/neucom
Neurocomputing
http://dx.doi.org/10.1016/j.neucom.2016.04.052
0925-2312/& 2016 Elsevier B.V. All rights reserved.
☆
This work was supported in part by the National Natural Science Foundation of
China under Grants 61573325, 61503351, 61210011, and the Hubei Provincial
Natural Science Foundation of China under Grant 2015CFA010.
n
Corresponding author at: School of Automation, China University of Geos-
ciences, Wuhan 430074, China.
E-mail address: ckzhang@cug.edu.cn (C.-K. Zhang).
Neurocomputing 205 (2016) 490–497