Delay-dependent stability for neural networks with time-varying
delays via a novel partitioning method
Bin Yang
a, Rui Wang
b,n, Georgi M. Dimirovski
c,d aSchool of Control Science and Engineering, Dalian University of Technology, Dalian 116024, PR China
bSchool of Aeronautics and Astronautics, State Key Laboratory of Structural Analysis of Industrial Equipment, Dalian University of Technology,
Dalian 116024, PR China
c
School of Engineering, Dogus University, Acibadem, TR-34722 Istanbul, Turkey
d
School FEIT, St. Cyril and St. Methodius University, Karpos 2, MK-1000 Skopje, Macedonia
a r t i c l e i n f o
Article history: Received 4 May 2015 Received in revised form 4 August 2015 Accepted 18 August 2015 Communicated by Hongyi Li Available online 1 September 2015 Keywords:
Neural networks Time-varying delay Stability
Lyapunov–Krasovskii functional Wirtinger integral inequality
a b s t r a c t
In this brief, a novel partitioning method for the conditions on bounding the activation function in the stability analysis of neural networks systems with time-varying delays is presented. Certain further improved delay-dependent stability conditions, which are expressed in terms of linear matrix inequal-ities (LMIs), are derived by employing a suitable Lyapunov–Krasovskii functional (LKF) and utilizing the Wirtinger integral inequality. Two well-known examples are investigated in a comparison mode with results to show the effectiveness and improvements achieved by the new results proposed.
& 2015 Elsevier B.V. All rights reserved.
1. Introduction
In recent years, an increasing number of research studies on the neural networks dynamics became apparent. Largely, this trend is due to their many successful applications in areas of pattern recognition, image processing, associative memories[1], optimi-zation problems and even mechanics of structures and materials
[2]. It should be noted that due to thefinite switching speed of electronics involved and the inherent communication time between the neurons, inevitably time-delay exists regardless of how small it may be. Precisely the time-delay is a main factor that can cause performance degradation and/or the instability of neural networks. It is therefore found that the stability problem of neural networks with time delays has attracted considerable attention of many researchers in the last few decades and con-siderable stability research results have emerged. The existing stability criteria may well be grouped into delay-independent and delay-dependent types of criteria. In general, the delay-dependent stability criteria are less conservative than the delay-independent ones. For the delay-dependent stability criteria, the maximum delay bound is an important index for checking and evaluating the
conservatism of the criteria. In turn, rather significant research efforts[3–16]have been devoted to the reduction of conservatism of the delay-dependent stability criteria for neural networks with time delays even when these are fairly small.
As known from Lyapunov stability theory, there are two effective ways to reduce the conservatism in stability analysis. One is the choice of suitable LKF and the other is the estimation of its time derivative. Recently, some new techniques of the construc-tion of a suitable LKF and the estimaconstruc-tion of its derivative for time-delay systems have been presented as seen from[3–39]. Methods for constructing a delicate LKF include delay-partitioning ideas, triple integral terms, augmented vectors, and involving more information of activation functions. Methods for estimating the time-derivative of LKF include P.Park’s inequality, Jensen’s inequality, free-weighing matrices, reciprocally convex optimiza-tion, quadratic convex combination method, and so on. Since Jensen’s inequality to introduce an undesirable conservatism in the stability conditions was noted, Seuret and Gouaisbaut [32]
introduced certain Wirtinger inequalities and overcame that con-servatism. An originally developed free-matrix-based inequality, which encompasses the Wirtinger-based inequality and was more tighter than existing ones, was presented by Zeng [40, 41 ]. The developments of the mentioned methods appeared very useful in the investigation of the stability problems for neural networks with time delays.
Contents lists available atScienceDirect
journal homepage:www.elsevier.com/locate/neucom
Neurocomputing
http://dx.doi.org/10.1016/j.neucom.2015.08.058
0925-2312/& 2015 Elsevier B.V. All rights reserved.
nCorresponding author.
For instance, here we refer to several of the recent develop-ments. Authors of[22,23]used the property offirst order convex combination and derived some less conservative criteria. J. H. Kim proposed quadratic convex function for linear system[24]and this quadratic convex combination method was developed further in
[30]. The approach of free-weighing matrices was used in[25,26]. Authors of [27–28] took a new augmented vector. Delay-partitioning methodology has been used in[25–28]and, in turn, the conservatism of delay-dependent stability criteria reduced. However, as the partitioning number of delay increases, the matrix formulation becomes more complex and the dimension of the stability criterion grows bigger, and thus the computational bur-den and time consumption growth become a tangible problem. In contrast to the delay-partitioning method, recently the approach of activation function dividing was proposed in [13], and new improved delay-dependent criteria for neural networks with time delays were established. Work[35]presented an improved delay-dependent stability criterion for recurrent neural networks with time-varying delays by adopting a more general method of acti-vation function dividing.
In this paper, following the above enlightenment discussion, we present a set of new contributions. Firstly, a new LKF is con-structed by taking more information of state and activation func-tions as augmented vectors, and then reciprocal convex approach and Wirtinger integral inequality is used to handle the integral term of quadratic quantities in the estimation of LKF's derivative. With the new LKF at hand, in Theorem 1, we derive the stability condition in terms of the convex combination with respect to the time-varying delays and its variation, and the delay-dependent stability criterion in which both the upper and lower bounds of delay derivative are available. Secondly, unlike the delay parti-tioning method, in Theorem 2, a more general dividing approach of the bounding conditions on activation function is employed. The bounding of activation functions of neural networks with time-varying delays is built within two subintervals, which can be either equal or unequal, in order to account for reducing the computing time and for improvement of the feasible region. By using the information of new bounding conditions for the two subintervals of activation functions, a new and not in-so-far pro-posed LKF is constructed for the proof of Theorem 3. Thirdly, by utilizing the results of Theorem 3, when only the upper bound of the delay derivative of the time-varying delay is available, Cor-ollary 1 presents the corresponding results. And when the infor-mation about the delay derivative of time-varying delay is unknown, this case can be readily derived from Corollary 1. Finally, our stability analysis method is applied to two well-known examples in the literature and our obtained results are com-pared with the existing corresponding results, respectively, to illustrate its effectiveness and demonstrate the improvements obtained.
Throughout this paper the following notation is used: CTrepresents the transposition of matrixC.ℝn denotes
n-dimen-sional Euclidean space and ℝnm is the set of all n m real
matrices. P40 means that P is a real symmetric positive-definite matrix. Symbol * represents symmetric term in a symmetric matrix and diagf⋯g denotes a block diagonal matrix. SymðXÞis defined as SymðXÞ ¼ X þXT.
2. Problem formulation
Consider the following class of neural networks with discrete time-varying delays:
_zðtÞ ¼ CzðtÞþAf ðzðtÞÞþBf ðzðt hðtÞÞÞþJ ð1Þ
In representation model (1), symbols denote: zðtÞ ¼ ½z1ðtÞ; :::; znðtÞTAℝnis a real-valued n-vector representing the state
variables associated with the neurons in the neural networks; f ðzÞ ¼ ½f1ðz1Þ; :::; fnðznÞTAℝnis n-vector of the neuron activation
functions; J ¼ ½J1; :::; JnTAℝn is a constant input vector; C ¼ diagf
c1; :::; cngAℝnn and A; B are the constant matrices of appropriate
dimensions completing the description of this class of neural networks.
The delay hðtÞ is assumed to be represented by a time-varying continuous function satisfying
C1: 0rhðtÞrhM; hlDr _hðtÞrh u Do1 C2: 0rhðtÞrhM; _hðtÞrhuD where hM40 and hlD; h u
Dare known constants.
The activation functions fiðziðtÞÞ; i ¼ 1; :::; n, are continuous,
bounded, and satisfy the inequalities kirfiðuÞfiðvÞ u v rk þ i ; u; vAℝ; uav; i ¼ 1; :::; n ð2Þ where ki and k þ i are constants.
For simplicity, in the stability analysis of the neural networks
(1), wefirst shift the equilibrium point zto the origin by letting
x ¼ z z; gðxÞ ¼ f ðxþzÞf ðzÞ. Then the network system model
(1)can be converted into
_xðtÞ ¼ CxðtÞþAgðxðtÞÞþBgðxðt hðtÞÞÞ ð3Þ where xðtÞ ¼ ½x1ðtÞ; :::; xnðtÞTAℝn is the state vector of the
trans-formed system, gðxðtÞÞ ¼ ½g1ðx1ðtÞÞ; :::; gnðxnðtÞÞT with gjðxjðtÞÞ ¼ fjðxj
ðtÞþz
jÞfjðzjÞ satisfying gjð0Þ ¼ 0ðj ¼ 1; :::; nÞ. Notice that functions
giðUÞði ¼ 1; :::; nÞ satisfy the following inequality conditions:
kir
giðuÞgiðvÞ
u v rk
þ
i ; u; vAℝ; uav; i ¼ 1; :::; n: ð4Þ
If there is v ¼ 0 in(4), then we have kir
giðuÞ
u rk
þ
i ; 8 ua0; i ¼ 1; :::; n: ð5Þ
The objective of this paper is to explore the analysis of the asymptotic stability of the considered class of neural networks with time-varying delays via utilizing the representation model
(3).
Before deriving the main results, we quote the following lem-mas that are used in the subsequent section.
Lemma 1. [17,32]. For given positive integers n; m, a scalar
α
in the interval ð0; 1Þ, a given n n matrix R40, and two matrices W1and Table 1Delay bounds hMwith different hD
Methods Condition of _hðtÞ hD¼ 0:8 hD¼ 0:9 Unknown or
Z1 Theorem 1[28] 0r _hðtÞrhD 3.0604 1.9956 – Corollary 1[28] – – – 1.7860 Theorem 1[11] _hðtÞrhD 3.0640 2.0797 – Corollary 1[11] – – – 1.9207 Theorem 1[13] hDr _hðtÞrhD 5.4741 3.7440 – Theorem 2[13] hDr _hðtÞrhD 6.5848 4.1767 – Theorem 3[13] hDr _hðtÞrhD 7.5173 5.3993 – Corollary 1[13] _hðtÞrhD 3.7236 2.9229 – Corollary 2[13] – – – 2.9208 Theorem 1 hDr _hðtÞrhD 7.1744 4.0356 – Theorem 2 (ρ¼0.66) hDr _hðtÞrhD 12.1028 6.3209 – Theorem 3 (ρ¼0.66) hDr _hðtÞrhD 16.8744 9.3778 – Corollary 1 (ρ¼0.66) _hðtÞrhD 4.6785 3.1153 – Corollary 1 ( ~Q ¼ 0; ρ¼0.66) – – – 2.3631
W2 inℝnm, for all vectors
ξ
inℝm let define the functionΘ
ðα
; RÞ given byΘ
ðα
; RÞ ¼1 αξ
TWT1RW1ξ
þ1 1αξ
T WT2RW2ξ
:Then, if there exists a matrix X inℝnnsuch that R X
R
40, the following inequality holds:
min αA ð0;1Þ
Θ
ðα
; RÞZ W1ξ
W2ξ
" #T R X R W 1ξ
W2ξ
" # :Lemma 2. [32]. For a given matrix R40, the following inequality holds for all continuously differentiable functions
σ
in ½a; b-ℝn:Z b a
σ
_ TðuÞ Rσ
_ ðuÞduZ 1 b aðσ
ðbÞσ
ðaÞÞ T Rðσ
ðbÞσ
ðaÞÞþ 3 b aδ
T Rδ
whereδ
¼σ
ðbÞþσ
ðaÞ 2 b a Rb aσ
ðuÞdu.Lemma 3. [33]. Let
ξ
Aℝn,Φ
¼Φ
TAℝnnand
Η
Aℝmnsuch thatrank ð
Η
Þon. Then, the following statements are equivalent: (1)ξ
TΦξ
o0;Ηξ
¼ 0;ξ
a0(2)ð
Η
?ÞTΦΗ
?o0, whereΗ
? is a right orthogonalcomple-ment of
Η
.Lemma 4. [34]. For symmetric matrices of appropriate dimensions R40;
Ω
; and a matrixΓ
, the following two statements are equivalent: (1)Ω
Γ
RΓ
To0 and (2) there exists a matrix of the appropriate dimensionΠ
such thatΩ
þΓΠ
T þΠΓ
TΠ
Π
T R " # o0 ð6Þ3. Main new results
In this section, by using a novel activation function partitioning method, a new asymptotic stability criterion for system (3) with time-varying delays is proposed. For simplicity of matrix representation, we setup block entry matrices e0¼ 013nn; eiði ¼ 1; :::; 13ÞAℝ13nn(for
example eT 2 ¼ 0 I 0 0 0 0 0 0 0 0 0 0 0 ) and we define
ξ
T ðtÞ ¼ x TðtÞ xTðt hðtÞÞ xTðt h MÞ _xTðtÞ _xTðt hðtÞÞ _xTðt hMÞ gTðxðtÞÞ gTðxðt hðtÞÞÞ gTðxðt h MÞÞ hðtÞ1 Rt t hðtÞxTðsÞds hM hðtÞ1 Rt hðtÞ t hM x TðsÞds Z t t hðtÞ gTðxðsÞÞdsZ t hðtÞ t hM gTðxðsÞÞds #ω
TðtÞ ¼ xTðtÞ xTðt h MÞ Z t t hM xTðsÞdsZ t t hM gTðxðsÞÞds xTðt hðtÞÞ ;Η
¼ C 0 n2n I 0n2n A B 0n5nα
Tðt; sÞ ¼ x TðtÞ xTðsÞ _xTðsÞ gTðxðsÞÞ xTðt hðtÞÞ;β
T ðsÞ ¼ x TðsÞ _xTðsÞ gTðxðsÞÞϒ
1ð _hÞ ¼ diagfI; I; I; I; ð1 _hðtÞÞIg;ϒ
2ð _hÞ ¼ diagfI; I; I; I; ð1 _hðtÞÞI; Ig;ϒ
3ð _hÞ ¼ diagfI; I; I; I; ð1 _hðtÞÞI; ð1 _hðtÞÞIgΠ
0 1¼ e 1 e3 e0 e12þe13 e2;Π
1 1¼ e 0 e0 e10 e0 e0;Π
2 1¼ e 0 e0 e11 e0 e0Π
2¼ e 4 e6 e1e3 e7e9 e5;Π
3¼ e 1 e4 e7;Π
4¼ e 2 e5 e8;Π
0 5¼ e0 e0 e1e2 e12 e0 e0Π
1 5¼ e1 e10 e0 e0 e2 e1e10 ;Π
6¼ e 4 e0 e0 e0 e5;Π
7¼ e 3 e6 e9Π
0 8¼ e0 e0 e2e3 e13 e0 e0 ;Π
2 8¼ e 1 e11 e0 e0 e2 e2e11;Π
0 9¼ h Me1 e0 e1e3 e12þe13 hMe2Π
1 9¼ e 0 e10 e0 e0 e0;Π
29¼ e 0 e11 e0 e0 e0;Π
0 10¼ e0 e1e2 e12 e0 e2e3 e13Π
1 10¼ e 10 e0 e0 e0 e0 e0;Π
2 10¼ e0 e0 e0 e11 e0 e0 Km¼ diag k1; k 2; …; k n ; KM¼ diag k1þ; k þ 2; …; k þ n ; Kρ¼ Kmþρ
ðKMKmÞ; ð0rρ
r1ÞΦ
¼Ξ
SΞ
;Ξ
¼ M 0nn 3M ; S ¼ S11 S12 S21 S22 " # ;Χ
¼ Z W Z ;Θ
¼ Symf½e7e1KmT1½e7e1KMTþ½e8e2KmT2½e8e2KMTþ½e9e3KmT3½e9e3KMTg;
Θ
a¼ Symf½e7e1KρT1½e7e1KmTþ½e8e2KρT2½e8e2KmTþ½e9e3KρT3½e9e3KmTg;
Θ
b¼ Symf½e7e1KMT4½e7e1KρTþ½e8e2KMT5½e8e2KρTþ½e9e3KMT6½e9e3KρTg;
Φ
1 ¼ Symf½e7e1KmD1e4Tþ½e1KMe7D2eT4gþSymf½e9e3KmD5eT6þ½e3KMe9D6eT6g;
Φ
2ðḣÞ ¼ ð1ḣðtÞÞSymf½e8e2KmD3e5Tþ½e2KMe8D4eT5g;Φ
3¼ e 1Π
3 e2 e0Q e1Π
3 e2 e0T; Fig. 1. State trajectories of the system of Example 1 with time-varying delayΦ
4ðḣÞ ¼ ð1ḣðtÞÞ e 1Π
4 e2 e1e2Q e 1Π
4 e2 e1e2T;Φ
5ðḣÞ ¼ ð1ḣðtÞÞ e 1Π
4 e2 e0R e 1Π
4 e2 e0T;Φ
6¼ e 1Π
7 e2 e2e3R e 1Π
7 e2 e2e3 T ;Φ
7¼ e1Π
3 e2N e 1Π
3 e2T;Φ
8¼ e 1Π
7 e2N e 1Π
7 e2 T ;Ω
a¼ Symf½e7e8ðe1e2ÞKρL1½e7e8ðe1e2ÞKmTþ½e8e9ðe2e3ÞKρL2½e8e9ðe2e3ÞKmTg
Ω
b¼ Symf½e7e8ðe1e2ÞKρL3½e7e8ðe1e2ÞKMTþ½e8e9ðe2e3ÞKρL4½e8e9ðe2e3ÞKMTg
Σ
ð _hÞ ¼ SymfΠ
0 1Pϒ
1ð _hÞΠ
T2gþΦ
2ð _hÞΦ
4ð _hÞ þ SymfΠ
0 5Qϒ
2ð _hÞΠ
6 e4Tg þSymfΠ
0 8Rϒ
3ð _hÞΠ
6 e5TgþΦ
5ð _hÞþSymfΠ
09Nϒ
1ð _hÞΠ
T6gΣ
1ðḣÞ ¼ SymfΠ
11Pϒ
1ðḣÞΠ
T2þΠ
1 5Qϒ
2ðḣÞΠ
6 e4T þΠ
1 9Nϒ
1ðḣÞΠ
T6g;Σ
2ðḣÞ ¼ SymfΠ
21Pϒ
1ðḣÞΠ
T2þΠ
2 8Rϒ
3ðḣÞΠ
6 e5T þΠ
2 9Nϒ
1ðḣÞΠ
T6g;Y ¼ e 1e2 e1þe22e10 e2e3 e2þe32e11T;
Ψ
¼Φ
1þΦ
3Φ
6þΦ
7Φ
8þh2MΠ
3ZΠ
T3þh2
Me4MeT4Y T
Φ
Y; ð7Þ Theorem 1. For given positive scalar hM, any scalars hlDand h
u Dwith
condition C1, diagonal matrices Km, KM, network system (3) is
asymptotically stable, if there exist positive definite matrices PAℝ 5n5n; Q Aℝ6n6n; RAℝ6n6n; N Aℝ5n5n; Z Aℝ3n3n; M A
ℝnn; positive diagonal matrices D
i¼ diagfdi1; di2; …; ding
ði ¼ 1; …; 6Þ, Ti¼ diagfti1; ti2; …; tingði ¼ 1; 2; 3Þ; and any matrix W A
ℝ3n3n; with matrices and matrix
Π
having appropriate dimensions,such that the following LMIs are feasible for h ¼ ð0; hMÞ and for
_h ¼ ðhl D; h u DÞ. ð
Η
?ÞTΩ
ðh; _hÞðΗ
?ÞþSymfðΗ
?ÞTΓ
ðhÞΠ
TgΠ
Χ
" # o0 ð8ÞΧ
40;Φ
40 ð9Þ whereΩ
ðh; _hÞ ¼Σ
ð _hÞþ hðtÞΣ
1ð _hÞþðhMhðtÞÞΣ
2ð _hÞþΨ
þΘ
,Γ
ðhÞ ¼Π
0 10þhðtÞΠ
110þðhMhðtÞÞ
Π
210, and the other matrices aredefined in(7), with
Η
? the right orthogonal complement ofΗ
. Proof. For positive diagonal matrices Diði ¼ 1; :::; 6Þ and positivedefinite matrices P; Q; R; N; Z; M, we consider the following LKF candidate: V ¼X 6 i ¼ 1 ViðxtÞ ð10Þ where: V1¼
ω
TðtÞPω
ðtÞ V2¼ 2 Xn i ¼ 1 ðd1i Z xiðtÞ 0 ðgiðsÞk i sÞds þ d2i Z xiðtÞ 0 ðkiþs giðsÞÞdsÞ þ2X n i ¼ 1 ðd3i Z xiðt hðtÞÞ 0 ðgiðsÞk i sÞds þ d4i Z xiðt hðtÞÞ 0 ðkiþs giðsÞÞdsÞ þ2X n i ¼ 1 ðd5i Z xiðt hMÞ 0 ðgiðsÞk i sÞds þ d6i Z xiðt hMÞ 0 ðkiþs giðsÞÞdsÞ V3¼ Z t t hðtÞα
ðt; sÞ Rt s_xðuÞdu " #T Q Rtα
ðt; sÞ s _xðuÞdu " # ds þZ t hðtÞ t hMα
ðt; sÞ Rt hðtÞ s _xðuÞdu " #T R Rt hðtÞα
ðt; sÞ s _xðuÞdu " # ds V4¼ Z t t hMα
Tðt; sÞNα
ðt; sÞds V5¼ hM Z t t hM Z t sβ
T ðuÞZβ
ðuÞduds V6¼ hM Z t t hM Z t s _xTðuÞ M_xðuÞdudsThe time derivative of V1can be represented as
_V1¼
ξ
TðtÞ SymððΠ
01þhðtÞΠ
1 1þðhMhðtÞÞΠ
21ÞPϒ
1ð _hÞΠ
T2Þ n oξ
ðtÞ ð11Þ Similarly, we get _V2¼ 2½gðxðtÞÞxðtÞKmTD1_xðtÞþ2½xðtÞKMgðxðtÞÞTD2_xðtÞ þð1 _hðtÞÞf 2½gðxðt hðtÞÞÞxðt hðtÞÞKmTD3_xðt hðtÞÞ þ2½xðt hðtÞÞKMgðxðt hðtÞÞÞTD4_xðt hðtÞÞg þ2½gðxðt hMÞÞxðt hMÞKmTD5_xðt hMÞ þ2½xðt hMÞKMgðxðt hMÞÞTD6_xðt hMÞ ¼ξ
T ðtÞ ½Φ
1þΦ
2ð _hÞ n oξ
ðtÞ ð12ÞFurther, calculation of _V3gives
_V3¼
α
ðt; tÞ 0n1 " #T Qα
ðt; tÞ 0n1 " # ð1 _hðtÞÞα
ðt; t hðtÞÞ xðtÞ xðt hðtÞÞ " #T Qα
ðt; t hðtÞÞ xðtÞ xðt hðtÞÞ " # þ2Z t t hðtÞα
ðt; sÞ Rt s_xðuÞdu " #T Qϒ
1ð _hÞη
ðtÞ _xðtÞ " # ds þð1 _hðtÞÞα
ðt; t hðtÞÞ 0n1 " #T Rα
ðt; t hðtÞÞ 0n1 " #α
ðt; t hMÞ xðt hðtÞÞ xðt hMÞ " #T Rα
ðt; t hMÞ xðt hðtÞÞ xðt hMÞ " # þ2Z t hðtÞ t hMα
ðt; sÞ Rt hðtÞ s _xðuÞdu " #T Rϒ
1ð _hÞη
ðtÞ ð1 _hðtÞÞ_xðt hðtÞÞ " # dsFig. 2. State trajectories of the system of Example 1 with time-varying delay hðtÞ ¼ 8:4778þ0:9 sin ðtÞ.
¼
ξ
T ðtÞΦ
3Φ
4ð _hÞþSymððΠ
05þhðtÞΠ
1 5ÞQϒ
2ð _hÞΠ
6 e4TÞ n þΦ
5ð _hÞΦ
6þSymððΠ
0 8þðhMhðtÞÞΠ
28ÞRϒ
3ð _hÞΠ
6 e5TÞ oξ
ðtÞ ð13Þ whereη
ðtÞ ¼ _xðtÞ 0 0 0 _xðt hðtÞÞh iT.The result of _V4, after an appropriate arrangement, leads to
_V4¼
α
Tðt; tÞNα
ðt; tÞα
Tðt; t hMÞNα
ðt; t hMÞ þ2 Z t t hMα
Tðt; sÞ Nϒ
1ð _hÞη
ðtÞds ¼ξ
T ðtÞΦ
7Φ
8þSymððΠ
09þhðtÞΠ
1 9 n þðhMhðtÞÞΠ
29ÞNϒ
1ð _hÞΠ
T 6Þ oξ
ðtÞ ð14ÞFurthermore, by using Jensen’s inequality and Lemma 1, _V5is
found bounded as _V5¼ h 2 M
β
T ðtÞZβ
ðtÞhM Rt t hðtÞβ
T ðsÞZβ
ðsÞdshM Rt hðtÞ t hMβ
T ðsÞZβ
ðsÞds rh2 Mβ
TðtÞZβ
ðtÞ hM hðtÞ Z t t hðtÞβ
ðsÞds T Z Z t t hðtÞβ
ðsÞds hM hM hðtÞ Z t hðtÞ t hMβ
ðsÞds !T Z Z t hðtÞ t hMβ
ðsÞds ! rh2 Mβ
T ðtÞZβ
ðtÞ Rt t hðtÞβ
ðsÞds Rt hðtÞ t hMβ
ðsÞds 2 4 3 5 T Z W Z Rt t hðtÞβ
ðsÞds Rt hðtÞ t hMβ
ðsÞds 2 4 3 5 ¼ξ
T ðtÞ h2 MΠ
3ZΠ
T3Γ
ðhÞΧΓ
T ðhÞ n oξ
ðtÞ ð15ÞFinally, _V6is easily obtained as follows:
_V6¼ h2M_x TðtÞM_xðtÞh M Z t t hM _xTðsÞM_xðsÞds ð16Þ
By applying Lemma 1 and Lemma 2, wefind hMRt ht M_x TðsÞM_xðsÞds ¼ h M Z t t hðtÞ _xTðsÞM_xðsÞds hM Z t hðtÞ t hM _xTðsÞM_xðsÞds r hM hðtÞ½xðtÞxðt hðtÞÞ TM½xðtÞ xðt hðtÞÞ hM hM hðtÞ½xðt hðtÞÞxðt hMÞ TM½xðt hðtÞÞ xðt h MÞ 3hM hðtÞ½xðtÞþxðt hðtÞÞ 2 hðtÞ Zt t hðtÞ xðsÞdsTM½xðtÞ þxðt hðtÞÞ 2 hðtÞ Z t t hðtÞ xðsÞds 3hM hM hðtÞ½xðt hðtÞÞþxðt hMÞ 2 hM hðtÞ Z t hðtÞ t hM xðsÞdsTM½xðt hðtÞÞ þ xðt h MÞhM hðtÞ2 Z t hðtÞ t hM xðsÞds ¼ hM hðtÞ xðtÞ xðt hðtÞÞ xðtÞþ xðt hðtÞÞ 2 hðtÞ Rt t hðtÞxðsÞds " #T
Ξ
xðtÞþ xðt hðtÞÞ 2xðtÞxðt hðtÞÞ hðtÞ Rt t hðtÞxðsÞds " # hM hM hðtÞ xðt hðtÞÞ xðt hMÞ xðt hðtÞÞ þ xðt hMÞhM hðtÞ2 Rt hðtÞ t hM xðsÞds 2 4 3 5 TΞ
xðt hðtÞÞ þ xðt hxðt hðtÞÞ xðt hMÞ MÞhM hðtÞ2 Rt hðtÞ t hM xðsÞds 2 4 3 5 r xðtÞ xðt hðtÞÞ xðtÞþ xðt hðtÞÞ 2 hðtÞ Rt t hðtÞxðsÞds xðt hðtÞÞ xðt hMÞ xðt hðtÞÞ þxðt hMÞhM hðtÞ2 Rt hðtÞ t hM xðsÞds 2 6 6 6 6 6 4 3 7 7 7 7 7 5 TΦ
xðtÞ xðt hðtÞÞ xðtÞþ xðt hðtÞÞ 2 hðtÞ Rt t hðtÞxðsÞds xðt hðtÞÞ xðt hMÞ xðt hðtÞÞ þ xðt hMÞh 2 M hðtÞ Rt hðtÞ t hM xðsÞds 2 6 6 6 6 6 4 3 7 7 7 7 7 5: Hence _V6ðxtÞrξ
T ðtÞðh2 Me4MeT4Y TΦ
YÞξ
ðtÞ ð17Þ From (5) it follows that for any positive diagonal matrices Ti¼ diagfti1; ti2;…; tingði ¼ 1; 2; 3Þ, the following inequality holds:0r 2X n i ¼ 1 t1i giðxiðtÞÞk i xiðtÞ giðxiðtÞÞk þ i xiðtÞ 2X n i ¼ 1 t2i giðxiðt hðtÞÞÞk i xiðt hðtÞÞ giðxiðt hðtÞÞÞ kiþxiðt hðtÞÞ 2X n i ¼ 1 t3igiðxiðt hMÞÞkixiðt hMÞgiðxiðt hMÞÞ kiþxiðt hMÞ ¼
ξ
T ðtÞΘξ
ðtÞ ð18ÞNow, with (11)–(18) at hand, we get _V r
ξ
T ðtÞΩ
ðh; _hÞΓ
ðhÞΧΓ
T ðhÞ n oξ
ðtÞ ð19Þ By virtue of Lemma 3,ξ
TðtÞnΩ
ðh; _hÞΓ
ðhÞΧΓ
TðhÞoξ
ðtÞ o0 with 0 ¼Ηξ
ðtÞis equivalent toð
Η
?ÞT½Ω
ðh; _hÞ
Γ
ðhÞΧΓ
TðhÞ ð
Η
?Þo0 ð20Þ By virtue of Lemma 4, inequality(20)is equivalent toð
Η
?ÞTΩ
ðh; _hÞðΗ
? ÞþSymfðΗ
?ÞTΓ
ðhÞΠ
TgΠ
Χ
" # o0: ð21Þ whereΠ
is a matrix of appropriate dimensions. The above con-dition is affine, and consequently convex, with respect to h(t) and _hðtÞ, and it is necessary and sufficient to ensure that inequality(21)holds at vertices of the intervals ½0; hM ½hlD; h u
D as shown in[32].
Based on thisfinding, we know that inequality (21)holds if and
Fig. 3. State trajectories of the system of Example 2 with time-varying delay hðtÞ ¼ 4:4106þ0:1 sin ðtÞ.
only if (8)and(9)hold as well, and then network system(3) is asymptotically stable. Therefore then neural network system(1)is stable too. This completes the proof.
Remark 1. Recently, the reciprocally convex optimization techni-que and Wirtinger integral inequality are proposed in[17]and[32]
respectively, and the two methods are utilized in (17)here. We point out that in Lemma 2, the term 1
b að
σ
ðbÞσ
ðaÞÞTRð
σ
ðbÞσ
ðaÞÞis equal to Jensen’s inequality, and the new term 3 b a
δ
T
R
δ
can reduce the enlargement of the estimation of LKF. The usage of reciprocally convex optimization method avoids the enlargement of hðtÞandhMhðtÞ, and only introduces two matrices W; S. Then,the convex optimization method is used to handle _VðxtÞ.
Remark 2. In Theorem 1, first, the terms 1 hðtÞ Rt t hðtÞx TðsÞds and 1 hM hðtÞ Rt hðtÞ t hM x
TðsÞds are used for the vector
ξ
ðtÞ. Secondly, thestates xðt hðtÞÞ and xðt hMÞ as interval of integral terms are
taken, as shown in the second and third terms of V2. Therefore,
more information on the cross terms in gðxðt hðtÞÞ;xðt hðtÞÞ;_xðt hðtÞ and gðxðt hMÞÞ;xðt hMÞ;_xðt hMÞis being utilized.
Thirdly, we introduce new terms xðtÞ;xðt hðtÞÞ,Rt hðtÞ
s _xðuÞdu in V3,
which is different from the previous works. Thus, the results of time-derivative of the proposed V3contain some cross-times such
as 2ð1 hðtÞ Rt t hðtÞxðsÞdsÞðhðtÞQ21_xðtÞÞ; 2ðhðtÞ1 Rt t hðtÞxðsÞdsÞðhðtÞQ26_xðtÞÞ; 2 ð1 hðtÞ Rt
t hðtÞxðsÞdsÞðhðtÞQ25ð1 _hðtÞÞ_xðt hðtÞÞÞ; which were presented
in(13)and does not be used in existing results. These considera-tions highlight the main differences in the construction of the LKF candidate in this paper.
Remark 3. The maximal order of the LMIs and total number of the scalar variables are usually considered as the index of the calcula-tion complexities, how to get novel stability criteria which spend less time on the calculation and large delay bound is another key work in this paper. In the stability analysis of neural networks with time delays, a number of works choose delay-partitioning method to investigate. Generally, the delay-partitioning number was taken as the range of two as a tradeoff between the computational burden and the improvement of the feasible region. However, when the condition 0rhðtÞrhM is divided into 0rhðtÞrhM=2 and
hM=2rhðtÞrhM, the matrix formulation becomes more complex
and the dimension of stability conditions grows larger because it has more augmented vectors. Inspired by the activation functions dividing method for neural networks with time-varying delays in
[13], we divide the bounding of activation function kirfiðuÞ=ur
kiþof neurons with time-varying delays into k i rfiðuÞ=urkρi and kρi rfiðuÞ=urk þ i , kρi ¼ k i þ
ρ
ðk þ i k i Þ; 0rρ
r1. The calculationcomplexity of this partitioning method is less than delay parti-tioning method because the stability condition has less augmented vectors. This new activation partitioning technique for neural net-works with time delays is more general and less conservative than the one in[13]. The new bounding partitioning approach is utilized instead of using delay-partitioning method, which is used in The-orem 2 further below. Thus, through TheThe-orem 1 and TheThe-orem 2, less conservative stability criteria are derived in this paper.
Now, based on the results of Theorem 1, an improved stability criterion for system (3) is introduced.
Theorem 2. For given positive scalars
ρ
r1and hM, any scalars hlDand huD with condition C1, diagonal matricesKm,KMandKρ, network
system (3) is asymptotically stable, if there exist positive definite matrices PAℝ 5n5n; Q Aℝ6n6n; RAℝ6n6n; N Aℝ5n5n; Z A
ℝ3n3n; M Aℝnn; positive diagonal matrices D
i¼ diagfdi1; di2; …;
dingZ0ði ¼ 1; :::; 6Þ; Ti¼ diagfti1; ti2;…; tingZ0ði ¼ 1; :::; 6Þ, Li¼ diagf
li1; li2;…; lingZ0ði ¼ 1; :::; 4Þ, and any matrix W Aℝ3n3n; with
matrices and matrix
Π
having appropriate dimensions, such that the following LMIs for h ¼ ð0; hMÞ and for _h ¼ ðhlD; hu DÞ ð
Η
?ÞTð ^Ω
ðh; _hÞþΘ
ΔþΩ
ΔÞðΗ
?ÞþSymfðΗ
?ÞTΓ
ðhÞΠ
TgΠ
Χ
" # o0Δ
¼ a; b ð22ÞΧ
40;Φ
40 ð23Þare satisfied, where ^
Ω
ðh; _hÞ ¼Σ
ð _hÞþ hðtÞΣ
1ð _hÞþðhMhðtÞÞΣ
2ð _hÞþ
Ψ
and other matrices are defined in(7), and whereΗ
? is the right orthogonal complement ofΗ
.Proof Consider the same LKF proposed in Theorem 1, we derive the actual results from the following two cases.
Case 1: kirðgiðuÞgiðvÞÞ=ðuvÞrkρi; ð24Þ by choosing v ¼ 0, it is equivalent to ½giðuÞk i u½giðuÞkρiuo0 ð25Þ
From (25)it is found that, for any positive diagonal matrices T1¼ diag tf11; ⋯; t1ng;T2¼ diag tf21; ⋯; t2ng;and T3¼ diag tf31; ⋯; t3ng; Table 2
delay bounds hMwith different hD
Methods Condition of _hðtÞ hD¼ 0:1 hD¼ 0:5 hD¼ 0:9 Unknown
orZ1 Theorem 1[26] (m ¼ 2) _hðtÞrhD 3.7525 2.7353 2.2760 – Theorem 1 with-out V2[26] – – – – 2.1326 Theorem 2[31] _hðtÞrhD 3.7857 3.0546 2.6703 – Theorem 1 with-out V3[31] – – – – 2.6575 Theorem 1[13] hDr _hðtÞrhD 3.9269 3.4072 2.8337 – Theorem 2[13] hDr _hðtÞrhD 3.9332 3.5277 3.2025 – Theorem 3[13] hDr _hðtÞrhD 3.9337 3.5307 3.2627 – Corollary 1[13] _hðtÞrhD 3.8102 3.1518 2.8402 – Corollary 2[13] - – – – 2.8379 Theorem 1 hDr _hðtÞrhD 4.5086 3.8091 3.2895 – Theorem 2 (ρ¼0.40) hDr _hðtÞrhD 4.5045 3.8334 3.4840 – Theorem 3 (ρ¼0.40) hDr _hðtÞrhD 4.5106 3.8716 3.5482 – Corollary 1 (ρ¼0.40) _hðtÞrhD 4.4338 3.5344 3.1410 – Corollary 1 ( ~Q ¼ 0;ρ¼0.40) – – – – 3.1122
Fig. 4. State trajectories of the system of Example 2 with time-varying delayhðtÞ ¼ 3:3716þ0:5 sin ðtÞ.
the following inequality holds: 0r 2X n i ¼ 1 t1i giðxiðtÞÞkixiðtÞ giðxiðtÞÞkρixiðtÞ 2X n i ¼ 1 t2igiðxiðt hðtÞÞÞkixiðt hðtÞÞgiðxiðt hðtÞÞÞ kρixiðt hðtÞÞ 2X n i ¼ 1 t3igiðxiðt hMÞÞkixiðt hMÞgiðxiðt hMÞÞ kρixiðt hMÞ ¼
ξ
TðtÞΘ
aξ
ðtÞ ð26ÞFrom inequality(24), the following conditions hold: kir giðxiðtÞÞgiðxiðt hðtÞÞÞ xiðtÞxiðt hðtÞÞ rk ρ i kir giðxiðt hðtÞÞÞgiðxiðt hMÞÞ xiðt hðtÞÞxiðt hMÞ rk ρ i ð27Þ
For i ¼ 1; ⋯; n; the above two conditions are equivalent to giðxiðtÞÞgiðxiðt hðtÞÞÞk i ðxiðtÞxiðt hðtÞÞÞ giðxiðtÞÞgiðxiðt hðtÞÞÞkρiðxiðtÞxiðt hðtÞÞÞ r0 ð28Þ giðxiðt hðtÞÞÞgiðxiðt hMÞÞkiðxiðt hðtÞÞxiðt hMÞÞ g iðxiðt hðtÞÞÞgiðxiðt hMÞÞkρiðxiðt hðtÞÞxiðt hMÞÞr0 ð29Þ Therefore, for any positive diagonal matrices L1¼ diag l11; ⋯;
l1ng;L2¼ diag l21; ⋯; l2n; the following inequality holds:
0r 2X n i ¼ 1 l1i giðxiðtÞÞgiðxiðt hðtÞÞÞk i ðxiðtÞxiðt hðtÞÞÞ giðxiðtÞÞgiðxiðt hðtÞÞÞkρiðxiðtÞxiðt hðtÞÞÞ 2X n i ¼ 1 l2i giðxiðt hðtÞÞÞgiðxiðt hMÞÞk i ðxiðt hðtÞÞ xiðt hMÞÞ g iðxiðt hðtÞÞÞgiðxiðt hMÞÞ kρiðxiðt hðtÞÞxiðt hMÞÞ ¼
ξ
T ðtÞΩ
aξ
ðtÞ ð30ÞThen, from the proof of Theorem 1, whenkirðgiðuÞ
giðvÞÞ=ðuvÞrkρi, an upper bound of _Vcan be shown
_V r
ξ
TðtÞf ^Ω
ðh; _hÞþΘ
aþΩ
aΓ
ðhÞΧΓ
TðhÞ gξ
ðtÞ ð31Þ Case 2: kρirðgiðuÞgiðvÞÞ=ðuvÞrk þ i ð32ÞFor this case, we define positive definite diagonal matrices T4
¼ diagft41; t42; …; t4ng; T5¼ diagft51; t52; …; t5ng;T6¼ diagft61; t62;
…; t6ng, and L3¼ diag l31; ⋯; l3n
;L4¼ diag l41; ⋯; l4n
, applying a similar procedure as the one used in Case 1, and therefore ulti-mately we obtain
_V r
ξ
TðtÞf ^
Ω
ðh; _hÞþΘ
bþΩ
bΓ
ðhÞΧΓ
TðhÞ g
ξ
ðtÞ ð33Þ Finally, we get an upper bound of _Vfor kirðgiðuÞgiðvÞÞ=ðuvÞr kiþas follows: _V r
ξ
TðtÞf ^
Ω
ðh; _hÞþΘ
ΔþΩ
ΔΓ
ðhÞΧΓ
TðhÞgξ
ðtÞ ð34Þwhere
Θ
Δ;Ω
ΔðΔ
¼ a and bÞare defined in (7). Similarly to the proof of Theorem 1, if inequalities (22) and (23) hold, then net-work system (3) is asymptotically stable for condition C1 fulfilled andkirðgiðuÞgiðvÞÞ=ðuvÞrkþ
i , and so is the neural network
system (1). This completes the proof.
Remark 4. In Theorem 2, a new activation function partitioning method was applied to derive a novel criterion for system(3). Here a different novel functional for the two subintervals of the bounding of the activation function was constructed to obtain a new result for system(3). By a procedure similar to the proof of Theorem 2, we also introduce Theorem 3 as follows.
Theorem 3. For given positive scalars
ρ
r1 and hM, any scalars hlDand huDwith condition C1, diagonal matrices Km, KMand Kρ, network
system (3) is asymptotically stable, if there exist positive definite matrices PAℝ 5n5n; QAℝ6n6n; RAℝ6n6n; N A
ℝ5n5n;Z Aℝ3n3n;M Aℝnn; positive diagonal matrices D i¼
diagfdi1; di2; …; dingði ¼ 1; :::; 12Þ, Ti¼ diagfti1; ti2;…; tingði ¼ 1; :::; 6Þ,
Li¼ diagfli1; li2;…; lingði ¼ 1; :::; 4Þ, and any matrix W Aℝ3n3n; with
matrices and matrix
Π
having appropriate dimensions, such that the following LMIs for h ¼ ð0; hMÞ and for _h ¼ hlD; hu D ð
Η
?ÞTΩ
Δðh; _hÞðΗ
?ÞþSymfðΗ
?ÞTΓ
ðhÞΠ
TgΠ
Χ
" # o0Δ
¼ a; b ð35ÞΧ
40;Φ
40 ð36Þare satisfied, where
Ω
Δðh; _hÞ ¼Σ
Δð _hÞþhðtÞΣ
1ð _hÞþðhMhðtÞÞΣ
2ð _hÞ þΨ
ΔþΘ
ΔþΩ
ΔΣ
Δð _hÞ ¼ SymðΠ
01Pϒ
1ð _hÞΠ
T2ÞþΦ
2Δð _hÞΦ
4ð _hÞ þ SymðΠ
0 5Qϒ
2ð _hÞΠ
6 e4TÞ þSymðΠ
0 8Rϒ
3ð _hÞΠ
6 e5TÞþΦ
5ð _hÞþSymðΠ
09Nϒ
1ð _hÞΠ
T6ÞΨ
Δ¼Φ
1ΔþΦ
3Φ
6þΦ
7Φ
8þh 2 MΠ
3ZΠ
T 3þh 2 Me4MeT4Y TΦ
YΦ
1a¼ Symf½e7e1KmD1eT4þ½e1Kρe7D2eT4þ½e9e3KmD5eT6þ½e3Kρe9D6eT6g
Φ
1b¼ Symf½e7e1KρD7eT4þ½e1KMe7D8eT4þ½e9e3KρD11eT6þ½e3KMe9D12eT6g
Φ
2að _hÞ ¼ ð1 _hðtÞÞSymf½e8e2KmD3e5Tþ½e2Kρe8D4eT5gΦ
2bð _hÞ ¼ ð1 _hðtÞÞSymf½e8e2KρD9eT5þ½e2KMe8D10eT5g ð37Þwhile other matrices are defined in (7), and
Η
? is the right orthogonal complement ofΗ
.Proof By considering the same LKF proposed in Theorem 1 except for the term V2, in Case 1 and Case 2, we choose different
V2 for the two subintervals of the bounding of the activation
function as follows:
Fig. 5. State trajectories of the system of Example 2 with time-varying delay hðtÞ ¼ 2:6482þ0:9 sin ðtÞ.
Case 1: when kirðgiðuÞgiðvÞÞ=ðuvÞrkρi; we choose V2as V2a¼ 2 Xn i ¼ 1 ðd1i Z xiðtÞ 0 ðgiðsÞk i sÞds þ d2i Z xiðtÞ 0 ðkρis giðsÞÞdsÞ þ2X n i ¼ 1 ðd3i Z xiðt hðtÞÞ 0 ðgiðsÞk i sÞdsþ d4i Zxiðt hðtÞÞ 0 ðkρis giðsÞÞdsÞ þ2X n i ¼ 1 ðd5i Z xiðt hMÞ 0 ðgiðsÞk i sÞds þ d6i Z xiðt hMÞ 0 ðkρis giðsÞÞdsÞ ð38Þ Case 2: when kρirðgiðuÞgiðvÞÞ=ðuvÞrk
þ i , we choose V2as V2b¼ 2 Xn i ¼ 1 ðd7i ZxiðtÞ 0 ðgiðsÞkρisÞds þ d8i Z xiðtÞ 0 ðkiþs giðsÞÞdsÞ þ2X n i ¼ 1 ðd9i Zxiðt hðtÞÞ 0 ðgiðsÞkρisÞds þd10i Z xiðt hðtÞÞ 0 ðkiþs giðsÞÞdsÞ þ2X n i ¼ 1 ðd11i Z xiðt hMÞ 0 ðgiðsÞkρisÞds þd12i Z xiðt hMÞ 0 ðkiþs giðsÞÞdsÞ ð39Þ
The calculation of _V2a; _V2bis similar to that of _V2in Theorem 1.
Thus, if inequalities(35)and(36)hold, then network system(3)is asymptotically stable, and so is neural network (1). This completes the proof.
Remark 5.. In Theorem 3, we consider that h(t) satisfies condition C1, but there are many systems satisfying condition C2. Therefore we introduced Corollary 1 in order to analyze the stability of neural networks satisfying the condition C2 by settingD3; D4; D9;
D10¼ 0, R ¼ 0 and changing V1; V2; V3; V4.
In Corollary 1, block entry matrices~e0¼ 012nn; ~eiAℝ12nnði ¼
1; :::; 12Þ are used and the following notations are defined for the sake of the simplicity of the matrix notation:
~
ξ
TðtÞ ¼ xTðtÞ xTðt hðtÞÞ xTðt h MÞ _xTðtÞ _xTðt h MÞ gTðxðtÞÞ gTðxðt hðtÞÞÞ gTðxðt hMÞÞ 1 hðtÞ Rt t hðtÞx TðsÞds 1 hM hðtÞ Rt hðtÞ t hM x TðsÞds Rt t hðtÞg TðxðsÞÞds Z t hðtÞ t hM gTðxðsÞÞds # ~ω
TðtÞ ¼ xTðtÞ xTðt h MÞ Rt t hMx TðsÞds Rt t hMg TðxðsÞÞds h iΗ
¼ C 0 n2n I 0nn A B 0n5n; Y ¼ ~e 1~e2 ~e1þ~e22~e9 ~e2~e3 ~e2þ~e32~e10T ~Π
0 1¼ ~e1 ~e3 ~e0 ~e11þ~e12 ;Π
~1 1¼ ~e0 ~e0 ~e9 ~e0 ; ~Π
2 1¼ ~e0 ~e0 ~e10 ~e0 ~Π
2¼~e4 ~e5 ~e1~e3 ~e6~e8; ~Π
3¼ ~e 1 ~e6 ~e0;Π
~4¼ ~e 2 ~e7 ~e1~e2; ~Π
0 5¼ ~e 0 ~e11 ~e0; ~Π
1 5¼ ~e 9 ~e0 ~e1~e9; ~Π
6¼ ~e 0 ~e0 ~e4;Π
~7¼ ~e 1 ~e4 ~e6; ~Π
8¼ ~e 3 ~e5 ~e8;Π
~ 0 9¼ ~e 0 ~e1~e2 ~e11 ~e0 ~e2~e3 ~e12; ~Π
1 9¼ ~e9 ~e0 ~e0 ~e0 ~e0 ~e0 ; ~Π
2 9¼ ~e0 ~e0 ~e0 ~e10 ~e0 ~e0 ~Θ
a¼ Symð½~e6~e1KρT1½~e6~e1KmT þ½~e7~e2KρT2½~e7~e2KmTþ½~e8~e3KρT3½~e8~e3KmTÞ;Θ
b¼ Symð½ e 6 e 1KMT4½ e 6 e 1KρT þ½ e7 e 2KMT5½ e 7 e 2KρTþ½ e 8 e 3KMT6½ e 8 e 3KρTÞ;Ω
a¼ Symð½ e 6 e 7ð e 1 e 2ÞKρL1½ e 6 e 7ð e 1 e 2ÞKmT þ½ e7 e 8ð e 2 e 3ÞKρL2½ e 7 e 8ð e 2 e 3ÞKmTÞ; ~Ω
b¼ Symð½~e6~e7ð~e1~e2ÞKρL3½~e6~e7ð~e1~e2ÞKMT þ½~e7~e8ð~e2~e3ÞKρL4½~e7~e8ð~e2~e3ÞKMTÞ ~Σ
¼ Symð ~Π
01~P ~Π
T 2Þ;Σ
~1¼ Symð ~Π
1 1~P ~Π
T 2ÞþSymð ~Π
1 5~Q ~Π
T 6Þ; ~Σ
2¼ Symð ~Π
2 1~P ~Π
T 2Þ ~Φ
1a ¼ Symð½~e6~e1KmD1~eT4þ½~e1Kρ~e6D2~eT4þ½~e8~e3KmD5~eT5 þ½~e3Kρ~e8D6~eT5Þ ~Φ
1b ¼ Symð½~e6~e1KρD7~eT4þ½~e1KM~e6D8~eT4þ½~e8~e3KρD11~eT5 þ½~e3KM~e8D12~eT5Þ ~Ψ
Δ¼ ~Σ
þ ~Φ
1Δþ ~Π
3~Q ~Π
T 3 ð1h u DÞ ~Π
4~Q ~Π
T 4þSymð ~Π
0 5~Q ~Π
T 6Þ þ ~Π
7~N ~Π
T 7 ~Π
8~N ~Π
T 8þh 2 MΠ
~7Z ~Π
T 7þh 2 M~e4M~eT4 ~Y TΦ
~Y ð40ÞCorollary 1. For given positive scalars
ρ
r1 and hM, any scalarhuDwith condition C2, diagonal matrices Km,KM and Kρ, network
system(3)is asymptotically stable, if there exist positive definite matrices ~PAℝ4n4n, ~NAℝ3n3n, ~Q Aℝ3n3n, ZAℝ3n3n, MAℝnn,
positive diagonal matrices Di¼ diagfdi1; di2; …; dingði ¼ 1; 2; 5; 6; 7;
8; 11; 12Þ, Ti¼ diagfti1; ti2;…; tingði ¼ 1; :::; 6Þ;Li¼ diagfli1; li2;…; ling
ði ¼ 1; :::; 4Þ, and any matrix W Aℝ3n3nwith matrices and matrix
~
Π
having appropriate dimensions, such that the following LMIs for h ¼ ð0; hMÞ ð ~Η
?ÞTΩ
~ ΔðhÞð ~Η
?ÞþSymfð ~Η
?ÞTΓ
~ðhÞ ~Π
TgΠ
~Χ
" # o0Δ
¼ a; b ð41ÞΧ
40;Φ
40 ð42Þare satisfied, where
Ω
~ΔðhÞ ¼ hðtÞ ~Σ
1þðhMhðtÞÞ ~Σ
2þ ~Ψ
Δþ~
Θ
Δþ ~Ω
Δ,Δ
¼ a; b, ~Γ
ðhÞ ¼ ~Π
0 9þhðtÞ ~Π
1 9þðhMhðtÞÞ ~Π
2 9, and othermatrices are defined in(7)or(40), while ~
Η
? is the right ortho-gonal complement of ~Η
.Proof Consider the following LKF candidate ~V ðxtÞ ¼ X6 i ¼ 1 ~ViðxtÞ where ~V1¼
ω
~TðtÞ ~Pω
~ðtÞ ~V2a¼ 2 Xn i ¼ 1 ðd1i Z xiðtÞ 0 ðgiðsÞk i sÞdsþ d2i ZxiðtÞ 0 ðkρis giðsÞÞdsÞ þ2X n i ¼ 1 ðd5i Z xiðt hMÞ 0 ðgiðsÞk i sÞdsþd6i Z xiðt hMÞ 0 ðkρis giðsÞÞdsÞ; ~V2b¼ 2 Xn i ¼ 1 ðd7i Z xiðtÞ 0 ðgiðsÞkρisÞds þ d8i Z xiðtÞ 0 ðkiþs giðsÞÞdsÞ þ2X n i ¼ 1 ðd11i Z xiðt hMÞ 0 ðgiðsÞkρisÞds þ d12i Z xiðt hMÞ 0 ðkþ i s giðsÞÞdsÞ; ~V3¼ Z t t hðtÞ xðsÞ gðxðsÞÞ Rt s_xðuÞdu 2 6 4 3 7 5 T ~Q xðsÞ gðxðsÞÞ Rt s_xðuÞdu 2 6 4 3 7 5ds ~V4¼ Z t t hM
β
TðsÞ ~ Nβ
ðsÞds; ~V5¼ V5; ~V6¼ V6The proof is very similar to the one of Theorem 3 so that we can see inequalities (41) and(42)to guarantee the asymptotic stability of network system(3)hence of the neural network(1).
In some circumstances, the information on the derivative of the delay may not be always available after all. The criterion for such a case can be readily derived from Corollary 1 by setting ~Q ¼ 0.
4. Numerical examples
In this section, two illustrative examples are presented and the respective results are summarized in comparison with relevant ones in the literature. These clearly show the effectiveness of the here proposed method and demonstrate the achieved improvements . Example 1. Consider the neural network system (3) with the following parameters C ¼ 2 0 0 2 ; A ¼ 11 11 ; B ¼ 01:88 11 ; Km¼ diagf0; 0g; KM¼ diagf0:4; 0:8g: ð43Þ
With the condition hDr _hðtÞrhD, our results obtained by
Theorems 1–3, are shown in Table 1. Also, when _hðtÞrhD and
unknown too, the corresponding results obtained by Corollary 1 are included inTable 1as well. It can be seen that our results based on Theorems 1–3, compared to the results of [11,13,28]
indeed improve the feasible region of stability criteria. The upper bounds obtained by Theorem 1 are larger than the results of Theorem 1 in [13], which shows that our new LKF, Wirtinger integral inequality and combined convex method can and do reduce the conservatism. The results based on Corollary 1 improve the feasible region of stability criteria compared to those of[11]
and[28], however, fall short compared to the results of[13]when hDis unknown. It is worth pointing out that the results of Theorem
2 and Theorem 3 clearly provide larger delay bounds than those of Theorem 1 when
ρ
¼ 0:66, which shows the effectiveness of technique of partitioning the bounding of the activation function conditions.The time evolution responses of the neural network with a time-varying delay in Example 1 when hM¼ 16:8744 for hD¼ 0:8;
hðtÞ ¼ 16:0744þ0:8 sin ðtÞr16:8744 and hM¼ 9:3778 for hD¼
0:9,hðtÞ ¼ 8:4778þ0:9 sin ðtÞr9:3778 are shown inFigs. 1and 2. These were obtained by setting xð0Þ ¼ ½2; 3T, gðxðtÞÞ ¼
½0:4 tanhðx1ðtÞÞ; 0:8 tanhðx2ðtÞÞT. The resulting responses clearly
indicate fast asymptotic stability of the simulated neural networks with time-varying delay.
Example 2. Next, let us consider the neural networks(3)with the following parameters: C ¼ 1:2769 0 0 0 0 0:6231 0 0 0 0 0:9230 0 0 0 0 0:4480 2 6 6 6 4 3 7 7 7 5; A ¼ 0:0373 0:4852 0:3351 0:2336 1:6033 0:5988 0:3224 1:2352 0:3394 0:0860 0:3824 0:5785 0:1311 0:3253 0:9534 0:5015 2 6 6 6 4 3 7 7 7 5; B ¼ 0:8674 1:2405 0:5325 0:0220 0:0474 0:9164 0:0360 0:9816 1:8495 2:6117 0:3788 0:8428 2:0413 0:5179 1:1734 0:2775 2 6 6 6 4 3 7 7 7 5; Km¼ diagf0; 0; 0; 0g; KM¼ diagf0:1137; 0:1279; 0:7994; 0:2368g: ð44Þ
For condition C1 satisfied, the acceptable maximal upper bounds of the time delay based on Theorems 1-3 are shown in
Table 2. Also, when _hðtÞrhDand it is unknown, the corresponding
results obtained by means of Corollary 1 are also included in
Table 2. It can be seen that our results based on Theorems 1-3 and Corollary 1 compare favorably to the results of [13,26,31] and hence improve the feasible region of stability criteria.
Fig. 3–5depict simulation results for the state curves of the neural network in Example 2 with the following settings: hðtÞ ¼ 4:4106þ0:1 sin ðtÞr4:5106 and hD¼ 0:1;hðtÞ ¼ 3:3716þ 0:5 sin ð
tÞr3:8716 and hD¼ 0:5;hðtÞ ¼ 2:6482þ0:9 sin ðtÞr3:5482 and
hD¼ 0:9. The chosen initial condition is given as xð0Þ ¼
½3; 1:5; 1; 2T
. The activation functions are gðxðtÞÞ ¼ ½0:1137 tanhðx1ðtÞÞ; 0:1279 tanhðx2ðtÞÞ; 0:7994 tanhðx3ðtÞÞ;
0:2368 tanhðx4ðtÞÞT. Although these state time responses have
longer transient oscillations, nonetheless, very rapidly fall within boundaries smaller than an amplitude of 0.5. These responses again demonstrate fast asymptotic stability of neural networks with time-varying delays (1) is guaranteed.
5. Conclusions
The problem of delay-dependent stability for neural networks with time-varying delays is investigated in this paper and new less conservative asymptotic stability criteria derived. By constructing a suitable LKF and using a novel partitioning method for the bounding of the activations functions and by employing Wirtinger integral inequality to deal with the derivative of LKF, less con-servative delay-dependent stability criteria expressed in terms of LMIs are presented. The results for two illustrative examples from the literature are given in comparison with the previous ones that clearly demonstrate the improvements achieved.
This paper only consider the stability problems of CNNs, other problems such as robust stability, exponential stability, synchroni-zation, and so on, can be investigated using the new methodological approach. Also, it is worth noting, constructing a more suitable LKF and reducing the calculation enlargement in estimating the deriva-tive also needs further investigation. Systematic stability analysis and controller design for Takagi–Sugeno (T–S) fuzzy systems have wit-nessed growing interests[42–46]. Generally, most of the approaches involve the employment of a simple LKF[47], and the application of some more or less tight techniques[48], such as Moon’s inequality, free-weighting matrix, or Jensen’s inequality, to derive the stability criteria for the time-delayed T–S fuzzy system. How to extend the
new constructed LKF and Wirtinger inequality to investigate the stability and controller design problems for time-delayed T–S fuzzy system is our another future work direction.
Acknowledgment
This work is supported by the National Nature Science Foun-dation of China under Grants 61374154 and 61374072.
References
[1]L.O. Chua, L. Yang, Cellular neural networks: applications, IEEE Trans. Circuits Syst. 35 (1988) 1273–1290.
[2]Z. Waszczyszyn, L. Ziemianski, Neural networks in mechanics of structures and materials-new results and prospects of applications, Comput. Struct. 79 (2001) 2261–2276.
[3]S. Xu, J. Lam, W.C. Daniel, Y. Zou, Delay-dependent exponential stability for a class of neural networks with time delays, J. Comput. Appl. Math. 183 (2005) 16–28. [4]T. Li, L. Guo, C. Sun, C. Lin, Further results on delay-dependent stability criteria
of neural networks with time-varying delays, IEEE Trans. Neural Netw. 19 (2008) 726–730.
[5]T. Li, A. Song, S. Fei, T. Wang, Delay-derivative-dependent stability for delayed neural networks with unbounded distributed delay, IEEE Trans. Neural Netw. 21 (2010) 1365–1371.
[6]T. Li, A. Song, M. Xue, H. Zhang, Stability analysis on delayed neural networks based on an improved delay-partitioning approach, J. Comput. Appl. Math. 235 (2011) 3086–3095.
[7]J. Qiu, H. Yang, J. Zhang, Z. Gao, New robust stability criteria for uncertain neural networks with interval time-varying delays, Chaos, Soliton, Fractals 39 (2009) 579–585.
[8]C. Li, G. Feng, Delay-interval-dependent stability of recurrent neural networks with time-varying delay, Neurocomputing 72 (2009) 1179–1183.
[9]L. Hu, H. Gao, W.X. Zheng, Novel stability of cellular neural networks with interval time-varying delay, Neural Netw. 21 (2008) 1458–1463.
[10]Z. Liu, J. Yu, D. Xu, Vector Writinger-type inequality and the stability analysis of delayed neural network, Commun. Nonlinear Sci. Numer. Simulat. 18 (2013) 1247–1257.
[11]Y. Wang, C. Yang, Z. Zuo, On exponential stability analysis for neural networks with time-varying delays and general activation functions, Commun. Non-linear Sci. Numer. Simul. 17 (2012) 1447–1459.
[12]P. Balasubramaniam, S. Lakshmanan, R. Rakkiyappan, Delay-interval depen-dent robust stability criteria for stochastic neural networks with linear frac-tional uncertainties, Neurocomputing 72 (2009) 3675–3682.
[13]O.M. Kwon, M.J. Park, S.M. Lee, J.H. Park, E.J. Cha, Stability for neural networks with time-varying delays via some new approaches, IEEE Trans. Neural Netw. Learn. Syst 24 (2013) 181–193.
[14]O.M. Kwon, S.M. Lee, J.H. Park, E.J. Cha, New approaches on stability criteria for neural networks with interval time-varying delays, Appl. Math. Comput. 213 (2012) 9953–9964.
[15]Y. He, M. Wu, J.H. She, An improved global asymptotic stability criterion for delayed cellular neural networks, IEEE Trans. Neural Netw. 17 (2006) 250–252. [16]Y. He, G. Liu, D. Rees, M. Wu, Stability Analysis for neural networks with
time-varying interval delay, IEEE Trans. Neural Netw. 18 (2007) 1850–1854. [17] P. Park, J.W. Ko, C. Jeong, Reciprocally convex approach to stability of systems
with time-varying delays, Automatica 47 (2011) 235–238.
[18]H. Li, B. Chen, Q. Zhou, W. Qian, Robust stability for uncertain delayed fuzzy hopfield neural networks with markovian jumping parameters, IEEE Trans. Syst., Man, Cybern. B, Cybern. 39 (2009) 94–102.
[19]P. Park, J.W. Ko, Stability and robust stability for systems with a time-varying delay, Automatica 43 (2007) 1855–1858.
[20]Q.L. Han, A discrete delay decomposition approach to stability of linear retarded and neutral systems, Automatica 45 (2009) 517–524.
[21]H. Li, Y. Gao, L. Wu, H.K. Lam, Fault detection for T-S fuzzy time-delay systems: delta operator and input-output methods, IEEE Trans. Cybern. 45 (2015) 229–241.
[22]D. Yue, E.G. Tian, Y.J. Zhang, C. Peng, Delay-distribution-dependent stability and stabilization of T-S fuzzy systems with probabilistic interval delay, IEEE Trans. Syst. Man Cybern. B, Cybern. 39 (2009) 503–516.
[23]T. Li, T. Wang, A.G. Song, S.M. Fei, Combined convex technique on delay-dependent stability for delayed neural networks, IEEE Trans, Neural Netw. Learn. Syst. 24 (2013) 1459–1466.
[24]J.H. Kim, Note on stability of linear systems with time-varying delay, Auto-matica 47 (2011) 2118–2121.
[25]Y. Zhang, D. Yue, E. Tian, New stability criteria of neural networks with interval time-varying delays: a piecewise delay method, Appl. Math. Comput. 208 (2009) 249–259.
[26]O.M. Kwon, J.H. Park, Improved delay-dependent stability criterion for neural networks with time-varying delays, Phys. Lett. A 373 (2009) 529–535. [27]J. Tian, X. Xie, New asymptotic stability criteria for neural networks with
time-varying delay, Phys. Lett. A 374 (2010) 938–943.
[28]J. Tian, S. Zhong, Improved delay-dependent stability criterion for neural networks with time-varying delay, Phys. Lett. A 373 (2009) 529–535. [29]H. Li, H. Gao, P. Shi, Passivity analysis for neural networks with discrete and
distributed delays, IEEE Trans. Neural Netw. 22 (2010) 1842–1847. [30]H. Zhang, F. Yang, X. Liu, Stability analysis for neural networks with
time-varying delay based on quadratic convex combination, IEEE Trans, Neural Netw. Learn. Syst 24 (2013) 513–521.
[31]O.M. Kwon, J.H. Park, S.M. Lee, E.J. Cha, Analysis on delay-dependent stability for neural networks with time-varying delays, Neurocomputing 103 (2013) 114–120.
[32]A. Seuret, F. Gouaisbaut, Wirtinger-based integral inequality: Application to time-delay systems, Automatica 49 (2013) 2860–2866.
[33]R.E. Skelton, T. Iwasaki, K.M. Grigoradis, A Unified Algebraic Approach to Linear Control Design, Taylor & Francis, New York, 1997.
[34]J. Liu, J. Zhang, Note on stability of discrete-time time-varying delay systems, IET Control Theory Appl. 6 (2012) 335–339.
[35]B. Yang, R. Wang, P. Shi, G.M. Dimirovski, New delay-dependent stability cri-teria for recurrent neural networks with time-varying delays, Neurocomput-ing 151 (2015) 1414–1422.
[36]P. Park, A delay-dependent stability criterion for systems with uncertain linear state-delayed systems, IEEE Trans. Autom. Control 35 (1999) 876–877. [37]K. Gu, An integral inequality in the stability problem of time-delay systems,
Proc. IEEE Conf. Decis. Control 3 (2000) 2805–2810.
[38]Y. He, M. Wu, J.H. She, G.P. Liu, Delay-dependent robust stability criteria for uncertain neutral systems with mixed delays, Syst. Control Lett. 51 (2004) 57–75.
[39]J. Qiu, Y. Wei, H.R. Karimi, New approach to delay-dependent H-infinity con-trol for continuous-time Markovian jump systems with time-varying delay and deficient transition descriptions, J. Frankl. Inst. 352 (2015) 189–215. [40] H.B. Zeng, Y. He, M. Wu, J. She, Free-matrix-based integral inequality for
sta-bility analysis of systems with time-varying delay, IEEE Trans. Autom. Control
http://dx.doi.org/10.1109/TAC.2015.2404271.
[41]H.B. Zeng, Y. He, M. Wu, J. She, Stability analysis of generalized neural net-works with time-varying delays via a new integral inequality, Neuro-computing 161 (2015) 148–154.
[42]L. Zhao, H. Gao, H.R. Karimi, Robust stability and stabilization of uncertain T–S fuzzy systems with time-varying delay: An input–output approach, IEEE Trans. Fuzzy Syst. 21 (2013) 883–897.
[43]J. Qiu, G. Feng, H. Gao, Fuzzy-model-based piecewise H1 static-output-feedback controller design for networked nonlinear Systems, IEEE Trans. Fuzzy Syst. 18 (2010) 919–934.
[44]B. Chen, X.-P. Liu, S.-C. Tong, and C. Lin, Observer-based stabilization of T–S fuzzy systems with input delay, IEEE Trans. Fuzzy Syst. 16 (2008) 652–663. [45]J. Qiu, G. Feng, H. Gao, Observer-based piecewise affine output feedback
controller synthesis of continuous-time T–S fuzzy affine dynamic systems using quantized measurements, IEEE Trans. Fuzzy Syst. 20 (2012) 1046–1062. [46]J. Qiu, G. Feng, H. Gao, Static-output-feedback H1 control of continuous-time T–S fuzzy affine systems via piecewise Lyapunov functions, IEEE Trans. Fuzzy Syst. 21 (2013) 245–261.
[47]C. Peng, D. Yue, Y.-C. Tian, New approach on robust delay dependent H1 control for uncertain T–S fuzzy systems with interval time-varying delay, IEEE Trans. Fuzzy Syst. 17 (2009) 890–900.
[48]M. Wu, Y. He, J.-H. She, Stability Analysis and Robust Control of Time-Delay Systems, Springer-Verlag, Berlin, Germany, 2010.
Bin Yang received the PhD degree in Control Theory and Control Engineering from Northeastern University, Shenyang, China, in 1998. From December 1998 to November 2000, he was a Postdoctoral Research Fellow with the Huazhong University of Science and Technol-ogy. He is currently an Associate Professor with the School of Control Science and Engineering, Dalian University of Technology, Dalian, China. His main research interests include time-delay systems, cellular neural networks, networked control systems, robust control.
Rui Wang received the B.S. and M.S. degrees in mathematics from Bohai University, Jinzhou, China, in 2001 and 2004, respectively, and the Ph.D. degree in control theory and applications from Northeastern University, Shenyang, China, in 2007. From March 2007 to December 2008, she was a Visiting Research Fellow with the University of Glamorgan, Pontypridd, U.K. She is currently an Associate Professor with the School of Aeronautics and Astronautics, Dalian University of Technology, Dalian, China. Her main research interests include switched systems, robust control, and net-worked control systems.
Dr Georgi Marko Dimirovski is a Research Professor (life-time) of Automation & Systems Engineering at the Faculty of Electrical-Electronics Engineering and Infor-mation Technologies of SS Cyril and Methodius Uni-versity of Skopje, R. Macedonia, and a Professor of Computer Science & Information Technologies at the Faculty of Engineering of Dogus University of Istanbul as well as an Invited Professor of Computer & Control Sciences at the Graduate Institutes of Istanbul Technical University, R. Turkey, and a‘Pro Universitas’ Professor at the Doctoral School of Obuda University in Budapest, Hungary. He is a Foreign Member of Serbian Academy of Engineering Sciences in Belgrade. He received his
Dipl.-Ing. degree in 1966 from SS Cyril and Methodius University of Skopje, Macedonia, M.Sc. degree in 1974 from University of Belgrade, R. Serbia, and Ph.D. degree in 1977 from University of Bradford, England, UK. He got his postdoctoral position in 1979 and subsequently was a Visiting Research Professor at the Uni-versity of Bradford in 1984, 1986 and 1988 as well as at the UniUni-versity of Wolver-hampton in 1990 and 1991. He was a Senior Research Fellow and Visiting Professor at Free University in Brussels, Belgium, in 1994 and also at Johannes Kepler Uni-versity in Linz, Austria, in 2000. His research interests include nonlinear systems and control, complex dynamical networks, switched systems, and applied com-putational intelligence to decision and control systems. Currently, as an associate editor, he serves Journal of the Franklin Institute, Asian J. of Control, and Intl. J. of Automation & Computing.