徐曉惠, 張繼業(yè), 施繼忠, 任松濤
(1.西華大學(xué) 汽車與交通學(xué)院,610039 成都;2. 牽引動(dòng)力國家重點(diǎn)實(shí)驗(yàn)室(西南交通大學(xué)),610031 成都; 3.浙江師范大學(xué) 工學(xué)院, 321004 浙江 金華)
?
脈沖干擾時(shí)滯復(fù)值神經(jīng)網(wǎng)絡(luò)的穩(wěn)定性分析
徐曉惠1, 張繼業(yè)2, 施繼忠3, 任松濤1
(1.西華大學(xué) 汽車與交通學(xué)院,610039 成都;2. 牽引動(dòng)力國家重點(diǎn)實(shí)驗(yàn)室(西南交通大學(xué)),610031 成都; 3.浙江師范大學(xué) 工學(xué)院, 321004 浙江 金華)
摘要:為分析脈沖干擾因素對(duì)復(fù)值神經(jīng)網(wǎng)絡(luò)動(dòng)態(tài)行為的影響,研究一類具有混合時(shí)滯和脈沖干擾的復(fù)值神經(jīng)網(wǎng)絡(luò)的平衡點(diǎn)的全局指數(shù)穩(wěn)定性. 在假定神經(jīng)元狀態(tài)、激活函數(shù)以及關(guān)聯(lián)矩陣定義在復(fù)數(shù)域的情況下,利用M矩陣?yán)碚摗⑾蛄縇yapunov函數(shù)法以及數(shù)學(xué)歸納法,分析確保該系統(tǒng)平衡點(diǎn)的存在性、唯一性以及全局指數(shù)穩(wěn)定性的充分條件,并給出了指數(shù)收斂率,最后通過一個(gè)數(shù)值仿真算例驗(yàn)證了所得結(jié)論的正確性. 結(jié)果表明:時(shí)滯和脈沖干擾均會(huì)降低神經(jīng)元狀態(tài)的指數(shù)收斂速度,所建立的穩(wěn)定性判據(jù)推廣了現(xiàn)有結(jié)論.
關(guān)鍵詞:復(fù)值神經(jīng)網(wǎng)絡(luò); 脈沖干擾; 混合時(shí)滯; 全局指數(shù)穩(wěn)定性; 矢量Lyapunov函數(shù)
近年來,復(fù)值神經(jīng)網(wǎng)絡(luò)在聯(lián)想記憶[1]、模式識(shí)別[2]、優(yōu)化問題求解[3]等領(lǐng)域得到了越來越多的應(yīng)用,關(guān)于復(fù)值神經(jīng)網(wǎng)絡(luò)的其他應(yīng)用參見文獻(xiàn)[4]. 神經(jīng)網(wǎng)絡(luò)平衡點(diǎn)的存在性以及收斂性是將其進(jìn)行硬件設(shè)計(jì)的前提條件,因此對(duì)復(fù)值神經(jīng)網(wǎng)絡(luò)平衡點(diǎn)的動(dòng)力學(xué)行為的研究是非常必要的. 文獻(xiàn)[5]研究了一類具有固定時(shí)滯的遞歸復(fù)值神經(jīng)網(wǎng)絡(luò),并利用LMI方法給出了判定該系統(tǒng)全局穩(wěn)定的充分條件. 文獻(xiàn)[6]研究了一類離散復(fù)值神經(jīng)網(wǎng)絡(luò),并給出了判定平衡點(diǎn)存在性、唯一性和指數(shù)穩(wěn)定的判定定理. 文獻(xiàn)[7]在假設(shè)復(fù)值激活函數(shù)關(guān)于神經(jīng)元狀態(tài)分別滿足有界或Lipschitz條件的情況下,利用LMI方法研究了一類具有固定時(shí)滯的復(fù)值神經(jīng)網(wǎng)絡(luò)平衡點(diǎn)的動(dòng)態(tài)行為. 文獻(xiàn)[8]研究了一類遞歸復(fù)值神經(jīng)網(wǎng)絡(luò)的多穩(wěn)態(tài)問題,但在模型中沒有考慮時(shí)滯. 時(shí)滯現(xiàn)象在實(shí)際系統(tǒng)中是不可避免的, 在神經(jīng)網(wǎng)絡(luò)中引入時(shí)間滯后參量,有利于移動(dòng)目標(biāo)的圖像處理、移動(dòng)物體速度的確定和模式分類. 文獻(xiàn)[9-10]在研究一類復(fù)值神經(jīng)網(wǎng)絡(luò)平衡點(diǎn)的多穩(wěn)態(tài)問題時(shí)在模型中引入了時(shí)滯,并得到了相應(yīng)的穩(wěn)定性充分判據(jù). 文獻(xiàn)[11-12]也初步研究了幾類具有混合時(shí)滯的復(fù)值神經(jīng)網(wǎng)絡(luò)的平衡點(diǎn)的動(dòng)態(tài)行為,并利用向量Lyapunov函數(shù)法的得到了確保系統(tǒng)平衡點(diǎn)存在性、唯一性以及指數(shù)穩(wěn)定性的充分判據(jù). 文獻(xiàn)[5-12]所考慮的復(fù)值神經(jīng)網(wǎng)絡(luò)都是確定型系統(tǒng),在實(shí)現(xiàn)復(fù)值神經(jīng)網(wǎng)絡(luò)硬件系統(tǒng)中,由于頻率轉(zhuǎn)換或者開關(guān)閉合等操作,使得系統(tǒng)的狀態(tài)在某些離散時(shí)刻會(huì)發(fā)生瞬間跳變,即系統(tǒng)狀態(tài)受到脈沖干擾. 關(guān)于具有脈沖干擾的實(shí)值神經(jīng)網(wǎng)絡(luò)的動(dòng)態(tài)行為分析,文獻(xiàn)[13-17]已經(jīng)作了大量的研究,然而目前尚未有學(xué)者對(duì)具有脈沖干擾的混合時(shí)滯復(fù)值神經(jīng)網(wǎng)絡(luò)平衡點(diǎn)的動(dòng)態(tài)行為進(jìn)行過相關(guān)研究. 基于以上分析,本文將在一類復(fù)值神經(jīng)網(wǎng)絡(luò)模型中同時(shí)考慮混合時(shí)滯和脈沖干擾,利用向量Lyapunov函數(shù)法和數(shù)學(xué)歸納法,研究該系統(tǒng)平衡點(diǎn)的模的全局指數(shù)穩(wěn)定性,并得到確保系統(tǒng)全局指數(shù)穩(wěn)定的充分條件.
1模型描述、基本假設(shè)以及引理

(1)
其中:zk∈C為第k個(gè)神經(jīng)元狀態(tài),k=1,2,…,n,m∈Ν,n為神經(jīng)元個(gè)數(shù);Δzk(tm)為在離散時(shí)刻tm系統(tǒng)狀態(tài)的突變量,離散集{tm}滿足0≤t0
假設(shè)連續(xù)函數(shù)θkj:[0,+)→[0,+),滿足

(2)
其中:μkj(β)為[0,δ)上的連續(xù)函數(shù),且μkj(0)=1,這里δ>0.
假設(shè)系統(tǒng)(1)的初始條件為zk(s)=φk(s),其中φk(s)為(-,0]上的有界連續(xù)函數(shù). 令為系統(tǒng)(1)的平衡點(diǎn).
定義1若存在常數(shù)Γ>0和λ>0,對(duì)所有U∈Cn及t≥0,有‖z(t)-z#‖≤sups∈(-,0]‖φ(s)-z#‖Γexp(-λt)成立,則稱系統(tǒng)(1)的平衡點(diǎn)z#是全局指數(shù)穩(wěn)定的.
假設(shè)1假設(shè)激活函數(shù)fk(·)滿足全局Lipschitz條件,即存在Lipschitz常數(shù)lk>0,使得對(duì)所有zk,vk∈C,有|fk(zk)-fk(vk)|≤lk|zk-vk|成立,k=1,2,…,n. 令L=diag(l1,l2,…,ln).

引理1[11]對(duì)于矩陣A=(akj)n×n∈Rn×n,如果所有非對(duì)角元素akj≤0,k≠j,則下面陳述是等價(jià)成立的:a) A為M矩陣;b) A的各階順序主子式均為正;c) 存在ξ∈Rn>0,使得Aξ>0;d) A的所有特征根的實(shí)部為正.
2主要結(jié)論

μkj(λ)|pkj|)·lj]<0 .
(3)
這里k=1,2,…,n,m∈N. 那么系統(tǒng)(1) 針對(duì)任意外部常輸入U(xiǎn)∈Cn,均存在唯一平衡點(diǎn)z#,且該平衡點(diǎn)是全局指數(shù)穩(wěn)定的,指數(shù)收斂率為0.5(λ-η).

(4)
其中k,j=1,2,…,n且k≠j.


(5)
定義曲線ζ={ω(χ):ωk=ξkχ,χ>0,k=1,2,…,n}和集合Ω(ω)={h:0≤h≤ω,ω∈ζ}. 顯然當(dāng)χ>χ′, Ω(ω(χ))?Ω(ω(χ′)). 令ξmax=max1≤k≤n{ξk},ξmin=min1≤k≤n{ξk},χ0=δ‖ψ‖2/ξmin,其中δ>1為一個(gè)常數(shù),則


這與假設(shè)D+Vk(t*)≥0是矛盾的. 因此有Vk(t)<ξkχ0,即|zk(t)|<(2ξkχ0exp(-λt))0.5,k=1,2,…,n,0 接下來,采用數(shù)學(xué)歸納法證明: k=1,2,…,n,tm-1≤t (6) 當(dāng)m=1時(shí),|zk(t)|2<2η0ξkχ0exp(-λt),k=1,2,…,n,t0≤t (7) 由于ηm≥1,進(jìn)而不等式(7)變?yōu)?/p> (8) 進(jìn)一步可以得到下面的不等式成立,即 k=1,2,…,n,tm≤t (9) 若上式不成立,那么存在某個(gè)子系統(tǒng)k′和時(shí)刻t′,使得D+Vk′(t′)≥0以及 tm≤t′ tm-τ 這與假設(shè)D+Vk′(t′)≥0是矛盾的. 因此不等式(9)是成立的. 根據(jù)數(shù)學(xué)歸納法,有 k=1,2,…,n,tm-1≤t (10) |zk(t)|2<2exp(η(t1-t0))exp(η(t2-t1))…exp(η(tm-1-tm-2))ξkχ0exp(-λt)=2exp(η(tm-1-t0))ξkχ0exp(-λt)<2ξkχ0exp(-(λ-η)(t-t0)),tm-1≤t 進(jìn)一步,有 ‖z(t)‖<(2δ‖ψ‖2ξmax/ξmin)0.5exp(-0.5(λ-η)(t-t0))=?!住琫xp(-0.5(λ-η)(t-t0)), 其中Γ=(2δξmax/ξmin)0.5. 通過定理1所建立的穩(wěn)定性判據(jù),可得出如下結(jié)論:1) 對(duì)于系統(tǒng)(1),當(dāng)系統(tǒng)中沒有脈沖干擾因素時(shí),該模型與文獻(xiàn)[11]中所研究的模型是相同的. 判定該系統(tǒng)平衡點(diǎn)的存在性、唯一性以及全局指數(shù)穩(wěn)定性的充分條件是假設(shè)1成立且矩陣Q為M矩陣. 該結(jié)論即為文獻(xiàn)[11]中的定理1和定理2. 本文所建立的判據(jù)推廣了現(xiàn)有結(jié)論. 2) 文獻(xiàn)[5,8,10-11]以及文獻(xiàn)[7]中的定理2在研究各類復(fù)值神經(jīng)網(wǎng)絡(luò)的動(dòng)態(tài)行為時(shí),繼續(xù)沿用了分析實(shí)值神經(jīng)網(wǎng)絡(luò)動(dòng)態(tài)行為的方法,即采用了將復(fù)值神經(jīng)網(wǎng)絡(luò)系統(tǒng)分解成實(shí)部系統(tǒng)和虛部系統(tǒng)的方法,得到了確保系統(tǒng)實(shí)部狀態(tài)和虛部狀態(tài)穩(wěn)定的充分判據(jù). 本文在研究該復(fù)值系統(tǒng)時(shí),并沒有對(duì)系統(tǒng)進(jìn)行實(shí)部與虛部的拆分,所建立的穩(wěn)定性判據(jù)為神經(jīng)元狀態(tài)的模的全局指數(shù)穩(wěn)定性. 此外,文獻(xiàn)[9]和文獻(xiàn)[7]中定理3也給出了判定一類復(fù)值神經(jīng)網(wǎng)絡(luò)系統(tǒng)神經(jīng)元狀態(tài)的模的穩(wěn)定性的充分條件,但在系統(tǒng)模型中未考慮無窮時(shí)滯和脈沖干擾因素. 3) 當(dāng)系統(tǒng)(1)中僅含有可變時(shí)滯或者無窮時(shí)滯時(shí),令定理1的不等式條件(3)中的bkj=0或pkj=0,其中k,j=1,2,…,n,其他假設(shè)條件不變,便可得到確保相應(yīng)系統(tǒng)平衡點(diǎn)全局指數(shù)穩(wěn)定的充分條件. 3算例 考慮如下復(fù)值神經(jīng)網(wǎng)絡(luò) (11) 其中z1(t)=x1(t)+y1(t)i,z2(t)=x2(t)+y2(t)i. 加權(quán)矩陣分別為 激活函數(shù)為 脈沖發(fā)生時(shí)刻為{0.2 s, 0.4 s, 0.6 s, 0.8 s,…},且 經(jīng)計(jì)算,有l(wèi)1=0.50,l2=0.25,η=6.93. 令系統(tǒng)(11)中的時(shí)延為τ1j=0.025-0.015sint,τ2j=0.03-0.01cost,j=1,2,t≥0. 令 令初始條件為 進(jìn)一步計(jì)算有 μ1j(λ)|p1j|)·lj]=-0.441<0, μ2j(λ)|p2j|)·lj]=-1.041<0. 根據(jù)定理1可得結(jié)論:系統(tǒng)(11)存在唯一平衡點(diǎn),且該平衡點(diǎn)是指數(shù)穩(wěn)定的,指數(shù)收斂率為0.535. 關(guān)于系統(tǒng)(11)的仿真結(jié)果見圖1、2,仿真結(jié)果驗(yàn)證了以上結(jié)論. 圖1 系統(tǒng)(11)狀態(tài)的模曲線 圖2 系統(tǒng)(11)狀態(tài)的實(shí)部曲線和虛部曲線 4結(jié)論 1)針對(duì)一類具有脈沖干擾的混合時(shí)滯復(fù)值神經(jīng)網(wǎng)絡(luò),在沒有沿用實(shí)值神經(jīng)網(wǎng)絡(luò)研究方法的情況下,對(duì)其平衡點(diǎn)的模的全局指數(shù)穩(wěn)定性進(jìn)行分析. 利用M矩陣?yán)碚?、向量Lyapunov函數(shù)法以及數(shù)學(xué)歸納法,得到了確保該系統(tǒng)平衡點(diǎn)的模全局指數(shù)穩(wěn)定性的充分條件. 2)穩(wěn)定性判據(jù)同時(shí)顯示出了時(shí)滯和脈沖干擾對(duì)系統(tǒng)平衡點(diǎn)指數(shù)收斂速度的影響,即時(shí)滯越大,脈沖干擾越強(qiáng)烈,系統(tǒng)神經(jīng)元狀態(tài)收斂的速度越慢. 所取得的研究成果推廣了現(xiàn)有結(jié)論. 3)通過數(shù)值算例驗(yàn)證了得的結(jié)論的可行性,同時(shí)算例仿真結(jié)果也顯示該結(jié)論的正確性. 參考文獻(xiàn) [1] LEE D L. Improvement of complex-valued Hopfield associative memory by using generalized projection rules [J]. IEEE Transactions on Neural Networks, 2006, 17(5): 1341-1347. [2] NAIT-CHARIF H. Complex-valued neural networks fault tolerance in pattern classification applications [C] //Proceedings of the Second WRI Global Congress on Intelligent Systems. Wuhan: IEEE, 2010: 154-157. [3] JIANG D. Complex-valued recurrent neural networks for global optimization of beamforming in multi-symbol MIMO communication systems [C] //Proceedings of International Conference on Conceptual Structurtion. Shanghai: Springer, 2008: 1-8.[4] HIROSE A. Recent progress in applications of complex-valued neural networks [C] // Proceedings of 10th International Conference on Artificiality Intelligence Soft Computing II. Zakopane: Springer, 2010: 42-46. [5] ZHANG Ziye, LIN Chong, CHEN Bing. Global stability criterion for delayed complex-valued recurrent neural networks [J]. IEEE Transactions on Neural Networks and Learning Systems, 2014, 25(9): 1704-1708. [6] SREE H, MURTHY G. Global dynamics of a class of complex valued neural networks [J]. International Journal Neural Systems, 2008, 18(2): 165-171. [7] HU Jin, WANG Jun. Global stability of complex-valued recurrent neural networks with time-delays [J]. IEEE Transactions on Neural Networks and Learning Systems, 2012, 23(6): 853-864. [8] HUANG Yujiao, ZHANG Huaguang, WANG Zhanshan. Multistability of complex-valued recurrent neural networks with real-imaginary-type activation functions [J]. Applied Mathematics and Computation, 2014, 229: 187-200. [9] ZHAO Zhenjiang, SONG Qiankun. Global exponential stability of complex-valued neural networks with time-varying delays on time scales [C] // Proceedings of the 33rd Chinese Control Conference. Nanjing: Control Systems IEEE, 2014: 5080-5085. [10]ZHOU Bo, SONG Qiankun. Boundedness and complete stability of complex-valued neural networks with time delay [J]. IEEE Transactions on Neural Networks and Learning Systems, 2013, 24(8): 1227-1238. [11]徐曉惠, 張繼業(yè), 趙玲. 一類混合時(shí)滯復(fù)值神經(jīng)網(wǎng)絡(luò)的動(dòng)態(tài)行為分析[J]. 西南交通大學(xué)學(xué)報(bào), 2014, 49(3): 470-476. [12]XU Xiaohui,ZHANG Jiye, SHI Jizhong. Exponential stability of complex-valued neural networks with mixed delays [J]. Neurocomputing, 2014, 128: 483-490. [13]SONG Qiankun, ZHANG Jiye. Global exponential stability of impulsive Cohen-Grossberg neural network with time-varying delays [J]. Nonlinear Analysis: Real World Applications,2008, 9(2): 500-510. [14]LI Liangliang, JIAN Jigui. Exponential convergence and Lagrange stability for impulsive Cohen-Grossberg neural networks with time-varying delays [J]. Journal of Computational and Applied Mathematics, 2015, 277: 23-35. [15]TOJTOVSKA B, JANKOVIC S. On some stability problems of impulsive stochastic Cohen-Grossberg neural networks with mixed time delays [J]. Applied Mathematics and Computation, 2014, 239: 211-226. [16]楊治國, 黃玉梅. 具有混合時(shí)滯的脈沖Cohen-grossberg神經(jīng)網(wǎng)絡(luò)的指數(shù)耗散性[J]. 四川大學(xué)學(xué)報(bào)(自然科學(xué)版), 2010, 47(3): 464-468. [17]施繼忠, 徐曉惠, 張繼業(yè). 擴(kuò)散反應(yīng)脈沖Cohen-Grossberg神經(jīng)網(wǎng)絡(luò)的魯棒穩(wěn)定性[J]. 西南交通大學(xué)學(xué)報(bào), 2010, 45(4): 596-602. (編輯魏希柱) Stability analysis of delayed complex-valued neural networks with impulsive disturbances XU Xiaohui1, ZHANG Jiye2, SHI Jizhong3, REN Songtao1 (1.School of Automobile and Transportation, Xihua University, 610039 Chengdu, China; 2.National Traction Power Laboratory(Southwest Jiaotong University), 610031 Chengdu, China; 3. College of Engineering, Zhejiang Normal University, 321004 Jinhua, Zhejiang,China) Abstract:To investigate the effect of impulsive disturbances on the dynamical behavior of the equilibrium point of complex-valued neural networks, the globally exponential stability of a class of the system with mixed delays and impulsive disturbances was studied in this paper. Assume that the neuron states, activation functions and interconnected matrix were defined in the complex domain. Some sufficient conditions for assuring the existence, uniqueness and globally exponential stability of the equilibrium point of the system were obtained by applying the M matrix theory, the mathematical induction and the vector Lyapunov function methods. Meanwhile, the exponential convergence rate was proposed. It can be concluded from the established sufficient conditions that the exponential convergence rate of the neurons is reduced by both time delays and the impulsive disturbances. The stability criteria established in this paper generalize the existing results. Finally, a numerical example with simulations was given to show the correctness of the obtained results. Keywords:complex-valued neural networks; impulsive disturbances; mixed delays; globally exponential stability; vector Lyapunov function 中圖分類號(hào):TP391 文獻(xiàn)標(biāo)志碼:A 文章編號(hào):0367-6234(2016)03-0166-05 通信作者:徐曉惠,xhxu@163.com. 作者簡(jiǎn)介:徐曉惠(1982—),女,副教授;張繼業(yè)(1965—),男,教授,博士生導(dǎo)師. 基金項(xiàng)目:國家自然科學(xué)基金(11402214, 51375402, 11572264, 61273021);四川省青年科技創(chuàng)新研究團(tuán)隊(duì)專項(xiàng)計(jì)劃 (2015TD0021);教育部“春暉計(jì)劃”合作科研項(xiàng)目(Z2014075);浙江省自然科學(xué)基金(LY14E08006). 收稿日期:2014-12-10. doi:10.11918/j.issn.0367-6234.2016.03.028

















