数学物理学报, 2026, 46(6): 2418-2432

具有时变时延与状态依赖切换的分数阶 Cohen-Grossberg 神经网络的多稳定性

董红勋,1, 万里光,1,*, 吴爱龙,2

1 湖北师范大学电气工程与自动化学院 湖北黄石 435002

2 湖北师范大学数学与统计学院 湖北黄石 435002

Multistability of Fractional-Order Cohen-Grossberg Neural Networks with Time-Varying Delays and State-Dependent Switching

Dong Hongxun,1, Wan Liguang,1,*, Wu Ailong,2

1 School of Electrical Engineering and Automation, Hubei Normal University, Hubei Huangshi 435002

2 School of Mathematics and Statistics, Hubei Normal University, Hubei Huangshi 435002

通讯作者: 万里光, E-mail: wanliguang@hbnu.edu.cn

收稿日期: 2025-04-14   修回日期: 2025-06-23  

基金资助: 湖北省自然科学基金(2024AFB837)
国家自然科学基金(62476082)
湖北省教育厅科研计划指导项目(B2023133)

Received: 2025-04-14   Revised: 2025-06-23  

Fund supported: NSF of Hubei Province(2024AFB837)
NSFC(62476082)
Guiding Project of Scientific Research Plan of Hubei Provincial Department of Education(B2023133)

作者简介 About authors

董红勋,E-mail:dhx3324@163.com;

吴爱龙,E-mail:hbnuwu@yeah.net

摘要

该文主要研究具有时变时延与状态依赖切换的分数阶 Cohen-Grossberg 神经网络 (SFCGNNs) 的多平衡点的共存性与稳定性. 利用状态空间划分的方法获得 SFCGNNs 具有 $5^n$ 个平衡点的存在性. 为了证明其中 $3^n$ 个平衡点是渐近稳定的, 采用 Lyapunov 方法, 此外还给出其余平衡点不稳定的充分条件. 最后通过两个数值算例验证结果的有效性.

关键词: 多稳定性; 分数阶 Cohen-Grossberg 神经网络; 时变时延; 状态依赖切换

Abstract

This paper primarily investigates the coexistence and stability of multiple equilibrium points for fractional-order Cohen-Grossberg neural networks (SFCGNNs) with time-varying delays and state-dependent switching. By employing the state space partition method, the existence of $5^n$ equilibrium points for SFCGNNs is established. To demonstrate the asymptotic stability of $3^n$ equilibrium points among them, the Lyapunov method is adopted. Additionally, sufficient conditions are derived to confirm the instability of the remaining equilibrium points. Finally, two numerical examples are provided to validate the effectiveness of the proposed results.

Keywords: multistability; fractional-order Cohen-Grossberg neural networks; time-varying delays; state-dependent switching

PDF (771KB) 元数据 多维度评价 相关文章 导出 EndNote| Ris| Bibtex  收藏本文

本文引用格式

董红勋, 万里光, 吴爱龙. 具有时变时延与状态依赖切换的分数阶 Cohen-Grossberg 神经网络的多稳定性[J]. 数学物理学报, 2026, 46(6): 2418-2432

Dong Hongxun, Wan Liguang, Wu Ailong. Multistability of Fractional-Order Cohen-Grossberg Neural Networks with Time-Varying Delays and State-Dependent Switching[J]. Acta Mathematica Scientia, 2026, 46(6): 2418-2432

1 引言

神经网络作为处理复杂问题的有效工具, 在模式识别, 参数优化, 联想记忆与信号处理[1,2,3,4] 等领域取得了广泛应用. 相较于仅存在单一全局稳定平衡点的单稳定性系统, 多稳定性系统在有限状态空间内存在多个局部稳定平衡点. 这一特性使得多平衡点的存在性验证与稳定性分析往往更具挑战性. 而且, 多平衡点的数目和位置与神经网络的存储和检索能力密切相关, 是模式识别与联想记忆等领域最重要的性能指标. 因此神经网络多稳定性的研究得到了广泛关注, 如文献[5] 讨论了具有不连续, 非单调, 分段且非线性激活函数的神经网络多稳定性, 文献[6] 则关注到无界时变延迟和周期激活函数对递归神经网络多稳定性的影响, 更多的神经网络多稳定性方向的研究工作可见参考文献[7,8,9,10].

作为一类广义递归神经网络, Cohen-Grossberg 神经网络自从在文献[11] 中被提出, 其优越的性能就得到了广泛关注, 如多种时延影响下的稳定性分析[12,13,14],

鲁棒性分析[15]和渐近周期解[16]. 此外, 文献[17] 讨论了混合时延影响下的 Cohen-Grossberg 神经网络的多模态稳定, 而文献[18] 研究了基于忆阻器的 Cohen-Grossberg 神经网络的多稳定性, 这为不连续的 Cohen-Grossberg 神经网络的稳定性研究提供了一个新的思路.

相比较于传统的整数阶微积分, 分数阶微积分可以更精确地描述实际演变过程. 换而言之, 利用分数阶微积分构造的动力学方程在物理系统, 工程系统与自动控制系统的建模中具有更高的建模精度. 这是因为分数微积分的记忆特性可以更好的描述时间内状态的总体行为. 这促使许多研究工作探讨了分数阶神经网络的多种动态行为, 如单稳定性与多稳定性[19], 多重渐近稳定性, $S$-渐近 $\omega$ 周期[20,21], Mittag-Leffler

输入稳定性[22]与预定时间同步[23]. 更多的分数阶 Cohen-Grossberg 神经网络的多稳定性的相关研究参见文献[24,25,26]. 需要注意的是上述参考文献中都没有考虑到切换在神经网络动态行为中的作用.

众所周知, 在工程, 化学, 生物医学等诸多领域中存在着许多具有状态依赖切换的系统, 如基于忆阻器的神经网络, 其连接权值会随着状态的变化而发生跳变. 而切换点的存在可能会增加平衡点的数目, 这代表着神经网络的存储容量和检索能力都会进一步得到增强[27], 因此对切换神经网络的研究工作是很有必要的. 如何结合状态依赖切换神经网络和分数阶神经网络的优点, 激发了很多研究者的热情,

如文献[28,29,30]分别考虑 Sigmoidal 函数与 Gaussian 函数作为激活函数时, 状态依赖切换对于分数阶神经网络多稳定性的影响.

值得注意的是, 上述文献均没有考虑到时延的问题. 在神经网络中时延一般表现为神经元的延迟激活, 这种异常激活会导致神经网络的非线性震荡乃至失稳. 文献[31,32]在递归神经网络中考虑了时延的影响, 分析了具有分段线性激活函数和 Sigmoidal 激活参数时的多稳定性. 现有的递归神经网络和 Hopfield 神经网络在结构上与 Cohen-Grossberg 神经网络相似, 均可视为 Cohen-Grossberg 神经网络中具有不同参数的一种特殊情况, 但关于具有时变时延和状态依赖切换的 Cohen-Grossberg 神经网络的多稳定性的研究较为稀少.

经由上述讨论, 本文研究的是具有时变时延与状态依赖切换的分数阶 Cohen-Grossberg 神经网络的多稳定性问题. 主要贡献概述如下, 首先本文考虑的分数阶系统中同时考虑的切换问题与时延问题更具一般性, 这是因为具有状态依赖的参数切换与时变时延的神经网络在实际工程问题中有着更广的适用范围. 其次, 本文使用状态空间划分的方法研究了 SFCGNNs, 所得到的推导结果包含了 Hopfield 神经网络与递归神经网络的研究结果. 第三, 利用 Brouwer's 不动点定理, 微分包含理论与 Lyapunov 稳定性理论, 给出了 SFCGNNs 具有 $5^n$ 个平衡点的充分条件, 其中 $3^n$ 个平衡点是渐近稳定的, 其余平衡点则是不稳定的. 相比较于前人的成果, 本文的结果是新的, 且更具一般性, 是对前人所做工作的概括与拓展.

本文其余内容结构如下: 在第 2 节中, 将介绍 SFCGNNs 的模型描述, 分数阶导数的定义与一些其他的预备知识; 第 3 节中, 利用状态空间划分的方法与分数阶 Caputo 导数分别推到了推导出了保证 SFCGNNs 具有 $5^n$ 个平衡点的充分条件, 证明了 $3^n$ 个平衡点是渐近稳定的, 同时也给出了其余的 $5^n-3^n$ 个平衡点不稳定的充分条件; 第 4 节中, 结合两个数值算例验证结果的有效性; 最后, 第 5 节是对本文的研究内容进行的总结.

2 预备知识

2.1 模型描述

考虑具有状态依赖切换的分数阶 Cohen-Grossberg 神经网络, 下面是其分数阶微分方程:

$\begin{equation} \begin{aligned} D^{q}x_{i}(t)&=\alpha_{i}(x_{i}(t))\left(\right.-\mu_{i}(x_{i}(t))x_{i}(t) +\sum_{j=1}^{n}\delta_{i j}(x_{j}(t))\mathcal{G}_{j}(x_{j}(t))\\ &\quad +\sum_{j=1}^{n}\rho_{ij}(x_{j}(t_\tau))\mathcal{G}_{j}(x_{j}(t_\tau))+I_{i}\left.\right). \end{aligned} \end{equation}$

此处 $x=(x_1,x_2,\cdots,x_n)^T$ 是状态向量, $I_{i}$ 表示外部输入, $t_\tau=t+\tau(t)$, 其中 $\tau(t)$ 表示延迟时间, 且有 $-\tau\le \tau(t) <0$. $\alpha_{i}(x_{i}(t))$ 是增益, $\mu_{i}(x_{i}(t))$ 是自抑制函数, 且满足 $\alpha_{i}^{'}$, $\alpha_{i}^{''}$, $\mu _{i}^{'}$$\mu_{i}^{''}$ 均为大于 0 的常数. $\delta_{ij}(x_{j}(t))$$\rho_{ij}(x_{j}(t_\tau))$ 是连接权值, 其中 $\delta _{i}^{'}, \delta _{ij}^{''}, \rho _{ij}^{'}, \rho _{ij}^{''}$ 均为常数. 本文所讨论的是具有状态依赖的绝对值切换, 即当系统状态的绝对值达到阈值时系统参数会发生切换, 对于切换跳变 $\mathcal{K}_i>0,\mathcal{K}_j>0$ 满足以下切换规则:

$\begin{equation*} \begin{aligned} \alpha_i(x_i(t)) = & \begin{cases} \alpha_{i}^{'}, |x_i(t)| < \mathcal{K}_i, \\ \alpha_{i}^{''}, |x_i(t)| \geq \mathcal{K}_i, \end{cases} \qquad \mu_i(x_i(t)) = & \begin{cases} \mu_{i}^{'}, |x_i(t)| < \mathcal{K}_i, \\ \mu_{i}^{''}, |x_i(t)| \geq \mathcal{K}_i, \end{cases} \\ \delta_{ij}(x_j(t)) = & \begin{cases} \delta_{ij}^{'}, |x_j(t)| < \mathcal{K}_j, \\ \delta_{ij}^{''}, |x_j(t)| \geq \mathcal{K}_j, \end{cases} \qquad \rho_{ij}(x_j(t_\tau)) = & \begin{cases} \rho_{ij}^{'}, |x_j(t_\tau)| < \mathcal{K}_j, \\ \rho_{ij}^{''}, |x_j(t_\tau)| \geq \mathcal{K}_j. \end{cases} \end{aligned} \end{equation*}$

系统(2.1)的初值定义为 $x_i(s) =\phi _i(s-t_0),s\in[t_0-\tau,t_0],\phi\in C((-\tau,0],\Re)$. 并且其中激活函数满足以下假设:

假设 1: 激活函数 $ \mathcal{G}_i$ 是连续函数具有饱和区间, 即存在两个常数 $m_{i}$$M_{i}$ 使得 $\lim_{u\rightarrow+\infty}G_{i}(u)=M_{i}$, $\lim_{u\rightarrow -\infty}G_{i}(u)=m_{i} $.

假设 2: 对于 $\forall u\in \Re$, 存在正常数 $k_i$, 使得 $0<\mathcal{G}_{i}^{'}(u)\le k_i$, 在此基础上若有 $u\ne 0$, 则 $u\mathcal{G}_{i}^{''}(u) <0$.$\check{k}_i=\sup_{|u|\geq \mathcal{K}_i}\mathcal{G}_{i}^{'}(u)$, $ k_i=\sup_{|u|<\mathcal{K}_i}\mathcal{G}_{i}^{'}(u)$.

2.2 符号

定义 $\Re ^n$ 是 n 维实空间. 向量 $ v=(v_1,v_2,\cdots,v_n )^T\in \Re ^n$ 是一个列向量. 定义 $C^n((t_0,+\infty),\Re^n) $ 是将 $(t_0,+\infty)$ 区间映射到 $\Re^n$ 的 n 次连续函数映射. $[\cdot,\cdot] $ 是代表区间, co$ [a,b] $ 表示由实数 $a, b$ 构成的闭凸包, 同样的 co$[\omega]$ 代表由给定集合 $\omega$ 构成的凸包, 其中 $\omega \in \Re$. 符号 $\mathcal{D}^{q} h_{i}(t)$$\mathcal{I}^{q}h_{i}(t)$ 分别表示 $h_{i}(t)$ 的阶数为 $q$ 的 Caputo 分数阶导数与分数阶积分. 另外, 本篇论文所考虑的所有系统的解均为 Filippov 意义下的解.

2.3 定义与性质

对于 SFCGNNs (2.1), 考虑如下集值映射

$\begin{equation*} \begin{array}{l} {\rm co}\left[\alpha _i\left( x_i\left( t \right) \right)\right] =\left\{ \begin{matrix} \alpha _{i}^{'},& \left| x_i\left( t \right) \right|<\mathcal{K}_i,\\ {\rm co}\left[ \alpha _{ij}^{'},\alpha _{ij}^{''} \right],& \left| x_i\left( t \right) \right|=\mathcal{K}_i,\\ \alpha _{i}^{''},& \left| x_i\left( t \right) \right|>\mathcal{K}_i,\\ \end{matrix} \right.\\ {\rm co}\left[\mu _i\left( x_i\left( t \right) \right)\right] =\left\{ \begin{matrix} \mu _{i}^{'},& \left| x_i\left( t \right) \right|<\mathcal{K}_i,\\ {\rm co}\left[ \mu _{ij}^{'},\mu _{ij}^{''} \right],& \left| x_i\left( t \right) \right|=\mathcal{K}_i,\\ \mu _{i}^{''},& \left| x_i\left( t \right) \right|>\mathcal{K}_i,\\ \end{matrix} \right.\\ {\rm co}\left[\delta _{ij}\left( x_j\left( t \right) \right)\right] =\left\{ \begin{matrix} \delta _{ij}^{'},& \left| x_j\left( t \right) \right|<\mathcal{K}_j,\\ {\rm co}\left[ \delta _{ij}^{'},\delta _{ij}^{''} \right],& \left| x_j\left( t \right) \right|=\mathcal{K}_j,\\ \delta _{ij}^{''},& \left| x_j\left( t \right) \right|>\mathcal{K}_j,\\ \end{matrix} \right.\\ {\rm co}[\rho_{ij}( x_j(t_\tau)] =\left\{ \begin{matrix} \rho _{ij}^{'},& |x_j(t_\tau)|<\mathcal{K}_j,\\ {\rm co}\left[ \rho _{ij}^{'},\rho _{ij}^{''} \right],& | x_j(t_\tau)|=\mathcal{K}_j,\\ \rho _{ij}^{''},& |x_j(t_\tau)|>\mathcal{K}_j.\\ \end{matrix} \right. \\ \end{array} \end{equation*}$

通过微分包含理论, 根据 SFCGNNs (2.1), 对于 $i \in 1,\ldots,n$

$\begin{equation*} \begin{aligned} D^{q}x_i(t)\in {\rm co}[ \alpha _i(x_i(t))]\left( \right.&-{\rm co}[\mu _i(x_i(t))] x_i(t) +\sum_{j=1}^n{{\rm co}}[\delta _{ij}(x_j(t))] \mathcal{G}_j( x_j(t))\\ &+\sum_{j=1}^n{{\rm co}}[\rho_{ij}(x_j(t_\tau))] \mathcal{G}_j(x_j(t_\tau)) +I_i\left. \right). \end{aligned} \end{equation*}$

定义 2.1 ([33]) 列向量 $ x(t) = \left( x_1(t), \ldots, x_n(t) \right)^T$ 是 SFCGNNs (1) 具有初值 $ x_i\left( s \right) =\phi _i\left( s-t_0 \right),s\in \left[ t_0-\tau,t_0 \right] $ 的 Filippov 解, 如果对于 $ \forall i \in \left(1, \ldots, n \right) $, 有 $ x_i(t) $$ t \in [0, +\infty) $ 的任意闭区间上绝对连续并且满足以下微分包含

$\begin{equation*} \begin{aligned} D^{q}x_i(t)\in {\rm co}[\alpha _i(x_i(t))] \left( \right. &-{\rm co}[\mu_i(x_i(t) )]x_i(t)+\sum_{j=1}^n{{\rm co}}[\delta _{ij}( x_j(t))]\mathcal{G}_j(x_j(t))\\ &+\sum_{j=1}^n{{\rm co}}[\rho_{ij}(x_j(t_\tau))] \mathcal{G}_j(x_j(t_\tau))+I_i\left. \right). \end{aligned} \end{equation*}$

定义 2.2 ([33]) 常数列向量 $ x^*=\left(x^*_{1}, \ldots, x^*_{n}\right)^{T}$ 可以被称为 SFCGNNs (1) 的平衡点, 如果 $\forall i \in \mathbb{N} $, $ x^*_{i} $ 满足

$\begin{equation*} 0\in -{\rm co}[\mu _i( x_{i}^{*})] x_{i}^{*} +\sum_{j=1}^n{{\rm co}}[\delta_{ij}(x_{j}^{*})] \mathcal{G}_j(x_{j}^{*}) \sum_{j=1}^n{{\rm co}}[\rho_{ij}(x_{j}^{*})] \mathcal{G}_j(x_{j}^{*})+I_i. \end{equation*}$

引理 2.1 ([34]) 如果 $ y\left( t \right) \in \Re ^n $ 可微, $ q\in \left( 0,1 \right] $, 而且 Q 是正定的, 那么

$\begin{equation*} D^{q}\left(y^{T}(t) Q y(t)\right) \leq 2 y^{T}(t) Q D^{q} y(t). \end{equation*}$

引理 2.2 ([35]) 假设 $ 0<q<1 $ 并且 $ h(t) $ 是可微的, 如果存在一些时刻 $ 0<t<t_1 $ 使得 $ h\left( t_1 \right) =0 $ 并且 $ h(t)<0 $ ($ h(t)\le 0 $) 那么 $ D^{q} h(t_1)) >0$ ($ D^{q} h(t_1)) \ge 0 $).

引理 2.3 ([36]) 设 $V(t)$ 是非负可微函数, 且满足

$\begin{equation*} \left\{\begin{array}{l} D^{q} V(t) \leq-\varsigma V(t)+\varpi V(t-\tau(t)), \\ V(v)=\phi(v) \geq 0, v \in[-\tau, 0], \end{array} \right. \end{equation*}$

其中 $ 0<q<1 $ 而且 $ 0 \leq \tau(t) \leq \tau $.$ t>0 $, 如果 $\varsigma>|\varpi|>0 $, 则 $\lim _{t \rightarrow+\infty} V(t)=0 $.

本文中, 神经网络的切换机制均依赖于系统状态的绝对值. 以第 $i$ 维为例, 任选一个切换跳变 $\mathcal{K}_i>0$ 的值, 均可根据切换跳变前后系统参数的不同, 将状态空间以切换点为界分为五个部分: 即 $|x_i(t)|<\mathcal{K}_i$ 时, 此时区间记为 $(-\mathcal{K}_i,\mathcal{K}_i)$; 当 $|x_i(t)|=\mathcal{K}_i$ 时, 视 $(-\mathcal{K}_i)$$(\mathcal{K}_i)$ 为两个特殊区间; 当 $|x_i(t)|>\mathcal{K}_i$ 时, 区间记为 $( -\infty,-\mathcal{K}_i)$$(\mathcal{K}_i, +\infty)$. 不失一般性的, 每维神经网络均具有五个不相交的状态子空间, 规定如下区间表示法:

$\begin{equation*} \begin{array}{c} ( -\infty,-\mathcal{K}_i) =( -\infty,-\mathcal{K}_i)^1\times(-\mathcal{K}_i) ^0\times (-\mathcal{K}_i,\mathcal{K}_i)^0 \times (\mathcal{K}_i)^0 \times (\mathcal{K}_i, +\infty)^0,\\ (-\mathcal{K}_i)=(-\infty,-\mathcal{K}_i)^0\times(-\mathcal{K}_i)^1\times(-\mathcal{K}_i,\mathcal{K}_i)^0\times (\mathcal{K}_i)^0\times(\mathcal{K}_i,+\infty )^0,\\ \vdots \\ (\mathcal{K}_i,+\infty)=( -\infty,-\mathcal{K}_i)^0\times(-\mathcal{K}_i)^0\times ( -\mathcal{K}_i,\mathcal{K}_i)^0\times(\mathcal{K}_i)^0\times(\mathcal{K}_i,+\infty)^1.\\ \end{array} \end{equation*}$

为了方便表示, 记集合 $\Theta$

$\begin{equation*} \begin{array}{l} \Theta =\{\prod_{i=1}^n{(-\infty,-\mathcal{K}_i)}_{1}^{\sigma_{1}^{(i)}}\cup\{-\mathcal{K}_i\}^{\sigma_{2}^{(i)}}\cup(-\mathcal{K}_i,\mathcal{K}_i)^{\sigma _{3}^{(i)}}\cup\{\mathcal{K}_i\}^{\sigma_{4}^{(i)}}\cup(\mathcal{K}_i, +\infty)^{\sigma_{5}^{(i)}},\\ \,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\, (\sigma_{1}^{(i)},\sigma_{2}^{(i)},\cdots,\sigma_{5}^{(i)})\in\{(1,0,0,0,0) \cup(0,1,0,0,0)\cup(0,0,1,0,0),\\ \,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\cup(0,0,0,1,0)\cup(0,0,0,0,1)\},i=1,\cdots,n\}, \end{array} \end{equation*}$

显然 $\Theta$ 具有 $5^n$ 个元素, 此外记 $\Theta ^{(1)}$

$\begin{equation*} \begin{aligned} \Theta^{(1)}=\{&\prod_{i=1}^{n}[-E,-\mathcal{K}_{i}-\varepsilon]^{\sigma_{1}^{(i)}} \cup\{-\mathcal{K}_{i}\}^{\sigma_{2}^{(i)}} \cup[-\mathcal{K}_{i}+\varepsilon, \mathcal{K}_{i}-\varepsilon]^{\sigma_{3}^{(i)}}\cup\{\mathcal{K}_{i}\}_{4}^{\sigma_{4}^{(i)}} \cup[\mathcal{K}_{i}+\varepsilon, E]^{\sigma_{5}^{(i)}},\\ & (\sigma_{1}^{(i)}, \sigma_{2}^{(i)}, \cdots, \sigma_{5}^{(i)})= (1,0,0,0,0) \text { or }(0,1,0,0,0) \text { or }(0,0,1,0,0) \text { or } (0,0,0,1,0), \\ &\text { or }(0,0,0,0,1), i=1, \cdots, n\}. \end{aligned} \end{equation*}$

此处 $\varepsilon$ 是一个足够小的常数而 $E$ 是一个常数, 可知每个子区域 $\Theta(k)\in \Theta^{(1)}$ 都是有界闭集, $k\in (1,\ldots,5)$.

3 主要命题的证明

定义边界函数如下

$\begin{equation*} \begin{aligned} B_{i}^{+}= & \sum_{j=1,j \neq i}^{n} \max \{\delta_{i j}^{\prime \prime} m_{j}, \delta_{i j}^{\prime \prime} M_{j}, \delta_{i j}^{\prime} \mathcal{G}_{j}(-\mathcal{K}_{j}), \delta_{i j}^{\prime} \mathcal{G}_{j}(\mathcal{K}_{j})\}\\ &+\sum_{j=1}^{n}\max\{\rho_{i j}^{\prime \prime} m_{j}, \rho_{i j}^{\prime \prime} M_{j}, \rho_{i j}^{\prime} \mathcal{G}_{j}(-\mathcal{K}_{j}), \rho_{i j}^{\prime} \mathcal{G}_{j}(\mathcal{K}_{j})\}+I_{i},\\ B_{i}^{-}= & \sum_{j=1, j \neq i}^{n} \min \{\delta_{i j}^{\prime \prime} m_{j}, \delta_{i j}^{\prime \prime} M_{j}, \delta_{i j}^{\prime} \mathcal{G}_{j}(-\mathcal{K}_{j}), \delta_{i j}^{\prime} \mathcal{G}_{j}(\mathcal{K}_{j})\}\\ & +\sum_{j=1}^{n} \min \{\rho_{i j}^{\prime \prime} m_{j}, \rho_{i j}^{\prime \prime} M_{j}, \rho_{i j}^{\prime} \mathcal{G}_{j}(-\mathcal{K}_{j}),\rho_{i j}^{\prime} \mathcal{G}_{j}(\mathcal{K}_{j})\}+I_{i}. \end{aligned} \end{equation*}$

3.1 SFCGNNs 平衡点的存在性

定义 3.1 对于任意 $ i=1,\cdots,n $, 下述条件成立

$\begin{equation} \left\{ \begin{array}{c} \alpha_{i} ''\left( \mu _{i}^{''}\mathcal{K}_i+\delta _{ii}^{''}\mathcal{G}_i\left( -\mathcal{K}_i \right) +B_{i}^{+} \right) <0,\\ \alpha_{i} '\left( \mu _{i}^{'}\mathcal{K}_i+\delta _{ii}^{'}\mathcal{G}_i\left( -\mathcal{K}_i \right) +B_{i}^{-} \right) >0,\\ \alpha_{i} '\left( -\mu _{i}^{'}\mathcal{K}_i+\delta _{ii}^{'}\mathcal{G}_i\left( \mathcal{K}_i \right) +B_{i}^{+} \right) <0,\\ \alpha_{i} ''\left( -\mu _{i}^{''}\mathcal{K}_i+\delta _{ii}^{''}\mathcal{G}_i\left( \mathcal{K}_i \right) +B_{i}^{-} \right) >0.\\ \end{array} \right. \end{equation}$

那么系统(2.1)在每个子集 $\Theta(k) \in \Theta$ 中均存在至少一个平衡点, 即系统(2.1)具有 $5^n$ 个平衡点.

根据(3.1), 显然我们可以找到 $ \varepsilon>0 $$ E>0 $使得

$\begin{equation} \left\{ \begin{array}{c} -\mu _{i}^{''}u+\delta _{ii}^{''}\mathcal{G}_i\left( u \right) +B_{i}^{-}>0,\forall u\leq -E,\\ \mu _{i}^{''}\left( \mathcal{K}_i+\varepsilon \right) +\delta _{ii}^{''}\mathcal{G}_i\left( -\mathcal{K}_i-\varepsilon \right) +B_{i}^{+}<0,\\ \mu _{i}^{'}\left( \mathcal{K}_i-\varepsilon \right) +\delta _{ii}^{'}\mathcal{G}_i\left( -\mathcal{K}_i+\varepsilon \right) +B_{i}^{-}>0,\\ -\mu _{i}^{'}\left( \mathcal{K}_i-\varepsilon \right) +\delta _{ii}^{'}\mathcal{G}_i\left( \mathcal{K}_i-\varepsilon \right) +B_{i}^{+}<0\\ -\mu _{i}^{''}\left( \mathcal{K}_i+\varepsilon \right) +\delta _{ii}^{''}\mathcal{G}_i\left( \mathcal{K}_i+\varepsilon \right) +B_{i}^{-}>0,\\ -\mu _{i}^{''}u+\delta _{ii}^{''}\mathcal{G}_i\left( u \right) +B_{i}^{+}<0,\forall u\geq E,\\ \end{array} \right. \end{equation}$

$\Theta(k)=\prod_{i=1}^n{L_{ik}} \in \Theta^{(1)}$, 其中, $k\in ( 1,...,5 )$. 因为此处将切换点视为一个特殊集合, 此处 $L_{ik}$ 分为: (1) $L_{i1}=[-E,-\mathcal{K}_{i}-\varepsilon]$$L_{i5} = [\mathcal{K}_{i}+\varepsilon,E]$; (2) $L_{i3}=[-\mathcal{K}_{i}+\varepsilon,\mathcal{K}_{i}-\varepsilon]$; (3) $L_{i2}=[-\mathcal{K}_{i}]$; (4)$L_{i4}=[\mathcal{K}_{i}]$ 等四种情形分别讨论.

情形 1: $x_i=[-E,-\mathcal{K}_{i}-\varepsilon]$$x_i=[\mathcal{K}_{i}+\varepsilon,E]$. 首先定义函数

$\begin{equation*} B_{i}(u)= -\mu _{i}^{''}x_{i}^{*}+\delta _{ii}^{''}\mathcal{G}_i(x_i)+\sum_{j=1,j\ne i}^n{\delta_{ij}^{''}\mathcal{G}_j(x_{j}^{*})}+\sum_{j=1}^n{\rho_{ij}^{''}\mathcal{G}_j( x_{j}^{*})}+I_i, \end{equation*}$

根据(3.2)可知, 当 $x_i=[-E,-\mathcal{K}_{i}-\varepsilon]$ 时, $B_i(-E)B_i(-\mathcal{K}_i-\varepsilon)<0 $; 因此存在 $\hat{x}_i \in (-E,-\mathcal{K}_i-\varepsilon )$, 使得 $B_i(\hat{x}_i) =0$. 同理, 当 $x_i=[\mathcal{K}_{i}+\varepsilon,E]$ 时, 存在 $\hat{x}_i \in(\mathcal{K}_i+\varepsilon,E)$, 使得 $B_i(\hat{x}_i)=0$ 成立.

情形 2: $x_i=[-\mathcal{K}_{i}+\varepsilon, \mathcal{K}_{i}-\varepsilon]$. 类似于情形 1, 定义函数

$\begin{equation*} B_{i}(u)= -\mu _{i}^{'}x_{i}^{*}+\delta _{ii}^{'}\mathcal{G}_i(x_i) +\sum_{j=1,j\ne i}^n{\delta _{ij}^{'}\mathcal{G}_j(x_{j}^{*})}+\sum_{j=1}^n{\rho _{ij}^{'}\mathcal{G}_j(x_{j}^{*})}+I_i, \end{equation*}$

对于 $x_i =[-\mathcal{K}_{i}+\varepsilon,\mathcal{K}_{i}-\varepsilon]$, 由(3.2)可知, $B_i(-\mathcal{K}_{i}+\varepsilon) B_i( \mathcal{K}_i-\varepsilon)<0$, 即存在 $\hat{x}_i\in(-\mathcal{K}_i+\varepsilon,\mathcal{K}_i-\varepsilon)$, 使得 $B_i(\hat{x}_i)=0$ 成立.

情形 3: $x_i \in [-\mathcal{K}_{i}]$. 由(3.1)可知有如下不等式成立:

$\begin{equation} 0\in {\rm co}[\mu_i( -\mathcal{K}_i)] \mathcal{K}_i+\sum_{j=1}^n{\rm co}[\delta_{ij}(-\mathcal{K}_i)] \mathcal{G}_j(-\mathcal{K}_i)+\sum_{j=1}^n{\rm co}[\rho_{ij}(-\mathcal{K}_i)] \mathcal{G}_j(-\mathcal{K}_i )+I_i, \end{equation}$

即对于任意 $x\in \Theta(k)$, 如果平衡点的第 i 个分量 $ x_i=-\mathcal{K}_{i}$, 平衡点显然存在, 记 $\hat{x}_i=-\mathcal{K}_i$.

情形 4: $x_i=[\mathcal{K}_{i}]$. 由(3.1)可知式(3.4)成立

$\begin{equation} 0\in -{\rm co}[\mu _i(\mathcal{K}_i)]\mathcal{K}_i+\sum_{j=1}^n{\rm co}[\delta_{ij}(\mathcal{K}_i)] \mathcal{G}_j(\mathcal{K}_i)+\sum_{j=1}^n{\rm co}[\rho _{ij}(\mathcal{K}_i)] \mathcal{G}_j(\mathcal{K}_i)+I_i, \end{equation}$

同理, 如果平衡点的第 $i$ 个分量 $x_i=\mathcal{K}_{i}$, 平衡点显然存在, 记 $\hat{x}_i=\mathcal{K}_i$.

定义一个连续函数 $\mathscr{F}:\Theta(k) \rightarrow \Theta(k)$ 使得 $\mathscr{F}(x_1,\cdots,x_n)=(\hat{x}_1,\cdots,\hat{x}_n)$$\Theta(k)$ 是一个紧凸集, 基于布劳威尔不动点理论, 此处存在一个不动点 $x^*=(x_{1}^{*},\cdots,x_{n}^{*})$, 使得 $\mathscr{F}( x_{1}^{*},\cdots,x_{n}^{*}) =( x_{1}^{*},\cdots,x_{n}^{*})$, $ x^* $ 也是系统(2.1)位于 $\Theta(k) \subset \Theta^{(1)} $的平衡点. 前文已说明 $\Theta$ 具有 $5^n$ 个元素, 所以系统(2.1)具有 $5^n$ 个平衡点.

3.2 SFCGNNs 的正不变集

为了便于分析, 首先定义 $ \bar{\Theta} \in \Theta $ 具有如下形式:

$\begin{equation*} \begin{aligned} \bar{\Theta}=\{\prod_{i=1}^n{(-\infty,-\mathcal{K}_i -\varepsilon )}_{1}^{\sigma _{1}^{(i)}}\cup(-\mathcal{K}_i +\varepsilon,\mathcal{K}_i-\varepsilon)^{\sigma _{2}^{(i)}}\cup(\mathcal{K}_i+\varepsilon,+\infty)^{\sigma _{3}^{(i)}},\\ ( \sigma_{1}^{(i)},\sigma _{2}^{(i)},\sigma_{3}^{(i)})\in\{(0,0,1)\cup(0,1,0) \cup ( 1,0,0) \},i=1,\cdots,n\}. \end{aligned} \end{equation*}$

定理 3.2 如果条件(3.1)成立, 则每个 $ \Omega _{\sigma}\in \bar{\Theta}$ 都是正不变集.

取集合 $\Omega _{\sigma}$$ \bar{\Theta} $ 中的任一元素, 对于任意初始状态 $ x(t_0)\in\Omega_{\sigma}$, 与之相对应的 SFCGNNs 的解 $x_i(t)$ 保持在区域 $\Omega_{\sigma}$ 内. 则可以声称对所有 $t>t_0$ 都有 $x_i(t) \in \Omega_{\sigma}$. 如果此推论不成立, 那么将会存在一个时刻 $t^*>t_0$, 使得 $x_i(t^*)$$x_i(t)$ 第一个逃离区域 $\Omega _{\sigma}$ 的分量. 为此, 讨论以下三种情况:

情形 1: 对 $x_i\in(-\mathcal{K}_i +\varepsilon,\mathcal{K}_i-\varepsilon)$, 存在 $t_1\ge t_0$ 使得 $ x_i(t_1)=\mathcal{K}_i -\varepsilon$, 即 $x_i(t_1)-\mathcal{K}_i +\varepsilon=0$ 且有 $x_i(t)-\mathcal{K}_i+\varepsilon<0$($ t_0 <t<t_1$), 根据引理(2.2), 可知 $ D^{q} z(t_1)=x_i(t_1) -\mathcal{K}_i +\varepsilon >0$, 又有 $\alpha(x_i(t_1))>0$, 则根据性质(3.2)有

$ D^q z_i(t_1) =D^qx_i(t_1) <\alpha(x_i(t_1))(-\mu _{i}^{'}(\mathcal{K}_i-\varepsilon ) +\delta _{ii}^{'}\mathcal{G}_i( \mathcal{K}_i-\varepsilon)+B_{i}^{+})<0, $

此处产生了矛盾. 同理, 对于另一端点假设 $x_i(t_1)=-\mathcal{K}_i +\varepsilon$, 同样会产生矛盾. 因此可以说这意味着对于任意 $ t>t_0$, 如果 $x_{i}(t)$ 的初值在区间内, $x_i(t)$ 不会离开区间 $( -\mathcal{K}_i+\varepsilon,\mathcal{K}_i -\varepsilon)$.

情形 2: 对 $x_i \in( -\infty,-\mathcal{K}_i -\varepsilon)$, 存在 $t_2 \ge t_0 $, 使得 $x_i(t_2)=-\mathcal{K}_i -\varepsilon$, 即 $x_i(t_2)+\mathcal{K}_i +\varepsilon=0$ 且有 $ x_i(t)+\mathcal{K}_i+\varepsilon<0$($ t_0 <t<t_2$), 根据引理(2.2), 可知 $D^{q} z(t_2)=x_i(t_2) +\mathcal{K}_i +\varepsilon>0$, 又有 $\alpha( x_i(t_2))>0$, 则根据性质(3.2)有

$ D^{q} z_{i}(t_2)=D^{q} x_{i}(t_2)<\alpha (x_i(t_2))(\mu _{i}^{''}(\mathcal{K}_i+\varepsilon) +\delta _{ii}^{''}\mathcal{G}_i( -\mathcal{K}_i-\varepsilon) +B_{i}^{+})<0, $

此处产生的矛盾意味着对于任意 $t>t_0$, 如果 $x_{i}(t)$ 的初值在区间内, $x_i(t)$ 不会离开区间 $( -\infty,-\mathcal{K}_i -\varepsilon)$.

情形 3: 对 $x_i\in (\mathcal{K}_i+\varepsilon,+\infty)$存在 $t_3 \ge t_0$ 使得 $x_i(t_3)=\mathcal{K}_i +\varepsilon$, 即 $ x_i(t_3)-\mathcal{K}_i-\varepsilon=0$ 且有 $x_i(t)-\mathcal{K}_i -\varepsilon>0$($t_0<t<t_3$), 根据引理(2.2), 可知 $D^{q}z(t_3)=x_i(t_3) -\mathcal{K}_i +\varepsilon<0$, 又有 $ \alpha ( x_i( t_2 ) )>0 $, 而由性质(3.2)则有

$ D^{q} z_{i}(t_3)=D^{q} x_{i}(t_3)>\alpha(x_i(t_3))(\mu_{i}^{''}(\mathcal{K}_i+\varepsilon )-\delta_{ii}^{''}\mathcal{G}_i(\mathcal{K}_i+\varepsilon)-B_{i}^{-})>0, $

这意味着对于所有的 $ t>t_0 $, 如果 $ x_{i}(t) $ 的初值在区间内, 则状态 $ x_{i}(t) $ 仍会保持在区间 $(\mathcal{K}_i +\varepsilon,+\infty )$.

综上所述, 对于所有的 $t>t_0$, 如果 $\phi \in C((-\tau,0 ],\Re)$, 那么 $x_{i}(t)$ 会保持在集合 $\Omega_{\sigma} \in \bar{ \Theta}$ 中,故 $\Omega_{\sigma}$ 是正不变集.

3.3 SFCGNNs 平衡点的渐近稳定性

定理 3.3 位于集合 $\Omega_{\sigma} \in \bar{ \Theta}$ 中的平衡点是渐近稳定的,如果条件(3.1)成立,并且存在 $c_1 >0$, $c_2 >0$ 使得

$\begin{equation} \begin{aligned} \min _{1 \leq i \leq n} \left\{\mu_{i}^{*} - \bar{\delta}_{i i}-\frac{1}{2}\left(c_{1}+\frac{k_{i}^{2}}{c_{1}}\right) \sum_{j=1}^{n} \hat{\delta}_{i j}^{s}-\frac{c_{2}}{2} \sum_{j=1}^{n} \hat{\rho}_{i j}^{s}\right\} -\frac{1}{2} \max _{1 \leq i \leq n}\left\{\frac{k_{i}^{2}}{c_{2}} \sum_{j=1}^{n} \hat{\rho}_{i j}^{s}\right\}\ge 0, \end{aligned} \end{equation}$

此处 $\mu _{i}^{*}=\min \{0,\mu _{i}^{'},\mu _{i}^{''} \} $, $\bar{\delta}_{ii}=\max \{0,\delta _{ii}^{'}k_i, \delta _{ii}^{''}\check{k}_i \} $, $ \hat{\rho}_{ij}^{s}=\frac{\hat{\rho}_{ij}^{}+\hat{\rho}_{ji}^{}}{2} $, $ \hat{\rho}_{ij}^{}=\max \{ | \rho _{ij}^{'} |,| \rho _{ij}^{''} | \} $, $\hat{\delta}_{ij}^{s}=\frac{\hat{\delta}_{ij}^{}+\hat{\delta}_{ji}^{}}{2},( i,j=1,...,n ) $, $\hat{\delta}_{ij}^{}=\max \{ | \delta _{ij}^{'} |,| \delta _{ij}^{''} |\}$, if $ i\ne j $; $ \hat{\delta}_{ij}^{}=0 $ if, $ i=j $.

定义 $ x^{*}=(x_{1}^{*},\cdots, x_{n}^{*})^{T} $ 为系统(2.1)在初值条件 $\phi\in C([-\tau, 0], \Omega_{\sigma})$ 下的解, 也是位于给定集合 $\Omega_{\sigma}$ 的平衡点.

$ e(t)=\left(e_{1}(t), \cdots, e_{n}(t)\right)^{T} = x(t) - x^* $, 对于所有 $ t>t_0 $ 均有

$\begin{equation} \begin{aligned} D^{q} e_{i}(t) \in {\rm co}[\alpha_{i}(x_{i}(t))] & \left(-{\rm co}[\mu_{i}(x_{i}(t))] x_{i}(t)+\sum_{j=1}^{n} {\rm co}[\delta_{ij}(x_{j}(t))] \mathcal{G}_{j}(x_{j}(t))\right. \\ & \left.+\sum_{j=1}^{n} {\rm co}[\rho_{ij}(x_{j}(t_\tau))]\mathcal{G}_{j}(x_{j}(t_\tau))\right) \\ & -{\rm co}[\alpha_{i}(x_{i}^{*})]\left(-{\rm co}[\mu_{i}(x_{i}^{*})]x_{i}^{*}+\sum_{j=1}^{n} {\rm co}[\delta_{ij}(x_{j}^{*})]\mathcal{G}_{j}(x_{j}^{*})\right. \\ & \left.+\sum_{j=1}^{n} {\rm co}[\rho_{ij}(x_{j}^{*})] \mathcal{G}_{j}(x_{j}^{*})\right) \end{aligned} \end{equation}$

在集合 $\Omega _{\sigma}$ 中, co$[\alpha_{i}(x_{i}(t))]={\rm co}[\alpha_{i}(x_{i}^{*})]=\alpha'$$\alpha''$, co$[\mu_{i}(x_{i}(t))]={\rm co}[\mu_{i}(x_{i}^{*})]=\mu_{i}'$$\mu_{i}''$, co$[\delta_{i}(x_{i}(t))]={\rm co}[\delta_{ij}(x_{j}^{*})]=\delta_{i}'$$\delta_{i}''$, co$[\rho_{i}(x_{i}(t))]={\rm co}[\rho_{ij}(x_{j}^{*})]=\rho_{i}'$$\rho_{i}''$. 所以(3.6)可以改写为

$\begin{equation} \begin{aligned} D^qe_i(t) \in {\rm co}[\alpha_i(x_{i}^{*})](x_{i}^{*})&\left(-{\rm co}[\mu_i(x_{i}^{*} )]e_i+\sum_{j=1}^{n} {\rm co}[\delta _{ij}(x_{j}^{*})] ( \mathcal{G}_j(x_j(t))-\mathcal{G}_j(x_{j}^{*})) \right.\\ & \left. +\sum_{j=1}^{n} {\rm co}[\rho_{ij}(x_{j}^{*})] (\mathcal{G}_j( x_j(t_\tau)) -\mathcal{G}_j(x_{j}^{*}))\right),\\ \end{aligned} \end{equation}$

换而言之, 存在 $\alpha_{i}^*\in {\rm co}[\alpha_{i}(x_{i}^{*})]$, $\mu_{i}^* \in {\rm co}[\mu_{i}(x_{i}^{*})]$, $\delta_{ij}^* \in {\rm co}[\delta_{ij}(x_{j}^{*})]$$\rho_{ij}^* \in {\rm co}[\rho_{ij}(x_{j}^{*})]$ 使得

$\begin{equation} \begin{aligned} D^qe_i(t)=\alpha_{i}^{*}(x_{i}^{*})\left(-\mu_{i}^{*}e_i+\sum_{j=1}^n{\delta}_{ij}^{*}(\mathcal{G}_j(x_j(t))-\mathcal{G}_j( x_{j}^{*})) +\sum_{j=1}^n{\rho}_{ij}^{*}(\mathcal{G}_j(x_j(t_\tau))-\mathcal{G}_j(x_{j}^{*}))\right).\\ \end{aligned} \end{equation}$

构造函数 $ V(e(t))=e^{T}(t) e(t) / 2 $, 根据引理(2.1)与(3.8)可知

$\begin{equation} \begin{aligned} D^q V(e(t)) &\leq e^T(t) D^{q}e(t)\\ &=\sum_{i=1}^n{e}_i(t) \alpha _{i}^{*}(x_{i}^{*})\left[-\mu _{i}^{*}e_i(t) +\sum_{j=1,j\ne i}^n{\delta}_{ij}^{*}(\mathcal{G}_j(x_j(t))-\mathcal{G}_j(x_{j}^{*}))\right.\\ &\quad +\delta _{ii}^{*}( \mathcal{G}_i(x_i(t))-\mathcal{G}_i(x_{i}^{*}))\left. +\sum_{j=1}^n{\rho}_{ij}^{*}( \mathcal{G}_j(x_j(t_\tau))-\mathcal{G}_j(x_{j}^{*}))\right],\\ \end{aligned} \end{equation}$

根据激活函数性质有 $e_i(t)(\mathcal{G}_i( x_i(t) )-\mathcal{G}_i(x_{i}^{*})) =\mathcal{G}_{i}^{'}(\xi_i(t))e_{i}^{2}(t)\geq 0$, 其中 $\xi _i(t) \in(\min\{x_{i}^{*},\\ x_i(t)\},\max\{x_{i}^{*},x_i(t)\})$. 而根据 $ \bar{\delta}_{ii} $ 的定义, 有

$\begin{equation} \delta _{ii}^{*}e_i(t)(\mathcal{G}_i(x_i(t)) -\mathcal{G}_i(x_{i}^{*})) =\delta _{ii}^{*}\mathcal{G}_{i}^{'}(\xi_i(t)) e_{i}^{2}(t) \leq \bar{\delta}_{ii}e_{i}^{2}(t). \end{equation}$

此外

$\begin{equation} \begin{aligned} &~~~~\sum_{i=1}^{n}\sum_{j=1, j \neq i}^{n} e_{i}(t) \delta_{i j}^{*}(\mathcal{G}_{j}(x_{j}(t))-\mathcal{G}_{j}(x_{j}^{*}))\\ &\leq \sum_{i, j=1}^{n} \hat{\delta}_{i j}|e_{i}(t)(\mathcal{G}_{j}(x_{j}(t))-\mathcal{G}_{j}(x_{j}^{*}))| \\ &\leq \frac{1}{2} \sum_{i, j=1}^{n}(\hat{\delta}_{i j}+\hat{\delta}_{j i}) \frac{1}{2}(c_{1} e_{i}^{2}(t)+\frac{l_{j}^{2}}{c_{1}} e_{j}^{2}(t))=\frac{1}{2} \sum_{i=1}^{n}((c_{1}+\frac{k_{i}^{2}}{c_{1}}) \sum_{j=1}^{n} \hat{\delta}_{i j}^{s}) e_{i}^{2}(t), \end{aligned} \end{equation}$

而且

$\begin{equation} \begin{aligned} &~~~~\sum_{i, j=1}^{n} e_{i}(t) \rho_{ij}^{*}(\mathcal{G}_{j}(x_{j}(t_\tau))-\mathcal{G}_{j}(x_{j}^{*}))\\ &\leq \frac{1}{2} \sum_{i, j=1}^{n}(\hat{\rho}_{i j}+\hat{\rho}_{j i}) \frac{1}{2}(c_{2} e_{i}^{2}(t)+\frac{l_{k}^{2}}{c_{2}} e_{j}^{2}(t_\tau)) \\ &=\frac{c_{2}}{2} \sum_{i=1}^{n}(\sum_{j=1}^{n} \hat{\rho}_{i j}^{s}) e_{i}^{2}(t)+\frac{1}{2} \sum_{i=1}^{n}(\frac{k_{i}^{2}}{c_{2}} \sum_{j=1}^{n} \hat{\rho}_{i j}^{s}) e_{i}^{2}(t_\tau). \end{aligned} \end{equation}$

根据(3.9)-(3.12), 有

$\begin{equation} D^{q} V(t) \leq-\varsigma V(t)+\varpi V(t_\tau), \end{equation}$

其中

$\begin{equation*} \varsigma =\min_{1\leq i\leq n}\left\{\alpha_{i}^*(2\mu _{i}^{*}-2\bar{\delta}_{ii}-(c_1+\frac{k_{i}^{2}}{c_1}) \sum_{j=1}^n{\hat{\delta}_{ij}^{*}}-c_2 \sum_{j=1}^n{\hat{\rho}_{ij}^{*}}) \right\}, \, \varpi =\max_{1\leq i\leq n}\left\{\alpha_{i}^* \frac{k_{i}^{2}}{c_2}\sum_{j=1}^n{\hat{\rho}_{ij}^{s}} \right\}, \end{equation*}$

根据 (3.5), 引理 (2.3) 可知 $\varsigma>\varpi$, $\underset{t\rightarrow \infty}{\lim}e_i(t)=0$, 所以称集合 $\Omega_\sigma$ 中的平衡点是渐近稳定的.

定理3.4 在(3.1)与(3.8)条件下, 若存在

$\begin{equation} \bar{\delta}_{ii}+\frac{1}{2} (c_1+ \frac{k_{i}^{2}}{c_1}) \sum_{j=1}^n{\hat{\delta}_{ij}^{*}}+ \frac{c_2}{2} \sum_{j=1}^n{\hat{\rho}_{ij}^{*}} -\frac{1}{2} \frac{k_{i}^{2}}{c_2}\sum_{j=1}^n{\hat{\rho}_{ij}^{s}} > \mu_{i}^{*}, \end{equation}$

则集合 $ \Omega_\sigma^{(1)} \in \Theta-\bar{\Theta} $ 中的平衡点唯一且不稳定.

在小节 3.1 中已说明集合 $ \Omega_\sigma^{(1)} \in \Theta-\bar{\Theta} $ 中平衡点的存在性, 此处采用反证法, 假设存在两个位于集合 $ \Omega_\sigma^{(1)} $ 中的平衡点 $ y^{*}=\left(y_{1}^{*}, \cdots, y_{n}^{*}\right)^{T} $$ x^{*}=\left(x_{1}^{*}, \cdots, x_{n}^{*}\right)^{T} $, 根据 $ \Omega_\sigma^{(1)} $ 的定义可知存在 $ j\in L_{j2}\cup L_{j4} $ 使得 $x_{j}^{*}=y_{j}^{*} \in \{-\mathcal{K}_j, \mathcal{K}_j\}$, 这表明 $ \mu_{i }\left(y_{j}^{*}\right)=\mu_{i}\left(x_{j}^{*}\right) $, $ {\rm co}\left[\delta_{i j}\left(y_{j}^{*}\right)\right]= {\rm co}\left[\delta_{i j}\left(x_{j}^{*}\right)\right] $$ {\rm co}\left[\rho_{i j}\left(y_{j}^{*}\right)\right]= {\rm co}\left[\rho_{i j}\left(x_{j}^{*}\right)\right] $. 因此

$\begin{equation*} \begin{aligned} 0\in &-\mu_{i}(x_{i}^{*})(y_{i}^{*}-x_{i}^{*})+\sum_{j=1}^{n} {\rm co}[\delta_{ij}(x_{j}^{*})](\mathcal{G}_{j}(y_{i}^{*})-\mathcal{G}_{j}(x_{j}^{*}))+\sum_{j=1}^{n}{\rm co}[\rho_{ij}(x_{j}^{*})](\mathcal{G}_{j}(y_{i}^{*})-\mathcal{G}_{j}(x_{j}^{*})), \end{aligned} \end{equation*}$

与前文所述类似, 存在 $\mu_{i}^*\in {\rm co}[\mu_{i}(x_{i}^{*})]$,$\delta_{ij}^*\in {\rm co}[\delta_{ij}(x_{j}^{*})]$$ \rho_{ij}^*\in {\rm co}[\rho_{ij}(x_{j}^{*})]$ 使得

$\begin{equation*} \mu_{i}^*(y_{i}^{*}-x_{i}^{*})=\sum_{j=1}^{n} \delta_{ij}^*(\mathcal{G}_{j}(y_{i}^{*})-\mathcal{G}_{j}(x_{j}^{*}))+\sum_{j=1}^{n}\rho_{ij}^*(\mathcal{G}_{j}(y_{i}^{*})-\mathcal{G}_{j}(x_{j}^{*})), \end{equation*}$

因此

$\begin{equation} \begin{aligned} ( \mu_{i}^*) \|y^{*}_{i}-x^{*}_{i}\|_{2}^{2} & =(y^{*}_{i}-x^{*}_{i})^{T}\sum_{j=1}^{n}\delta_{ij}^{*}(\mathcal{G}_{j}(y_{j}^{*})-\mathcal{G}_{j}(x_{j}^{*})) +\sum_{j=1}^{n} \rho_{i j}^{*}(\mathcal{G}_{j}(y_{j}^{*})-\mathcal{G}_{j}(x_{j}^{*}))\\ & \leq (\bar{\delta}_{ii}+\frac{1}{2} (c_1+ \frac{k_{i}^{2}}{c_1})\sum_{j=1}^n{\hat{\delta}_{ij}^{*}}+ \frac{c_2}{2} \sum_{j=1}^n{\hat{\rho}_{ij}^{*}} -\frac{1}{2} \frac{k_{i}^{2}}{c_2}\sum_{j=1}^n{\hat{\rho}_{ij}^{s}} )\|y^{*}_{i}-x^{*}_{i}\|_{2}^{2}. \end{aligned} \end{equation}$

根据条件(3.18),(3.14)与(3.15), 可知 $ \|y^{*}_{i}-x^{*}_{i}\|_{2}^{2} = 0$, 即位于给定集合 $\Omega_\sigma^{(1)}$ 中的平衡点具有唯一性.

定义 $ \varepsilon _0=\min_{y^*,x^*\in \mathcal{K},y^*\ne x^*}= \frac{1}{2} \|y^{*}_{i}-x^{*}_{i}\|_{2} $, 此处 $ \mathcal{K} $ 表示平衡点所在集合, 若要证明集合 $\Omega_\sigma^{(1)} \in \Theta-\bar{\Theta} $ 中的平衡点 $\bar x^* $ 是不稳定的, 只需要证明在初始条件对于所有的 $ s\in (t_0-\tau,t) $, $x(s)=\phi(s-t_0) \equiv \bar x$, 当 $ t_1>t_0 $, 对任意足够小的 $\delta_0 >0$ 都满足 $\|\bar{x}^{*}-\bar{x}\|_{2}<\delta_0 $ 时, 有 $\|x(t_{1})-\bar{x}^{*}\|_{2}>\varepsilon_{0}$. 对于 $\bar{x}_i^* \in\Omega_\sigma^{(1)}$, 取 $\bar{x}_{i}^{*} \in\{-\bar{\mathcal{K}}_{i}, \bar{\mathcal{K}}_{i}\}$, 此时定义 $\bar{x}_{i}=\bar{x}_{i}^{*}+\delta_0 /\sqrt{2n}$, 对于 $ \bar{x}_i^*\in \Omega_\sigma^{(1)}$, 定义 $\bar{x}_{j}=\bar{x}_{j}^{*}$, 这样可以保证 $ \|\bar{x}^{*}-\bar{x}\|_{2} \leq \delta_0 /2<\delta_0$. 理论 (3.3) 证明了, 当 $x(s)$ 的初始状态 $\bar{x}$ 位于一个正不变集 $\Omega_{\sigma}^{(1)}$ 内时, 系统的解会收敛到该正不变集内的平衡点处, 记作 $\check{x}^*$, 因此存在时间 $t_1>t_0$, 使得 $\|x(t_{1})-\check{x}^{*}\|_{2}<\varepsilon_{0}$, 对所有 $t\ge t_1$ 有以下不等式成立

$ \|x(t_{1})-\bar{x}^{*}\|_{2} \geq\|\check{x}^{*}-\bar{x}^{*}\|_{2}-\|x(t_{1})-\check{x}^{*}\|_{2}>2 \varepsilon_{0}-\varepsilon_{0}=\varepsilon_{0}, $

因此位于集合 $ \Omega_\sigma^{(1)} \in \Theta-\bar{\Theta} $ 中的平衡点 $\bar x^*$ 是不稳定的.

4 仿真结果

为了验证前文所推导出的理论结果的有效性, 本节将给出两个数值算例.

例 1 虑一个二维的 SFCGNNs 满足以下分数阶微分方程,

$\begin{equation} \left\{\begin{array}{r} \mathcal{D}^{0.98} x_{1}(t)=\alpha_{1}(x_{1}(t))(-\mu_{1}(x_{1}(t))x_{1}(t)+\delta_{11}(x_{1}(t))f(x_{1}(t))+\delta_{12}(x_{2}(t))f(x_{2}(t))\\ +\rho_{11}(x_{1}(t)) f(x_{1}(t+(0.5\sin(t)-1)))+\rho_{12}(x_{2}(t))f(x_{2}(t+(0.5\sin(t)-1)))+0.2), \\ \mathcal{D}^{0.98} x_{2}(t)=\alpha_{2}(x_{2}(t))(-\mu_{2}(x_{2}(t))x_{2}(t)+\delta_{21}(x_{1}(t)) f(x_{1}(t))+\delta_{22}(x_{2}(t)) f(x_{2}(t)) \\ +\rho_{21}(x_{1}(t)) f(x_{1}(t+(0.5\sin(t)-1)))+\rho_{22}(x_{2}(t)) f(x_{2}(t+(0.5\sin(t)-1)))+0.5), \end{array}\right. \end{equation}$

对于 $t\ge 0$, 其中参数如下所示:

$\begin{equation*} \begin{array}{l} \alpha_{1}(x_{1}(t))=\left\{\begin{array}{ll} 1.2, & |x_{1}(t)| \leq 1.5, \\ 0.8, & |x_{1}(t)|>1.5, \end{array} \right.\qquad\qquad\quad \alpha_{2}(x_{2}(t))=\left\{\begin{array}{ll} 1, & |x_{2}(t)| \leq 1.5, \\ 1, & |x_{2}(t)|>1.5, \end{array}\right. \\ \mu_{1}(x_{1}(t))=\left\{\begin{array}{ll} 1.5, & |x_{1}(t)| \leq 1.5, \\ 1.5, & |x_{1(t)}|>1.5, \end{array}\right. \qquad\qquad\quad \mu_{2}(x_{2}(t))=\left\{\begin{array}{ll} 0.7, & |x_{2}(t)| \leq 1.5, \\ 0.9, & |x_{2}(t)|>1.5, \end{array}\right. \\ \delta_{11}(x_{1}(t))=\left\{\begin{array}{cc} -4, & |x_{1}(t)| \leq 1.5, \\ 4, & |x_{1}(t)|>1.5, \end{array} \right.\qquad\qquad\quad \delta_{21}(x_{1}(t))=\left\{\begin{array}{cc} -0.2, & |x_{1}(t)| \leq 1.5, \\ 0.2, & |x_{1}(t)|>1.5, \end{array}\right. \\ \delta_{12}(x_{2}(t))=\left\{\begin{array}{cc} -0.3, & |x_{2}(t)| \leq 1.5, \\ 0.3, & |x_{2}(t)|>1.5, \end{array} \right.\qquad\qquad\; \delta_{22}(x_{2}(t))=\left\{\begin{array}{cc} -3, & |x_{2}(t)| \leq 1.5, \\ 3, & |x_{2}(t)|>1.5, \end{array}\right. \\ \rho_{11}(x_{1}(t))=\left\{\begin{array}{cc} -0.2, & |x_{1}(t+(0.5\sin(t)-1))| \leq 1.5, \\ 0.2, & |x_{1}(t+(0.5\sin(t)-1))|>1.5, \end{array} \right. \\ \end{array} \end{equation*}$
$\begin{equation*} \begin{array}{l} \rho_{21}\left(x_{1}(t)\right)=\left\{\begin{array}{cc} -0.1, & |x_{1}(t+(0.5\sin(t)-1))| \leq 1.5, \\ 0.1, & |x_{1}(t+(0.5\sin(t)-1))|>1.5, \end{array}\right. \\ \rho_{12}(x_{2}(t))=\left\{\begin{array}{cl} -0.2, & |x_{2}(t+(0.5\sin(t)-1))| \leq 1.5, \\ 0.2, & |x_{2}(t+(0.5\sin(t)-1))|>1.5, \end{array} \right. \\ \rho_{22}\left(x_{2}(t)\right)=\left\{\begin{array}{cl} -0.3, & |x_{2}(t+(0.5\sin(t)-1))| \leq 1.5, \\ 0.3, & |x_{2}(t+(0.5\sin(t)-1))|>1.5. \end{array}\right. \end{array} \end{equation*}$

此时激活函数为 $ f(r) = {\rm tanh}(r) $, 几何形状如图1 所示, 且有

$\begin{equation*} \begin{cases} \alpha_{1}'' \left( \mu_{1}''\mathcal{K}_1 + \delta_{11}''\mathcal{G}_1(-\mathcal{K}_1) + B_{1}^{+} \right) = -0.3763 < 0, \\ \alpha_{1}' \left( \mu_{1}'\mathcal{K}_1 + \delta_{11}'\mathcal{G}_1(-\mathcal{K}_1) + B_{1}^{-} \right) = 6.4445 > 0, \\ \alpha_{1}' \left( -\mu_{1}'\mathcal{K}_1 + \delta_{11}'\mathcal{G}_1(\mathcal{K}_1) + B_{1}^{+} \right) = -5.9645 < 0, \\ \alpha_{1}'' \left( -\mu_{1}''\mathcal{K}_1 + \delta_{11}''\mathcal{G}_1(\mathcal{K}_1) + B_{1}^{-} \right) = 0.6963 > 0, \\ \alpha_{2}'' \left( \mu_{2}''\mathcal{K}_2 + \delta_{22}''\mathcal{G}_2(-\mathcal{K}_2) + B_{2}^{+} \right) = -0.65 < 0, \\ \alpha_{2}' \left( \mu_{2}'\mathcal{K}_2 + \delta_{22}'\mathcal{G}_2(-\mathcal{K}_2) + B_{2}^{-} \right) = 3.85 > 0, \\ \alpha_{2}' \left( -\mu_{2}'\mathcal{K}_2 + \delta_{22}'\mathcal{G}_2(\mathcal{K}_2) + B_{2}^{+} \right) = -3.05 < 0, \\ \alpha_{2}'' \left( -\mu_{2}''\mathcal{K}_2 + \delta_{22}''\mathcal{G}_2(\mathcal{K}_2) + B_{2}^{-} \right) = 1.45 > 0, \end{cases} \end{equation*}$

图1

图1   例1中的激活函数


此外,有 $\mu _{1}^{*}=\mu _{2}^{*}=0$, $\bar{\delta}_{11}=0.68$, $\bar{\delta}_{22}=0$, $\hat{\rho}_{11}^{s}=0.2$, $\hat{\rho}_{12}^{s}=\hat{\rho}_{21}^{s}=0.15$, $\hat{\rho}_{22}^{s}=0.3$, $\hat{\delta}_{11}^{s}=\hat{\delta}_{22}^{s}=0$, $\hat{\delta}_{12}^{s}=\hat{\delta}_{21}^{s}=0.25$, 验证定理 3.1-定理 3.4 的所需条件可以被满足, 因此根据定理 3.1-定理 3.4 可知 SFCGNNs 具有 $25$ 个平衡点, 其中 $9$ 个平衡点是渐近稳定的, 而其余 $16$ 个平衡点不稳定.

图2

图2   例2中的激活函数


为了验证时变时延对神经网络产生的影响, 考虑 SFCGNNs(4.1)具有条件 $\rho_{11}(x_{1}(t))=\rho_{12}(x_{2}(t))=\rho_{21}(x_{1}(t))=\rho_{22}(x_{2}(t))=0$. 在此条件下,随机选取了 150 个初始值 $x(0) \in [-6, 6] \times [-6, 6]$ 进行数值仿真, 仿真结果如图3 所示, 其中使用红色符号 $*$ 标记稳定的平衡点所在位置. 作为对比, 考虑时变时延为 $t_\tau=(0.5\sin(t)-1)$, $t\ge 0$, 此时时变时延对 SFCGNNs(4.1)产生的影响如 SFCGNNs(4.1)中的参数所示, 选取同样的初始条件, 最后数值仿真结果表示在图4 并使用蓝色符号 $*$ 标记稳定的平衡点所在位置. 仿真结果表明, 时变时延的存在不会影响稳定的平衡点的数量, 这说明本文的推导条件和结果在时延存在的情况下仍然成立.

图3

图3   例1中 $(x_{1}(t),x_{2}(t))$ 的相图, 无时延.


图4

图4   例1中 $(x_{1}(t),x_{2}(t))$ 的相图, 时变时延为 $(0.5\sin(t)-1)$


需要注意的是, 在例 1 中有 $\alpha_2'=\alpha_2''=1$, 此时 Cohen-Grossberg 神经网络就变成了 Hopfield 神经网络, 所以本文的结果同样适用于切换分数阶 Hopfield 神经网络. 而对于切换分数阶递归神经网络, 接下来将给出例 2, 说明在 $\alpha_i'=\alpha_i''=1$$\mu_i'=\mu_i''=1$ 时, SFCGNNs (1) 的结果同样适用于分数阶递归神经网络.

例 2. 考虑一个二维的 SFCGNNs, 其激活函数如图2 所示,

$\begin{equation} \left\{\begin{array}{r} \mathcal{D}^{0.98} x_{1}(t)=\alpha_{1}(x_{1}(t)) (-\mu_{1}(x_{1}(t)) x_{1}(t)+\delta_{11}(x_{1}(t)) f(x_{1}(t))+\delta_{12}(x_{2}(t)) f(x_{2}(t))\\ +\rho_{11}(x_{1}(t-1)) f(x_{1}(t-1))+\rho_{12}(x_{1}(t-1)) f(x_{2}(t-1))+0.1), \\ \mathcal{D}^{0.98} x_{2}(t)=\alpha_{2}(x_{2}(t)) (-\mu_{2}(x_{2}(t))x_{2}(t)+\delta_{21}(x_{1}(t)) f(x_{1}(t))+\delta_{22}(x_{2}(t))f(x_{2}(t)) \\ +\rho_{21}(x_{1}(t-1)) f(x_{1}(t-1))+\rho_{22}(x_{2}(t-1))f(x_{2}(t-1))+0.2), \end{array}\right. \end{equation}$

对于 $t\ge 0$, 其中

$\begin{equation*} \begin{array}{l} \alpha_{1}(x_{1}(t))=\left\{\begin{array}{ll} 1, & |x_{1}(t)| \leq 1, \\ 1, & |x_{1}(t)|>1, \end{array} \right.\qquad\qquad\qquad \alpha_{2}(x_{2}(t))=\left\{\begin{array}{ll} 1, & |x_{2}(t)| \leq 1, \\ 1, & |x_{2}(t)|>1, \end{array}\right. \\ \mu_{1}(x_{1}(t))=\left\{\begin{array}{ll} 1, & |x_{1}(t)| \leq 1, \\ 1, & |x_{1(t)}|>1, \end{array}\right. \qquad\qquad\qquad \mu_{2}(x_{2}(t))=\left\{\begin{array}{ll} 1, & |x_{2}(t)| \leq 1, \\ 1, & |x_{2}(t)|>1, \end{array}\right. \\ \end{array} \end{equation*}$
$\begin{equation*} \begin{array}{l} \delta_{11}(x_{1}(t))=\left\{\begin{array}{cc} -3, & |x_{1}(t)| \leq 1, \\ 3, & |x_{1}(t)|>1, \end{array} \right.\qquad\qquad\quad \delta_{21}(x_{1}(t))=\left\{\begin{array}{cc} -0.3, & |x_{1}(t)|\leq 1, \\ 0.3, & |x_{1}(t)|>1, \end{array}\right. \\ \delta_{12}(x_{2}(t))=\left\{\begin{array}{cc} -0.4, & |x_{2}(t)| \leq 1, \\ 0.4, & |x_{2}(t)|>1, \end{array}\right. \qquad\qquad\; \delta_{22}(x_{2}(t))=\left\{\begin{array}{cc} -4, & |x_{2}(t)| \leq 1, \\ 4, & |x_{2}(t)|>1, \end{array}\right. \\ \rho_{11}(x_{1}(t-1))=\left\{\begin{array}{cc} -0.3, & |x_{1}(t-1)| \leq 1, \\ 0.3, & |x_{1}(t-1)|>1, \end{array}\right. \quad \rho_{21}(x_{1}(t-1))=\left\{\begin{array}{cc} -0.1, & |x_{1}(t-1)| \leq 1, \\ 0.1, & |x_{1}(t-1)|>1, \end{array}\right. \\ \rho_{12}(x_{2}(t-1))=\left\{\begin{array}{cl} -0.2, &|x_{2}(t-1)| \leq 1, \\ 0.2, & |x_{2}(t-1)|>1, \end{array} \right.\quad \rho_{22}(x_{2}(t-1))=\left\{\begin{array}{cl} -0.3, & |x_{2}(t-1)| \leq 1, \\ 0.3, & |x_{2}(t-1)|>1. \end{array}\right. \end{array} \end{equation*}$

根据上述参数, 可以计算出 $\mu_{i}^{*},\bar{\delta}_{ij},\hat{\rho}_{ij}^{s},\hat{\delta}_{ij}^{s}$, 从而确定在 $c_1=c_2=1$ 时, 定理 3.3 的条件对 SFCGNNs 均成立. 我们可以得出结论: SFCGNNs 具有 $ 5\times5=25$ 个平衡点, 其中 $3\times3=9$ 个平衡点是渐近稳定的, 而其余 $16$ 个平衡点是不稳定的. 随机选取 150 个初始条件 $x(0) \in [-5, 5] \times [-5, 5]$ 的仿真图像表示在图5, 其中蓝色符号 $*$ 为稳定的平衡点所在位置. 因此通过上述算例, 可知本论文所推导的的理论成果不仅可用于 Hopfield 神经网络的多稳定分析, 还可用于递归神经网络[32]的多稳定性分析, 更具有一般性.

图5

图5   例2中 $(x_{1}(t),x_{2}(t))$ 的相图


5 总结

在本文中, 研究了具有时变时延的 SFCGNNs 的多重平衡点的稳定性与不稳定性, 得到了确定 SFCGNNs 存在 $ 5^n $ 个平衡点的充分条件, 其中 $ 3^n $ 个平衡点是渐近稳定的, 确定了其余 $ 5^n - 3^n $ 个平衡点不稳定的充分条件. 并且本文所给出的 SFCGNNs 的参数若取特殊值, 如在 $ \alpha_1=\alpha_2=1 $ 时, SFCGNNs (1) 变为切换分数阶 Hopfield 神经网络, 在此条件的基础上再有 $ \mu _1=\mu_2=1 $, SFCGNNs (1) 则是分数阶递归神经网络. 本文推导的理论结果仍然成立. 在未来, 可能会探索具有更强非线性激活函数的 SFCGNNs 的多稳定性.

参考文献

Liu Y, Wang Z, Xiao M, et al.

$\mu$-stability of multiple equilibria in Cohen-Grossberg neural networks and its application to associative memory

Science Technological Sciences, 2023, 66(9): 2611-2624

[本文引用: 1]

Shen Y, Zhu S, Liu X, et al.

Multistability and associative memory of neural networks with Morita-like activation functions

Neural Networks, 2021, 142: 162-170

DOI:10.1016/j.neunet.2021.04.035      PMID:34000563      [本文引用: 1]

This paper presents the multistability analysis and associative memory of neural networks (NNs) with Morita-like activation functions. In order to seek larger memory capacity, this paper proposes Morita-like activation functions. In a weakened condition, this paper shows that the NNs with n-neurons have (2m+1) equilibrium points (Eps) and (m+1) of them are locally exponentially stable, where the parameter m depends on the Morita-like activation functions, called Morita parameter. Also the attraction basins are estimated based on the state space partition. Moreover, this paper applies these NNs into associative memories (AMs). Compared with the previous related works, the number of Eps and AM's memory capacity are extensively increased. The simulation results are illustrated and some reliable associative memories examples are shown at the end of this paper.Copyright © 2021 Elsevier Ltd. All rights reserved.

Deng K, Zhu S, Bao G, et al.

Multistability of dynamic memristor delayed cellular neural networks with application to associative memories

IEEE Transactions on Neural Networks and Learning Systems, 2021, 34(2): 690-702

DOI:10.1109/TNNLS.2021.3099814      URL     [本文引用: 1]

Ding D, Xiao H, Yang Z, et al.

Coexisting multi-stability of Hopfield neural network based on coupled fractional-order locally active memristor and its application in image encryption

Nonlinear Dynamics, 2022, 108(4): 4433-4458

DOI:10.1007/s11071-022-07371-0      [本文引用: 1]

Du W, Xiang J, Tan M.

Multistability of quaternion-valued recurrent neural networks with discontinuous nonmonotonic piecewise nonlinear activation functions

Neural Processing Letters, 2023, 55(5): 5855-5884

DOI:10.1007/s11063-022-11116-w      [本文引用: 1]

Wang J, Zhu S, Ma Q, et al.

Multistability of recurrent neural networks with general periodic activation functions and unbounded time-varying delays

Journal of the Franklin Institute, 2024, 361(18): 107236

DOI:10.1016/j.jfranklin.2024.107236      URL     [本文引用: 1]

Ou S, Guo Z, Wen S, et al.

Multistability and fixed-time multisynchronization of switched neural networks with state-dependent switching rules

Neural Networks, 2024, 180: 106713

DOI:10.1016/j.neunet.2024.106713      URL     [本文引用: 1]

Yang L, Gong W, Li Q, et al.

Multistability analysis of complex-valued recurrent neural networks with sine and cosine activation functions

Neurocomputing, 2024, 577: 127382

DOI:10.1016/j.neucom.2024.127382      URL     [本文引用: 1]

Ke J W, Zhang J E, Zhang J X.

Multimode function multistability of Cohen-Grossberg neural networks with Gaussian activation functions and mixed time delays

AIMS Mathematics, 2024, 9(2): 4562-4586

DOI:10.3934/math.2024220      URL     [本文引用: 1]

Zhu S, Zhang J, Liu X, et al.

Multistability and robustness of competitive neural networks with time-varying delays

IEEE Transactions on Neural Networks and Learning Systems, 2024, 35(12): 18746-18757

DOI:10.1109/TNNLS.2023.3321434      URL     [本文引用: 1]

Cohen M A, Grossberg S.

Absolute stability of global pattern formation and parallel memory storage by competitive neural networks

IEEE Transactions on Systems, Man, and Cybernetics, 1983 (5): 815-826

[本文引用: 1]

Faydasicok O, Arik S.

A novel Lyapunov stability analysis of neutral-type Cohen-Grossberg neural networks with multiple delays

Neural Networks, 2022, 155: 330-339

DOI:10.1016/j.neunet.2022.08.022      PMID:36099666      [本文引用: 1]

The major target of this research article is to conduct a new Lyapunov stability analysis of a special model of Cohen-Grossberg neural networks that include multiple delay terms in state variables of systems neurons and multiple delay terms in time derivatives of state variables of systems neurons in the network structure. Employing some proper linear combinations of three different positive definite and positive semi-definite Lyapunov functionals, we obtain some novel sufficient criteria that guarantee global asymptotic stability of this type of multiple delayed Cohen-Grossberg type neural systems. These newly derived stability results are determined to be completely independent of the involved time delay terms and neutral delay terms, and they are totally characterized by the values of the interconnection parameters of Cohen-Grossberg neural system. Besides, the validation of the obtained stability criteria can be justified by applying some simple appropriate algebraic equations that form some particular relations among the constant system elements of the considered neutral neural systems. A useful and instructive numerical example is analysed to exhibit some major advantages and novelties of these newly proposed global stability results in this paper over some previously reported corresponding asymptotic stability conditions.Copyright © 2022 Elsevier Ltd. All rights reserved.

Dong Z, Wang X, Zhang X.

A nonsingular M-matrix-based global exponential stability analysis of higher-order delayed discrete-time Cohen-Grossberg neural networks

Applied Mathematics and Computation, 2020, 385: 125401

DOI:10.1016/j.amc.2020.125401      URL     [本文引用: 1]

Zhou L, Zhao Z.

Asymptotic stability and polynomial stability of impulsive Cohen-Grossberg neural networks with multi-proportional delays

Neural Processing Letters, 2020, 51: 2607-2627

DOI:10.1007/s11063-020-10209-8      [本文引用: 1]

Ketencigil M M, Faydasicok O, Arik S.

Novel criteria for robust stability of Cohen-Grossberg neural networks with multiple time delays

Discrete and Continuous Dynamical Systems-S, 2022, 15(11): 3189-3203

DOI:10.3934/dcdss.2022081      URL     [本文引用: 1]

Bohner M, Stamov G T, Stamova I M.

Almost periodic solutions of Cohen-Grossberg neural networks with time-varying delay and variable impulsive perturbations

Communications in Nonlinear Science and Numerical Simulation, 2020, 80: 104952

DOI:10.1016/j.cnsns.2019.104952      URL     [本文引用: 1]

Wan L, Liu Z.

Multimode function multistability for Cohen-Grossberg neural networks with mixed time delays

ISA Transactions, 2022, 129: 179-192

DOI:10.1016/j.isatra.2021.11.046      URL     [本文引用: 1]

Yao C L H.

Exponential multistability of memristive Cohen-Grossberg neural networks with stochastic parameter perturbations

Applied Mathematics and Computation, 2020, 386: 125483

DOI:10.1016/j.amc.2020.125483      URL     [本文引用: 1]

Wan P, Sun D, Zhao M, et al.

Monostability and multistability for almost-periodic solutions of fractional-order neural networks with unsaturating piecewise linear activation functions

IEEE Transactions on Neural Networks and Learning Systems, 2020, 31(12): 5138-5152

DOI:10.1109/TNNLS.5962385      URL     [本文引用: 1]

Kao Y, Li H.

Asymptotic multistability and local S-asymptotic $\omega$-periodicity for the nonautonomous fractional-order neural networks with impulses

Science China Information Sciences, 2021, 64(1): 1-13

[本文引用: 1]

Wu Z.

Multiple asymptotic stability of fractional-order quaternion-valued neural networks with time-varying delays

Neurocomputing, 2021, 448: 301-312

DOI:10.1016/j.neucom.2021.03.079      URL     [本文引用: 1]

Sene N.

Mittag-Leffler input stability of fractional differential equations and its applications

Discrete & Continuous Dynamical Systems-S, 2020, 13(3): 867-880

[本文引用: 1]

Cheng Y, Hu T, Xu W, et al.

Fixed-time synchronization of fractional-order complex-valued neural networks with time-varying delay via sliding mode control

Neurocomputing, 2022, 505: 339-352

DOI:10.1016/j.neucom.2022.07.015      URL     [本文引用: 1]

Hui M, Wei C, Zhang J, et al.

Finite-time synchronization of memristor-based fractional order Cohen-Grossberg neural networks

IEEE Access, 2020, 8: 73698-73713

DOI:10.1109/Access.6287639      URL     [本文引用: 1]

Jiang W, Li Z, Zhang Y.

Global Mittag-Leffler stability and global asymptotic $\omega$-period for fractional-order Cohen-Grossberg neural networks with time-varying delays

International Journal of Pattern Recognition and Artificial Intelligence, 2022, 36(12): Art 2259023

[本文引用: 1]

Wan L, Liu Z.

Multiple $O(t^{-q})$ stability and instability of time-varying delayed fractional-order Cohen-Grossberg neural networks with Gaussian activation functions

Neurocomputing, 2021, 454: 212-227

DOI:10.1016/j.neucom.2021.05.018      URL     [本文引用: 1]

Cao B, Nie X, Zheng W X, et al.

Multistability of state-dependent switched fractional-order Hopfield neural networks with Mexican-hat activation function and its application in associative memories

IEEE Transactions on Neural Networks and Learning Systems, 2023, 36(1): 1213-1227

DOI:10.1109/TNNLS.2023.3334871      URL     [本文引用: 1]

Guo Z, Ou S, Wang J.

Multistability of switched neural networks with sigmoidal activation functions under state-dependent switching

Neural Networks, 2020, 122: 239-252

DOI:S0893-6080(19)30334-X      PMID:31726330      [本文引用: 1]

This paper presents theoretical results on the multistability of switched neural networks with commonly used sigmoidal activation functions under state-dependent switching. The multistability analysis with such an activation function is difficult because state-space partition is not as straightforward as that with piecewise-linear activations. Sufficient conditions are derived for ascertaining the existence and stability of multiple equilibria. It is shown that the number of stable equilibria of an n-neuron switched neural networks is up to 3 under given conditions. In contrast to existing multistability results with piecewise-linear activation functions, the results herein are also applicable to the equilibria at switching points. Four examples are discussed to substantiate the theoretical results.Copyright © 2019 Elsevier Ltd. All rights reserved.

Guo Z, Ou S, Wang J.

Multistability of switched neural networks with Gaussian activation functions under state-dependent switching

IEEE Transactions on Neural Networks and Learning Systems, 2021, 33(11): 6569-6583

DOI:10.1109/TNNLS.2021.3082560      URL     [本文引用: 1]

Liu Y, Wang Z, Huang X.

Multistability analysis of state-dependent switched Hopfield neural networks with the Gaussian-wavelet-type activation function

Mathematics and Computers in Simulation, 2022, 196: 232-250

DOI:10.1016/j.matcom.2022.01.021      URL     [本文引用: 1]

Wan L, Liu Z.

Multiple exponential stability and instability for state-dependent switched neural networks with time-varying delays and piecewise-linear radial basis activation functions

Neurocomputing, 2023, 524: 1-16

DOI:10.1016/j.neucom.2022.12.040      URL     [本文引用: 1]

Liu P, Xu M, Li Y, et al.

Multistability analysis of switched fractional-order recurrent neural networks with time-varying delay

Neural Computing and Applications, 2022, 34(23): 21089-21100

DOI:10.1007/s00521-022-07592-9      [本文引用: 2]

Dong H, Wan L, Wu A.

Multistability of fractional-order Cohen-Grossberg neural networks with state-dependent switching and time-varying delays via Gaussian-like activation functions

Asian Journal of Control, 2025

[本文引用: 2]

Duarte-Mermoud M A, Aguila-Camacho N, Gallegos J A, et al.

Using general quadratic Lyapunov functions to prove Lyapunov uniform stability for fractional order systems

Communications in Nonlinear Science and Numerical Simulation, 2015, 22(1-3): 650-659

DOI:10.1016/j.cnsns.2014.10.008      URL     [本文引用: 1]

Liu P, Zeng Z, Wang J.

Multiple Mittag-Leffler stability of fractional-order recurrent neural networks. IEEE Transactions on Systems

Man, and Cybernetics: Systems, 2017, 47(8): 2279-2288

[本文引用: 1]

He B B, Zhou H C, Chen Y Q, et al.

Asymptotical stability of fractional order systems with time delay via an integral inequality

IET Control Theory & Applications, 2018, 12(12): 1748-1754

DOI:10.1049/cth2.v12.12      URL     [本文引用: 1]

/