面向突触故障的回声状态网络容错方法研究  

Research on Fault Tolerance Method for Echo State Networks Oriented to Synapse Faults

作　　者：孙晓川 高佳慧 王宇 李莹琦 SUN Xiaochuan;GAO Jiahui;WANG Yu;LI Yingqi(College of Artificial Intelligence,North China University of Science and Technology,Tangshan 063210,China;Hebei Key Laboratory of Industrial Perception,Tangshan 063210,China)

机构地区：[1]华北理工大学人工智能学院,河北唐山063210 [2]河北省工业智能感知重点实验室,河北唐山063210

出　　处：《郑州大学学报（理学版）》2024年第5期39-46,共8页Journal of Zhengzhou University:Natural Science Edition

基　　金：河北省高等学校科学技术研究重点项目(ZD2021088)。

摘　　要：在实际应用部署过程中,回声状态网络中储备池的内部突触容易受到应用系统故障的影响而发生故障,偏离预期的性能。因此,提出一个面向突触故障的回声状态网络容错模型。分析模型中一些故障的突触(权重扰动)的行为特征,考虑发生故障的实际场景,采用检测机制检测并定位这些故障的突触,最后通过容错策略使其恢复。仿真实验结果表明,针对不同特点的时间序列数据,通过容错性能评估、统计分布和短期记忆能力对比,提出的容错模型是有效的。此外,通过数学理论推导证明出储备池突触故障容错的边界条件。During practical application deployment,the internal synapses of reservoirs in echo state networks were susceptible to failure and deviation from the expected performance due to application system failures.Therefore,a fault-tolerant model for synaptic faults in echo state networks was proposed.There were some faulty synapses(weight perturbations)in this model.Their behavioral characteristics were analyzed,the actual scenarios in which the failures occurred were considered.A detection mechanism was used to detect and locate these faulty synapses,and finally,they were recovered by a fault-tolerance strategy.The simulation results showed that the effectiveness of the fault-tolerance model was verified by fault-tolerance performance evaluation,statistical distribution,and short-term memory capacity comparison for time series data with different characteristics.In addition,the boundary conditions of fault tolerance for reservoir synapses failures were proved by mathematical theory derivation.

关 键 词：回声状态网络 突触故障 容错性 时间序列预测 短期记忆能力 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]