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

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.