基于神经网络多步时序预测的非线性系统故障诊断(英文)

Fault Diagnosis of Nonlinear Systems Using Multistep Prediction of Time Series Based on Neural Network

提出一种新的基于神经网络多步时序预测的非线性系统故障诊断方法 .该方法先利用回归神经网络对多个传感器检测序列并行进行多步预测 ,再由多步预测序列和传感器检测序列生成历史残差序列和预测残差序列 .最后 ,根据统计定义的几个决策指标进行故障检测与诊断 .与其它方法相比 ,本文方法所需信息较少、可诊断的故障较多 .仿真表明该方法是有效的 ,可有效地增强故障信息、抑制非故障信息 .

A new approach to fault diagnosis of nonlinear systems, which uses multistep prediction of time series based on neural network, is presented in this paper. By using recurrent neural network, multistep predictions for sampling series from multiple sensors are performed simultaneously, then two series, namely historical residual series and predicting residual series, are separately constructed from predicting series and sampling series. Lastly, several evaluation indexes used to detect whether fault(s) exist or not in a nonlinear system are derived. Compared with other methods, the proposed method not only needs less information about diagnosis objects, but also can detect more classes of faults. Simulation results show that the method is effective, and can strengthen the fault information and restrain the non-fault information effectively.