基于LSTM循环神经网络的故障时间序列预测

Exploring LSTM based recurrent neural network for failure time series prediction

有效地预测使用阶段的故障数据对于合理制定可靠性计划以及开展可靠性维护活动等具有重要的指导意义。从复杂系统的历史故障数据出发,提出了一种基于长短期记忆(LSTM)循环神经网络的故障时间序列预测方法,包括网络结构设计、网络训练和预测过程实现算法等,进一步以预测误差最小为目标,提出了一种基于多层网格搜索的LSTM预测模型参数优选算法,通过与多种典型时间序列预测模型的实验对比,验证了所提出的LSTM预测模型及其参数优选算法在故障时间序列分析中具有很强的适用性和更高的准确性。

Effectively forecasting the failure data in the usage stage is essential to reasonably make reliability plans and carry out reliability maintaining activities. Beginning with the historical failure data of complex system,a long short-term memory（LSTM） based recurrent neural network for failure time series prediction is presented,in which the design of network structure,the procedures and algorithms of network training and forecasting are involved. Furthermore,a multilayer grid search algorithm is proposed to optimize the parameters of LSTM prediction model. The experimental results are compared with various typical time series prediction models,and validate that the proposed LSTM prediction model and the corresponding parameter optimization algorithm have strong adaptiveness and higher accuracy in failure time series prediction.