基于DCDAE-BiLSTM模型的转辙机故障诊断方法研究  

Research on Switch Fault Diagnosis Based on DCDAE-Bi LSTM Model

作　　者：刘可兴 黄海于[1] Liu Kexing;Huang Haiyu

机构地区：[1]西南交通大学信息科学与技术学院,成都611756

出　　处：《铁道通信信号》2021年第8期87-91,95,共6页Railway Signalling & Communication

基　　金：国家自然科学基金青年基金.复杂环境下铁路轨道在线故障诊断的迁移学习方法研究.62003313。

摘　　要：在分析道岔转辙机常见故障类型的基础上,提出一种深度卷积降噪自编码器与双向循环神经网络相结合的故障诊断方法。以某城市地铁S700K型道岔转辙机采集的道岔功率数据作为训练集和测试集,利用Keras框架搭建上述网络模型;将训练数据集输入到深度卷积降噪自编码器提取高阶抽象特征,然后作为双向循环神经网络的输入进行训练;通过测试数据集从准确率、误报率、ROC曲线等指标进行试验验证。结果表明采用上述模型故障诊断准确率达到98.82%;通过对不同算法的对比分析,证明了该方法的有效性。Based on the analysis of common types of switch fault,a switch fault diagnosis method combining deep convolution noise reduction self encoder and bidirectional recurrent neural network is proposed.Taking the switch power data collected from S700 K switch machines in one city subway as the training set and test set,the neural network model is built in use of Keras framework.The training data set is input into the deep convolution de-noising self encoder to extract high-order abstract features,and then used as the input for training the bidirectional recurrent neural network.Finally,the test data set is used to verify the accuracy,false alarm rate,ROC curve of the switch fault diagnosis method.And the results show that the accuracy of fault diagnosis using the switch fault diagnosis method is up to 98.82%.Through comparative analysis of different algorithms,the effectiveness of the method is proved.

关 键 词：转辙机 故障诊断 卷积神经网络 循环神经网络 

分 类 号：U283[交通运输工程—交通信息工程及控制]