Multiple Convolutional Recurrent Neural Networks For Fault Identification Of High-speed Train Bogie

With the rapid development of high-speed railway,the traffic density has increases year by year,and the security operation of high-speed train(HST)has become particularly important.As the key moving component of HST,bogie is composed of frame,wheel set,axle box,braking device,driving device and spring-suspension device,bearing various loads and forces,reducing abnormal vibration and impact,and ensuring safe and stable operation of train During HST’s service period,bogies must withstand the disturbance of track irregularities for a long time,presenting abnormal conditions such as performance degradation and even complete failure.For monitoring the safe operation state of HST,the train collects state information by installing sensors,but the fusion degree of multi-sensor information is low,the multi-source heterogeneous information can not be fully utilized,and the reliability of traditional fault diagnosis is poor.In view of the above problems,this paper mainly studies the fault type identification and performance degradation evaluation of bogie based on the multiple convolutional recurrent neural network(M-CRNN),the main contents include:1)fault analysis of HST bogie based on signal processing methodFirstly,the mechanical structure,fault mechanism and fault effect of bogie are introduced Secondly,the bogie model and data collection scheme of SIMPACK simulation experiment are described in detail.In addition,the time-domain and frequency-domain characteristics of bogie fault singles are analyzed.Finally,the time-frequency characteristics of the bogie signals of fault state and performance degradation are extracted by the combination method of empirical mode decomposition(EEMD)and autoregressive(AR)spectrum analysis.Then the fault characteristic frequencies of fault state and performance degradation are revealed,and the fault analysis of HST bogie is realized2)fault identification of bogie based on 1 dimensional convolutional neural network(1D-CNN)methodIn order to optimize the feature extraction efficiency of bogie fault signal and improve the accuracy of fault recognition,an intelligent fault recognition method of bogie based on 1D-CNN is proposed.1D-CNN method takes multi-source and heterogeneous bogie signals as network input,adaptively extracts high-dimensional features of fault signals,completes information fusion among multi-sensors,and realizes intelligent fault identification of bogie The accuracy of 1D-CNN method in bogie fault identification task is 96.4%,and experiment result verifies the feasibility of 1D-CNN method used in bogie fault identification task3)fault identification of bogie based on convolutional recurrent neural network(CRNN)methodTo make full use of the multi-source and heterogeneous features in bogie fault signals and improve the recognition accuracy of 1D-CNN method,the CRNN framework is designed to realize the bogie fault recognition with high recognition accuracy and fast training speed The core structures of CRNN fault recognition method are the structural feature extraction module of 1D-CNN and the sequential feature extraction module of simple recurrent unit(SRU).Compared with 1D-CNN,SRU,random forest,gradient boosting decision tree and other deep learning or machine learning frameworks,CRNN has the highest efficiency for the recognition of bogie fault:the recognition accuracy and training time of CRNN are 97.4%and 24m35s respectively.Furthermore,under the visual experiment of T-distributed stochastic neighbor embedding(T-SNE),the clustering bias of fault signal samples of bogie in the feature space is less,and the clustering quality is high.Therefore,CRNN has good applicability in bogie fault identification task4)fault identification and performance degradation evaluation of bogie based on M-CRNN methodThe fault state and performance degradation features of bogie signal are interrelated and overlapped,making the fault diagnosis of bogie is more challenging than the fault identification of bogie.For overcoming the above difficulties,the multi-task fault diagnosis method based on M-CRNN is proposed to realize fault state identification and performance degradation evaluation of bogie.M-CRNN method consists of two CRNN modules,learns the internal relationship between fault type and performance degradation,extracts the features of fault state and performance degradation respectively,and ensures the bogie fault diagnosis with high accuracy.For multi-speed comparison experiments,the multi-task accuracy of M-CRNN is 94.6%,the accuracy of fault state and performance degradation identification are 99.1%and 95.4%respectively.Experiment results prove that M-CRNN has better fault diagnosis accuracy and generalization ability.Moreover,for the visualization experiment of T-SNE,the signal features extracted by M-CRNN also have better clustering qualityFinally,based on PyQt,Visual Studio 2015 and other softwares,the intelligent fault diagnosis human-machine interface of bogie is developed,which provides an effective solution for the safe operation and maintenance of HST.