Research On LSTM-based Fault Diagnosis Of High-speed Railway Bogies

In recent years,China has vigorously carried out the construction of high-speed railway.The rapid growth of high-speed railway operating mileage and the improvement of operating speed make the safety technology of high-speed trains face great challenges.Abnormal state of bogie is often reflected in the abnormal vibration of bogie and car body.Recurrent neural network(RNN)is a kind of network that can extract its internal correlation characteristics according to the time series data.Since the bogie vibration data is the time series data,the use of RNN can extract the internal information of the data essentially.Long Short Term-Memory(LSTM)is an upgraded version of RNN,which can control the amount of data input,memory and output,connect previous data to analyze and extract the characteristics of current data,so as to improve the efficiency of RNN.In this paper,based on LSTM network,a fault diagnosis model and a fault regression fitting model for vibration signal of high-speed train bogie are established.It is the first application of LSTM network in the field of fault diagnosis and regression fitting of high-speed train bogies.The main contributions of the thesis are as follows:Firstly,for the existing fault vibration data of key components of high-speed train bogie,a fault diagnosis classification model based on a variety of neural networks such as LSTM network is built,and multi-channel fusion fault diagnosis is conducted for a variety of fault data under the condition of constant speed.Secondly,in addition to the 58 channels inherent in the original vibration data,a certain amount of other neural network fault diagnosis classification results are added as new channels,which are used as new data for classification operation to improve the robustness of the model and the confidence of classification results.Then,in view of the performance-degradation vibration data of the key components of the existing high-speed train bogie,a series of LSTM based multiple regression fitting model are set up.Further,single-channel and multi-channel-integration-based regression analysis are performed,where a variety of indices are provided in the experiment to evaluate the resultant performanceFinally,the experimental results of the above fault diagnosis and fault regression fitting were demonstrated.In addition,the difficulties encountered in the experiment and the points to be improved were summarized.