| The point machine equipment is one of the most important components of the railway signal system.With the rapid development of high-speed railways in our country,the people’s growing demand for fast and safe railway operation has led to the direct diagnosis of the point machine equipment.It affects the safety of railway operation and the efficiency of transportation in China.At present,the fault diagnosis of railway signal equipment in China is still at the stage of fault repair and planned repair.How to use the new technology and the existing large amount of maintenance log data to improve the efficiency of railway signal equipment fault diagnosis is a still difficult problem for the current researchers in the railway field.The field maintenance log data of the point machine is a detailed record of the phenomenon of the point machine failure,the cause of the failure,and the maintenance situation recorded by the employees working on site.Due to the lack of effective processing technology,these important data generated from the front line have been put on hold for a long time.Therefore,this paper makes full use of the existing fault log data to study the fault diagnosis method of the ZYJ7 point machine.The work done is mainly as follows:First,before fault diagnosis,pre-process the faulty text data.In this thesis,we use the jieba word segmentation tool that incorporates a custom dictionary for Chinese word segmentation.An improved LDA theme model based on a priori knowledge is proposed,and a fault feature vocabulary more in line with the actual situation is obtained,and then the word2 vec model is used for word vector training to provide input data for the diagnostic model.Secondly,a fault diagnosis model of the ZYJ7 point machine based on the deep residual CNN network is established,and the first-level fault and the second-level fault are diagnosed based on the layered fault diagnosis idea.The analysis of the experimental results shows that the deep residual network has a good diagnosis effect on the first-level failure mode.Because the CNN model has poor ability to learn time series data,the obtained second-level fault diagnosis effect is not ideal.Finally,based on the shortcomings of the deep residual network model,a CNN +LSTM model was built to diagnose the secondary failure mode.This model solves the problem that the deep residual network has a poor ability to learn the correlation between two levels of fault labels.Because of the category imbalance in the second-level fault data,an integrated learning framework is introduced,and a CNN +LSTM fault diagnosis model based on boosting integration is established to improve the generalization ability of the model.At last,through comparison experiments with SVM,BP and LSTM models,it is verified that the first-level and second-level fault diagnosis models proposed in this thesis have high diagnostic accuracy. |
