Methodologies For Fault Prediction And Diagnosis Of Railway Turnout System

Railway turnout system(RTS)is indispensable to control the steering of trains and ensure the safety operation.Once the fault occurs,it will greatly affect the transportation efficiency and even endanger the life of passengers.With the railway transportation network is becoming denser and denser,the railway department pays more and more attention to ensure the high reliability of RTS.At present,periodic maintenance is generally adopted according to the maintenance plan in the routine safeguard work of RTS.Such a "planned maintenance" pattern is unable to perform "condition-based maintenance" according to the remaining useful life(RUL)of RTS,which is prone to cause "under maintenance" and "over maintenance".In the work of fault diagnosis and location,the alarm accuracy of centralized monitoring system(CSM)is difficult to meet the filed requirements,and the application of intelligent diagnose algorithm has been restricted by the problem of data imbalance,so it still mainly relies on manual troubleshooting.The purpose of this dissertation is to apply the concept of prognostics and health management(PHM)to the maintenance work of RTS,i.e.,predict its RUL in the normal working stage to provide guidance for "on-condition repair",and effectively deal with the problem of unbalanced fault data in the diagnosis work to improve the efficiency and accuracy of diagnosis.To achieve the above goal,this dissertation makes a research on the fault prediction and diagnosis of RTS.A fault prediction method based on improved sparse auto encoder(ISAE)and gated recurrent unit(GRU)is proposed in the prediction work.The experiment on field data sets show that the method can be well applied to RTS fault prediction.To solve the data imbalance problem,i.e.,turnout’s fault data is much less than the normal data in the diagnosis work,a model based on Borderline-SMOTE(Synthetic Minority Over-Sampling Technique)is performed to generate railway turnout feature data.The fault diagnosis accuracy is validated by combining the generated data and actual data.The main work of this thesis can be summarized as follows:(1)In the work of fault prediction,a prediction scheme of the RUL for RTS based on ISAE and GRU network is constructed.The feature of turnout degradation is extracted by ISAE,and then the health indicator(HI)of RTS is obtained by feature selection and feature fusion.Finally,considering the strong correlation of HI in time series,the GRU network is introduced to "integrate" the health indicators of historical time and current time for fault prediction.The proposed model’s effectiveness and good performance are verified by comparing with different prediction algorithms on multiple indicators.(2)In the research of fault diagnosis,aiming at the imbalance between fault data and normal data,a model based on Borderline-SMOTE(Synthetic Minority Over-Sampling Technique)is utilized for data enhancement after feature extraction.The high similarity of generated feature data and actual feature data are examined in a variety of ways.Then,the fault diagnosis is performed on all the data,and the diagnosis accuracy is more than97%.Finally,the comparative experiments with a variety of algorithms show that this method has certain advantages.(3)The fault prediction and diagnosis software of RTS is designed,and each function is tested by using field data.The results show that the software meets the requirements of field application.Figure 49,Table 12,References 75...