Research On Intelligent Fault Diagnosis Method Of Railway Train Wheelset Bearing Based On Deep Learning

With the increase in passenger and freight volume,railway trains are developing in the direction of enormous capacity,complication,high speed and automation.Hidden hazards and failures are more likely to endanger critical components related to operational safety,and further,cause severe accidents due to the increment of frequency and intensity of equipment utilization.Rolling bearings are the most widely used power transmission elements in the running part of railway trains,and their performance has a direct impact on the operational safety of rail vehicles.Therefore,it is of great significance to study stable and reliable wheelset bearing fault diagnosis methods at this stage.This paper takes the wheelset bearings of railway rolling stock as the research object,and studies the fault diagnosis problems in the case of unbalanced data during the analysis of vehicle wheelset bearings under different application scenarios and changing conditions.Correspondingly,two fault diagnosis methods are proposed.The main research contents include:First,the conditional adversarial generative networks(CGAN)are introduced into the data expansion process of rolling bearing vibration signals,aiming at unbalanced data distribution caused by the scarcity of fault samples in the offline analysis application scenarios of railway trains.The framework of CGAN is employed to build a bearing vibration data expansion model based on the analysis of the new sample generation mechanism of CGAN.The model is trained with the laboratory bearing data and rolling stock wheelset bearing datasets of Case Western Reserve University.Subsequently,the effectiveness of the data augmentation method is verified by generating a signal quality test.Second,an end-to-end bearing diagnosis model is constructed by the residual network combined with the attention mechanism regarding the problem that the noise in the field datasets is difficult to effectively extract the signal features.To fix the problem of unbalanced train wheel bearing offline data,CGAN is implemented in the data expansion to reach the equilibrium,and then the diagnosis model is proposed to complete the fault diagnosis task.Hence,the rolling bearing fault diagnosis model under the data imbalance state is created.Two data sets are used to verify that the method solves the problem of unbalanced data in the fault diagnosis of rolling stock wheelset bearings in offline analysis application scenarios to a certain extent.Third,focusing on the problem of reduced accuracy of traditional intelligent fault diagnosis methods and lack of labeled data due to changing working conditions in railway train online monitoring application scenarios,a bearing fault diagnosis method based on improved domain-adversarial transfer learning is proposed.The domain-adversarial neural network(DANN)is enhanced by adopting the central moment difference measure,and the rolling bearing diagnosis model under the variable working conditions is constructed based on the research of transfer learning method of the DANN.Two kinds of actual data are applied to train the improved DANN,and the labeled source domain dataset is used to diagnose and identify the unmarked variable condition target domain data set to realize the variable condition bearing fault diagnosis task in the online monitoring application scenario.The following experiments verify that this method solves the problem of end-to-end fault diagnosis of rolling stock wheelset bearings under variable operating conditions to a certain extent.There are 45 pictures,16 tables and 65 references in the body.