Research And Application Of Remaining Useful Life Prediction Of Railway Wagon Wheels Based On Domain Adaption Adversarial Learning

With the continuous updating and upgrading of modern heavy-haul high-speed railway and freight car transportation equipment and technology,the life cycle of freight car operation is becoming shorter and shorter.Problems such as wear,crack and deformation are easy to occur in the running process.At present,the maintenance system of railway freight cars is mainly based on "planned preventive repair",and is gradually turning to "state repair" mode.The mode of "condition repair" is based on the data of truck mileage tracking,wear monitoring,shaft temperature measurement and so on,and adopts a variety of monitoring methods to carry out all-round dynamic monitoring of the truck in operation.As the key parts of truck operation,if the failure is not found in time,the operation performance of the truck will be affected at least,and the safety accident will be caused at worst.Therefore,the research on wheel life prediction of heavy-duty truck has become a key problem to be solved in the field of track equipment maintenance.In this paper,an adversarial learning network is proposed to solve the domain adaptive problem of label poverty in practice.Adversarial Discriminative Domain Adaptation(ADDA)network combined with Long Short-Term Memory(LSTM)network was used to extract features,and the overall framework of the remaining service life of truck wheels was proposed.The main research contents of this paper are as follows:(1)Due to the large demand of data for model training in practical application,and the large amount of energy and financial resources for collection,the unsupervised learning method is proposed.LSTM network is used for feature extraction,and crossdomain prediction is carried out by admittedly differentiated method.Using labeled source domain data and unlabeled target domain data as inputs to the model,the ADDA network model shows an accurate prediction rate.At the same time,in the unsupervised domain adaptive methods,compared with other domain adaptive structures,the proposed LSTM-ADDA has advantages in the prediction of the remaining service life.(2)In the actual operation of freight cars,it is difficult to obtain a large number of label data of freight wheels due to less equipment on line and short running time,which brings challenges to the study of the remaining service life.Therefore,the network model of LSTM-ADDA is proposed to predict the data of wheels.It has been proved that the proposed broad adaptive domain network is superior to the non-domain adaptive method for life prediction,and the cross-domain capability of the proposed method in residual service life prediction is verified.(3)Based on the research of the residual service life prediction model of heavy-duty truck wheels,the wheel life management system was integrated into the comprehensive discriminant model system of condition repair diagnosis decision.In the process of truck operation,the monitoring station dynamically uploads the wear data of heavy-duty truck wheelsets and inputs it into the wheel life prediction module to give the predicted maintenance time,reduce the cost of operation and maintenance,and increase the effective gain brought by the wheels for railway freight transportation.