Fault Diagnosis Of Jointless Track Circuit Based On Deep Learning

In recent years,with the development of society and economy,higher requirements are put forward for railway transportation capacity.As the basic equipment of railway signal,the track circuit is used to send the train operation information from train control system to the train,and complete the train occupation detection and integrity inspection,so it plays an important role in ensuring the safe and reliable operation of the train.With the increase of railway operating mileage in China,the number of track circuit increases.Track circuit equipment works outdoors for a long time,its environment is complex,so it is easy to fail.However,manual maintenance has the disadvantages of low real-time and high risk.Aiming at the problems existing in the existing diagnosis system and research,this thesis realizes the track circuit fault diagnosis based on the existing monitoring data.Based on the above background,the main work of this thesis includes the following aspects:First of all,consulting the relevant standards and field Investigation,which determines the monitoring data that different diagnosis systems rely on,and propose corresponding solutions according to different monitoring systems.After determining the monitoring data,according to the track circuit transmission line model,the monitoring data at different positions are simulated and compared with the actual test data to determine the validity of the model,which is used to generate training data other than the actual measured data.Secondly,in the aspect of ground monitoring data fault diagnosis,for the diagnosis system with outdoor ground monitoring data,considering the noise in the monitoring data,we use stacked denoising auto-encoders to realize fault diagnosis.By adjusting the algorithm structure and parameters,the diagnosis accuracy reaches 99.63%.This thesis focuses on the research of the diagnosis system without outdoor ground monitoring data.The difficulty lies in the real-time online fault diagnosis of the tuning area.At present,the problem is usually realized by the on-board data,but the timeliness is relatively low.By analyzing the working principle of the tuning area,this thesis proposes a solution to extend the fault diagnosis data to the centralized monitoring data of the adjacent area.Due to the lack of fault characteristics,when the algorithm structure and parameters are the same as the outdoor monitoring system,the fault type within the area cannot be identified.The convolution operation is further used to optimize the trestle noise reduction self coding,improve the ability of local feature extraction of the algorithm,and realize the real-time online fault diagnosis including tuning area fault when there is no outdoor monitoring system.Finally,in the aspect of on-board data fault diagnosis,the single wheelset and multi wheelset train shunt model are established.The single wheelset shunting model is used when the normal shunt in whole the section,and the multi wheelset shunting model is used when the poor shunt in the local section.The locomotive signal voltage data of compensation capacitance fault under different operation scenarios is formed.The algorithm adopts convolutional neural network,in the diagnosis test,there is a problem that the accuracy of diagnosis is reduced due to the label error.The above problem is solved by analyzing the principle of algorithm classification and establishing the wrong label data set.The diagnosis effect is good and the validity of the model is verified.