Research On Intelligent Diagnosis Method Of Turnout Fault Based On Deep Learning

With the country’s vigorous development of railways,by the end of 2020,the total length of railways in China was about 146,300 km,of which the total length of high-speed railways was more than 37,900 km,accounting for more than two-thirds of the total length of high-speed railways in the world The total mileage of urban rail transit is about 7978.19 km.Turnout is one of the basic equipment of the outdoor equipment of the railway signal system.It controls the train’s steering,ensures the train’s moving safety and arriving on time.Whether the turnout is out of order or not has a direct influence on the train’s safe operation and running efficiency,so it is very meaningful to study the fault diagnosis of turnout.When the turnout is in trouble,it is usually the staff members who observe the curve data collected by the centralized monitoring system according to their own work experience and put forward the corresponding solution,but it is easy to miss or misjudge,and the fault identification time is long,can not achieve the real-time processing,may affect the normal operation of the train,this kind of turnout fault diagnosis method is no longer suitable for the current high-speed railway automation,the intelligent development direction.Therefore,a deep learning model based on LSTM and 1DCNN is proposed to diagnose the fault of turnout.Firstly,taking S700 K switch machine as an object,the structure and action principle of switch machine,power curve collection principle and process,fault type and cause analysis are analyzed.Secondly,it introduces some knowledge about deep learning,the structure of recurrent neural network and long-term and short-term memory network and their reverse propagation process,establishes the structure model of long-term and short-term memory network,integrates feature extraction and fault classification,the time characteristic of the power curve is extracted as the feature,and then the fault diagnosis is carried out by Softmax.Thirdly,the structure of the convolutional neural network and its back propagation process are introduced,and the structure model of the convolutional neural network is established,which integrates feature extraction and fault classification,and optimizes the network parameters.The convolution layer and the pooler layer extract the adaptive features of the original time domain signals and capture the spatial dimension information of the signals.The regularization Dropout is used to enhance the generalization ability of the model and improve the accuracy of the real-time fault diagnosis of the turnout,t-SNE visualization method is used to reflect the validity of feature extraction.Finally,the structure model of LSTM-1DCNN is established.In the feature extraction,the power curve is extracted with self-adaptive time and space features,and then the fault is classified with Softmax.After experimental simulation,the accuracy of LSTM model is not high,the LSTM-1DCNN model is better than LSTM-1DCNN model in fault accuracy,and the training time of the model is close to each other,so the LSTM-1DCNN model can be used for fault diagnosis of turnouts.