Research On Turnout Fault Prediction Method Based On Data Mining

The rapid development of Chinese Railway puts forward higher requirements for the maintenance of railway signal equipment.Trains can be converted to different approaches through turnouts.Therefore,turnout as one of the important signal basic equipment in the railway signal system,the failure will seriously affect the safety and operation efficiency of the train.In the current railway field,it is mainly through the centralized signal monitoring system to monitor the turnout equipment,including the action current,voltage and power data of the switch machine.In recent years,the centralized signal monitoring system in some stations also includes the monitoring of the gap in the turnout.When the turnout failure,mainly rely on experienced staff in the field to judge the turnout failure.In order to reduce the turnout failures,the research focus is on the state prediction of turnout.When the turnout does not fail,the operating state of the turnout is known in advance,the turnout failure is predicted based on the model,and the turnout is maintained in advance.On this basis,this paper uses the method of deep learning to combine the convolutional neural network(CNN)with the gated recurrent units(GRU)to realize the state prediction of the turnout.Through the information mining and analysis of non-fault power data of turnout,the non-fault data of turnout is classified by using Mini Batch K-Means and spectral clustering algorithm.Combining Euclidean distance,Pearson Correlation Coefficient and Fletcher Distance,a new evaluation index is established to excavate the relationship between normal,fault and non-fault of turnout.Finally,the convolutional neural network is used to automatically extract the features of the turnout action power curve,which is connected with the GRU to establish the state prediction model of the turnout,so as to realize the fault diagnosis and prediction of the turnout.The main work contents are as follows:(1)Preprocessing of turnout power data.By analyzing the action process of turnout equipment,master the action characteristics of turnout power curve,and break the turnout power curve into four action stages.Export the power data of the turnout in the centralized signal monitoring system.To analyze the characteristics and causes of fault curves for the common faults of the turnout,and to establish a sample data set of turnout faults.(2)Processing of turnout non-fault data.The processing of non-fault data of turnout includes feature extraction,feature selection,pseudo-label construction of non-fault data and so on.Feature extraction is mainly through calculating the time-domain statistical characteristics of turnout power curve.The average value,variance and waveform factor of data points in different operation stages of turnout can represent the calculation method of power curve.At the same time,the average value sum and difference sum of the whole operation curve after Haar wavelet transform are calculated to form the 47-dimensional features of turnout power curve.The pseudo-label of the non-fault data of turnout are constructed by using the Mini Batch K-Means clustering algorithm,and then the features are selected by using the model-based feature ranking method according to the pseudo-label,to screen out the 35-dimensional model-sensitive features.Finally,the classification of non-fault data of turnout is obtained by spectral clustering algorithm,and five non-fault action curves for the turnout are summarized.(3)Establish the connection between various states of turnout.In order to realize the state prediction of turnout,it is necessary to establish the relationship among normal state,non-fault state and fault state.Using the curve similarity method,the correlation between the states of the turnout cannot be analyzed when the Euclidean distance,Pearson correlation coefficient and Fréche distance are used alone.The T-SNE verifies that there is a numerical difference between the states of the turnout and that the states cannot be grouped together.So combining the Euclidean distance,Pearson Correlation Coefficient and Fréche Distance,a new evaluation index is established to calculate the correlation between the normal state,fault state and non-fault state of turnout,excavate the correlation between the states of turnout,according to the conclusions obtained to achieve turnout state prediction.(4)Research on the method of turnout state prediction.According to the mixed data of turnout,CNN-GRU model is established to realize the state prediction and fault diagnosis of turnout.Compared with the feature matrix of other dimensions through experimental simulation,the model accuracy is high when the feature matrix is 40 dimensions,and compared with other iterations,the accuracy of model training is better when iterates 50 times.Therefore,the accuracy of turnout state prediction is 94.2% when the feature matrix is input in 40 dimensions and iterated 50 times.The model simulation of this paper is compiled by PyCharm platform and Python language,and the model is trained and verified in Keras framework.After the establishment of the algorithm model,it has the advantages of simple call,high reuse rate,flexible response to various requirements.It has better superiority.