Railway Turnout Fault Diagnosis Method Based On Deep Learning

The railway transportation is in a key position in our country’s economy and development.As the scale and the speed of the railway increased,safety-related issues have become increasingly prominent.In the railway signal equipment,turnout is one of the important components.Due to the changing environment together with the frequent usage,the possibility of failures in turnout is also greatly increased,thus it is necessary to research on fault diagnosis methods of railway turnout.At present,fault detection achieved by the most existing methods relies on the manual judgment of the monitored current curve.On the one hand,this kind of method costs large human resources.On the other hand,the result is greatly affected by manual experience,which may cause misjudgment.With the development of big data,deep learning has stood out due to its advantages in feature extraction and generalization capabilities.Aiming to solve the problems,such as difficult to select fault feature of the current,insufficient feature extraction of the current,hard to obtain labels of the current,three railway turnout fault diagnosis methods based on deep learning are proposed in this paper.The main work is as follows:(1)Turnout fault diagnosis method based on LSTM and attention mechanism.Firstly,segments are obtained by implemented overlapping sampling on the collected turnout currents.Secondly,LSTM is used to extract features from each segment.Then,an attention mechanism network is constructed to calculate the contribution degree of the segments,and the weight is assigned to feature of each segment.Finally,features of each segment are linearized based on the assigned weight to obtain the label of each current.This method does not require manual feature selection,and an end-to-end railway turnout fault diagnosis model is constructed to achieve a high accuracy.(2)Feature extraction method of railway turnout current signal based on graph autoencoder.Firstly,cosine similarity is used to model the collected turnout current signal into a graph.Then,a graph autoencoder model is constructed to extract the feature of the current signal,which includes an encoder constructed based on the graph convolutional neural network layer to achieve feature extraction of the graph structure and the dot product is used as decoder to reconstruct the graph.Next,the graph autoencoder model is trained to obtain the current signal features.Finally,the clustering results of each sample are obtained by using hierarchical clustering.This method can fully extract the features of railway turnout current signals without labeled samples,and carry out preliminary fault diagnosis by clustering algorithm,which can help realize preliminary analysis and category representation in the case of massive data.(3)Semi-supervised railway turnout fault diagnosis based on graph convolutional neural network and LSTM.Firstly,LSTM is used to extract the feature of the current signal and the feature matrix is obtained.Based on the extracted features,KNN method is used to construct the current signal samples into a graph with an adjacency matrix is obtained.Then,a graph convolutional neural network fault diagnosis model is constructed to classify the fault samples based on the feature matrix and adjacency matrix.This method combines the advantages of LSTM for feature extraction of time series data and graph convolutional neural network for sufficient feature extraction of data,which improves the problem of low accuracy in unsupervised fault diagnosis and overcomes the problem of difficulty acquiring labeled samples in practical situations.