A Fault Detection System For Track Switch Machines Based On Deep Learning

The switch machine,which controls the direction of the train,is the key equipment that directly relates to the safety of the train.Therefore,when a switch machine is out of order,it is very important to detect the type of the fault and repair it as soon as possible.With the advancement of informationization of railway control system,more and more state data of switch machines are being collected.In this paper,the methods of deep learning are used to classify the state data to determine whether a fault exists and the type of the fault.There are several typical fault modes of switch machines in which the variations of the output power in the entire process of switch action have different characteristics.Traditional intelligent diagnosis methods use artificial features to which there are several disadvantages such as the high cost of feature construction and the insufficiency of feature design.In this paper,we use autoencoder and convolutional neural network?CNN?,two kinds of deep learning models,to automatically extract features and do classification,and the classification accuracy of these models are compared.As to the fault classification method based on the combination of autoencoder and BP neural network,we compare the classification performance of original power data and the encoded power data with different dimensions using BP neural network,finding that the higher the dimension of the data,the better the classifier.As for the CNN-based approach,we draw the sampled power data of switch machines onto graphs,which are then fed into CNN models.Four CNN models?CNN₃₁,CNN₄₁,CNN₅₁,CNN₃₂?different in the number of convolutional layers,in the length of convolutional step and in the number of convolutional kernels are designed.Through experiments,we find that CNN₅₁ is the best and is much better than the combination of autoencoder and BP neural network.Furthermore,the CNN-based ensemble learning is introduced to improve the accuracy and stability of classification.We compare the ensemble method of different CNN models and the ensemble method of different number of the same CNN model,finding that ensemble learning has “the longest plate effect” or “the second longest plate effect” and that the ensemble method based on CNN₅₁ is the best.Finally,we develop a simple and easily-used fault detection system for the switch machine based on the ensemble learning of the model of CNN₅₁.