Fault Diagnosis of On-board Equipment in CTCS-3 Based on CNN-LSTM Model

Abstract

Data-driven methods based on deep learning have achieved remarkable results of fault diagnosis of train control system due to their superiority in feature extraction. However, it still faces uneven data distribution problem, which afects the detection accuracy of fault diagnosis. In this paper, by considering different failures both in system and subsystem level of train control system, we propose a novel two-stages fault diagnosis method based on a combination of Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM). Firstly, samples are obtained by segmenting and vectorizing the text form faulty data set, and fed into the proposed CNN-LSTM model. Then, in the first stage, the features of the processed data are extracted through the CNN layer, whereas the correlation between the sample data are derived through the LSTM layer. Thus, the classification of first-level faults, respect as system level, are realized with high accuracy of diagnosis. Finally, in the second stage, to solve the problem of data imbalance, we reconsider part of data from the CNN layer, and put them into the new LSTM layer for secondary faults diagnosis. We apply this method on a real CTCS-3 On-board equipment and the experimental results show that the accuracy rate of our proposed model reaches 96.7% and the accuracy of small data faults is also higher when compare with other neural network models,such as TextCNN, ANN, LSTM and RNN.