Fault Diagnosis of CTCS-3 On-Board Logic Function Based on SSGAN Framework

As a typical safety-critical system, the CTCS-3 on-board system has complex logic functions, in which any faults happened may lead to huge loss of life and properties. Fault diagnosis is an important means to improve the safety and reliability of the train control system, thus can gain insight into its logical operating mechanism. Because of the overfitting of fault data with low sampling frequency, the small amount of CTCS-3 on-board dataset makes it a great bottleneck for traditional fault diagnosis method application. Based on the high time-series characteristic in logic function fault data, this paper proposes a novel Semi-Supervised Generative Adversarial Networks(SSGAN) fault diagnosis framework that introduces One-dimensional convolutional neural network(1D CNN) and Long and Short Term Memory Network(LSTM) to overcome the problem. In this framework, the semi-supervised training method and the constant fitting and generation of the data by the generator, the problem of over-fitting can be solved effectively. The 1D CNN introduced in the discriminator follows time to convolve the data, and LSTM achieves long-term memory of relevant information by selectively remembering and forgetting prior information, thus extracting temporal features. We apply this framework to real train operation fault data to diagnose on-board logic function faults. The experiment indicates that despite a 50% labeled rate, the test set achieves an accuracy of 97.2%, effectively alleviating the over-fitting problem caused by small data volumes. Compared with other commonly used supervised fault diagnosis methods such as SVM, CNN-LSTM, etc., our model achieves a better F1-score.