Unsupervised anomaly detection system for railway turnout based on GAN

With the rapid development of society, the railway system plays an important role in human life, and the safety of railways has become an extremely important task. As we all know, the switch is one of the important equipment to ensure the safe operation of trains. Real-time detection of the turnout current plays a vital role in train safety. However, the previous signal-based processing methods require a large number of feature engineering, which greatly increases the workload of pre-processing. Some neural network-based methods show good performance, but for the time series data of the switch, these methods cannot fully extract its local features, resulting in poor information loss and poor prediction accuracy. Based on the generational confrontation network, this paper proposes a kind of unsupervised anomaly detection system. We combine the one-dimensional convolution with the Generative Adversarial Networks (GAN). The one-dimensional convolution network can effectively extract the local features of the time series. The GAN can self-game learning to sample distribution and is better than self-encoder and other models, which improves the accuracy of prediction. In the real railway system, abnormal data is extremely rare and varied, while unsupervised learning does not require label data, and it can well learn the distribution of normal samples. The system improves the efficiency of the staff, accurately diagnoses the switch, greatly shortens the processing time, and avoids the blindness in maintenance. Using our model on the turnout data, the accuracy rate is 0.992, the recall rate is 0.815, and the F1 score is 0.895.