Semi-supervised Anomaly Detection for Real-World Railway Surveillance via Restraining the Overgeneralization

Railway plays an increasingly important role in interurban journeys. The operation state of railway components is particularly related to railway safety. The traditional methods rely on manually inspecting the components periodically is labor-consuming. Recently, some supervised learning methods have been applied in the railway industry to detect potential defects in components. However, the lack of defective data and the sample imbalance limit the application of the relevant methods. In this paper, we propose a novel semi-supervised anomaly detection method for railway surveillance. We train an autoencoder with specific reconstruction capability using only normal samples, and the samples that cannot be well reconstructed are considered as anomalies. It can be used to detect potential and undefined anomalies of railway components. The proposed method imposes multi-space constraints on a modified Adversarial autoencoder to restrain the overgeneralization. We provide the results of our approach on some public datasets as well as one that we gathered from the real-world high-speed railway and named HRBDD. Experiments show that the proposed method yields a strong performance for railway anomaly detection in real-world surveillance.