Anomaly Detection Method for MVB Network Based on Variational Autoencoder

Multifunction Vehicle Bus (MVB) is used to transfer information among devices in the train communication network, whose anomaly will endanger the safety of train operation. Based on the analysis of MVB common faults, an anomaly detection method was proposed for the MVB network based on variational autoencoder (VAE), where the physical layer waveforms of MVB signals collected were directly used as input of VAE model and the reconstruction error of the VAE model was defined as the basis of the MVB network anomaly detection. In the training phase, the VAE model was trained by only using the normal data in the semi-supervised learning manner, which can solve the problem of the lack of labeled anomaly samples in practice. The health indicator of the MVB network node was designed according to the reconstruction error of the normal data of the MVB network, and the kernel density estimation method was applied to determine the decision threshold in this case only normal samples were provided without relying on expert experience. The experimental results show that the proposed method, capable of handling the high-dimensional samples and learning the internal features of the MVB waveforms effectively, has higher performance than the traditional methods in anomaly detection. © 2022, Department of Journal of the China Railway Society. All right reserved.