A Dynamic Graph Attention-Based Unsupervised Bi-Learning Framework for Anomaly Detection of Urban Rail Vehicle Door Systems

The anomaly detection of urban rail vehicle door systems is crucial for the security and reliability of rail vehicles, and unsupervised methods have received particular attention in anomaly detection due to the scarcity of anomaly data. However, existing studies have the following limitations: (1) The dynamic interrelationships between different sensor variables are not explicitly considered, which is detrimental to the learning of normal patterns. (2) The detection performance is limited due to the reliance on manual threshold settings and the lack of confidence measures for different sensor variables. To overcome the abovementioned limitations, a dynamic graph attention-based unsupervised bi-learning framework for the anomaly detection of rail vehicle door systems is proposed in this paper. First, a dynamic graph attention block (DGAT block) is designed to model dynamic interactions between sensor variables and fuse information during signal coding. Then, a bi-learning framework is developed to automatically learn thresholds and quantify the confidence of sensor variables for anomaly detection. The proposed method is verified on a real rail vehicle door system. Compared with the state-of-the-art autoencoder-based methods that are applicable to multivariate time series data, the proposed method improves the best competing precision and F1-score by 7.7% and 4.3%, respectively, while reducing the false positive rate (FPR) by 88.9%.