STAD: Multivariate Time Series Anomaly Detection Based on Spatio-Temporal Relationship

Anomaly detection for multivariate time series is a very complex problem that requires models not only to accurately identify anomalies, but also to provide explanations for the detected anomalies. However, the majority of existing models focus solely on the temporal relationships of multivariate time series, while ignoring the spatial relationships among them, which leads to the decrease of detection accuracy and the defects of anomaly interpretation. To address these limitations, we propose a novel model, named spatio-temporal relationship anomaly detection (STAD). This model employs a novel graph structure learning strategy to discover spatial features among multivariate time series. Specifically, Graph Attention Networks (GAT) and graph structure are used to integrate each time series with its neighboring series. The temporal features of multivariate time series are jointly modeled by using Transformers. Furthermore, we incorporate an anomaly amplification strategy to enhance the detection of anomalies. Experimental results on four public datasets demonstrate the superiority of our proposed model in terms of anomaly detection and interpretation.