Spatial and Temporal Correlation Aware Network for Anomaly Detection in Multivariate Time Series

With the development of the industrial field, cyber-physical systems generate an increasing amount of multivariate time series data(MTS), which reflects the operational state of these systems. Effectively capturing the underlying temporal dependencies and inter-variable correlation in such data is critical for accurate anomaly detection. However, some existing methods assume that all variables are interconnected, resulting in the extraction of redundant correlation. In addition, while some methods consider temporal dependence and inter-variable correlation, the spatio-temporal correlation is ignored, which limits the model's capability to extract normal patterns in the time series. To address the above problems, We propose STCA which is designed to be aware of spatial and temporal correlations within the time series. STCA contains a graph structure learning module to help the model extract more accurate dependencies between variables. Besides, it designs a spatio-temporal correlation aware module for extracting normal patterns between temporal dependencies and inter-variable correlation. We conducted experiments with baselines on three different real datasets, which demonstrate the superior anomaly detection performance of our proposed method STCA.