USTG:Multivariate Time Series Anomaly Detection via Unsupervised Spatial-Temporal Graph Learning

Multivariate time series anomaly detection is crucial in many applications, including finance, healthcare, transportation, power grids, and water treatment plants. However, due to the complex interdependencies among sequences,We propose a novel approach called Unsupervised Spatio-Temporal Graph Learning (USTG) for multivariate time series anomaly detection. USTG captures the temporal and variable correlations through the learning of multivariate time series correlations and self-attention mechanisms. It then utilizes graph convolutional neural network to develop a spatio-temporal graph neural network. The data distribution is obtained through a data distribution standardization flow, which compares the distribution differences between anomaly points and normal points to detect anoma-lies. Experimental results demonstrate that USTG can effectively detect and diagnose anomalies, and it can achieve anomaly detection at different time delays.