Multivariate Anomaly Detection Based on Spatio-Temporal Graph Structure Perception

Multivariate anomaly detection is essential for ensuring stability in complex systems, where existing methods often fall short in addressing time-delay causal effects—the lag between anomaly occurrence and detection that allows systems to persist in faulty states, leading to escalated failures—and microscopic anomalies, subtle low-intensity deviations that evade capture due to insufficient modeling of nonlinear temporal dependencies and spatial structures. In this work, we introduce ST-GSP, a novel spatio-temporal graph structure perception model that constructs adaptive multivariate coupling networks via heuristic strategies integrating directed similarity measures and clustering algorithms, overcoming limitations of predefined graphs. By fusing convolutional long short-term memory (Conv-LSTM) networks with multichannel graph attention encoding layers, ST-GSP extracts deep spatio-temporal features to explicitly capture lagged causal effects and spatial couplings, while incorporating complex network structural measures such as degree centrality, betweenness centrality, closeness centrality, PageRank, and eigenvector centrality to quantify subtle structural deviations for enhanced microscopic anomaly perception. Reconstruction via variational autoencoders (VAE) combined with peak-over-threshold (POT) dynamic thresholding further bolsters robustness and real-time performance in noisy environments. Extensive experiments on six real-world datasets, including SMD, SMAP, MSL, ASD, SWaT and WADI, demonstrate ST-GSP’s superiority over benchmarks like MTAD-GAT, achieving average improvements of 5.2%, 4.8%, and 4.6% in precision, recall and F1-score, respectively, gaining AUCs of 0.9989 and 0.9986 on SMD and SMAP. This framework not only breaks bottlenecks in lag modeling and anomaly sensing but also provides a more robust, interpretable paradigm for multivariate time series anomaly detection in industrial IoT applications.