A Novel Spatio-Temporal Reconstruction Approach for Multivariate Time Series Anomaly Detection in Heavy Industry Equipment Monitoring

Aiming at the challenges of critical equipment condition monitoring in heavy industry, including the lack of anomaly labels, high monitoring performance requirements, and unclear correlations between sensor data, a spatio-temporal reconstruction network named MRGSTR is proposed. MRGSTR integrates the Mamba network and Residual Graph Convolutional Network (ResGCN) to reconstruct multivariate time series data from equipment sensors, enabling condition monitoring without anomaly labels. The Mamba network improves computational efficiency and reduces resource consumption, whereas ResGCN captures the correlations among sensor variables. By analyzing reconstruction errors and three anomaly scoring metrics, MRGSTR identifies single-series outlier anomalies, single-series periodic anomalies, and cross-series correlation anomalies, corresponding to the three primary failure types of critical equipment. Finally, the accuracy of the MRGSTR algorithm is verified by the actual case of continuous caster mold.