S-DGAT: A Spatially Enhanced Dynamic Graph Attention Network for Multi-Level Anomaly Detection

As industrial systems generate increasingly large and structurally complex multivariate time-series data, anomaly detection is essential for ensuring operational safety and effective management. However, these data often exhibit high dimensionality, strong inter-variable coupling, and temporal dynamics, which pose substantial challenges for traditional methods of modeling complex spatiotemporal dependencies and variable interactions. To address these challenges, this study proposes a Spatially Enhanced Dynamic Graph Attention Network (S-DGAT). The proposed model integrates node attributes, spatial position information, and temporal dynamics and introduces a spatial bias into the graph attention mechanism to enhance the influence of physical topology on neighborhood information exchange. Initially, graph structures are built based on statistical correlations, and adjacency relations are dynamically updated within sliding time windows by integrating attention weights with PPMI-based co-occurrence information. This mechanism improves the model’s ability to capture the dynamic interactions and structural evolution among the variables. Furthermore, S-DGAT introduces a multigranularity anomaly detection framework encompassing point-level, segment-level, and pattern-level analyses, enabling hierarchical detection and interpretable characterization of diverse anomaly types. The experimental results show that the proposed approach consistently outperforms the representative models on multiple benchmark datasets in terms of accuracy, robustness, and interpretability. These findings confirm the effectiveness and generalization capability of the proposed model, offering a promising direction for anomaly detection in multivariate time series.