<p>With the rapid development of the Industrial Internet of Things, industrial control systems (ICS) generate substantial amounts of multivariate time series data characterized by complex spatiotemporal coupling. This data not only encompasses dynamic temporal evolution patterns but also includes spatial correlation characteristics induced by the system’s topological structure. This paper addresses the anomaly detection problem of multivariate time series in industrial scenarios by proposing an anomaly detection framework based on spatiotemporal joint modeling. By introducing a graph structure driven by mutual information and integrating graph attention networks with Transformer encoders, we achieve collaborative modeling of multiscale spatiotemporal features. In feature space modeling, we propose a sampling method for the adjacency matrix based on mutual information estimation, effectively capturing the nonlinear dependencies among variables. In temporal modeling, we employ multiscale positional encoding and an adaptive attention mechanism to enhance the representation capability of long-range dependencies in industrial data. Experimental results demonstrate that GTR outperforms other methods in industrial anomaly detection benchmarks. Ablation experiments validate the effectiveness of graph structure learning and Transformer temporal modeling, while case studies further highlight the interpretability advantages of the constructed graph structure in elucidating physical mechanisms and characterizing anomaly propagation paths.</p>