Industrial Control Systems (ICSs) are critical to infrastructure security, but detecting anomalies in their multivariate time series remains challenging due to significant limitations in existing methods for modeling hierarchical multi-scale spatio-temporal dependencies and capturing long-range temporal dynamics. This paper introduces a novel framework integrating a Multi-Scale Graph Relational Learner and a DC-Mamba Block to address these gaps. The former employs convolutional feature extraction and self-attention to model hierarchical dependencies across different resolutions, while the latter fuses diffusion convolution with Mamba to efficiently capture long-range temporal dynamics in industrial processes. Experiments on two benchmark ICS dataset, Secure Water Treatment (SWaT) and Water Distribution (WADI), demonstrate the framework’s superiority. On SWaT, it achieves a precision of 88.64% and an F1-score of 0.86; on the more challenging WADI dataset, it yields a precision of 94.41% and an F1-score of 0.56. Ablation studies validate the indispensable roles of both modules in enhancing detection accuracy. This work provides a robust solution for ICS anomaly detection, emphasizing the effectiveness of multi-scale graph learning and adaptive temporal modeling in industrial cyber-physical systems.