A Multi-Scale Spatiotemporal Attention Attack Detection Model in Industrial Cyber-Physical Systems

To address the challenge of accurately identifying attacks on Industrial Cyber Physical Systems, this paper investigates a multi-scale spatiotemporal attention model. The model employs wavelet convolution to extract multi-scale local time-frequency information and utilizes multi-branch Transformer attention heads to capture global dependencies across different temporal scales. A dynamic graph structure is constructed based on multi-scale temporal features, and a multi-stage, multi-modal graph attention network is designed to capture the multi-scale spatial dependencies among ICPS device nodes. The multi-scale spatiotemporal features are then fused and fed into a multi-layer perceptron for prediction, with the prediction error serving as the anomaly score for anomaly detection. Simulations results in PyCharm demonstrate that the proposed model outperforms existing models, verifying its effectiveness in ICPS attack detection.