Knowledge-enhanced spatial-temporal multi-frequency fusion transformer for plant-wide industrial process monitoring

Plant-level industrial process monitoring is critical for ensuring production safety and efficiency. Existing state monitoring models predominantly focus on temporal modeling, neglecting the significance of underlying process structural knowledge and frequency domain information. To address these limitations, we propose a novel knowledge-enhanced spatial-temporal multi-frequency fusion transformer (KSTFformer) for comprehensive plant-wide process monitoring. Unlike existing data-driven methods, KSTFformer integrates prior knowledge with process data to enhance both modeling performance and interpretability. Initially, we decompose the entire process into multiple operating units based on equipment connection relationships, constructing variable correlations as a directed graph. We then design a graph convolution-based structural learning module that captures spatial-temporal dependencies using graph convolutional networks, leveraging the Gumbel-softmax sampling method to learn variable interconnections. Furthermore, a novel multi-band frequency domain feature extractor was innovatively proposed, which refines frequency domain sequences into multiple local subsequences and extracts frequency domain characteristics of variables through channel-independent networks. Finally, the model integrates spatial-temporal and frequency domain features through attention mechanism, enhancing anomaly detection capabilities. The proposed method's effectiveness is validated through two real-world industrial process case studies.