Spatio-temporal interactive network: A knowledge-guided federated learning for fault propagation path identification

As a cutting-edge technology, federated learning (FL) demonstrates great potential in industrial fault propagation path identification, which uncovers the intrinsic relationships between variables by collaborating across multiple clients while preserving privacy, contributing to process safety and system optimization. However, existing FL methods do not take into account the interactive fusion of multi-scale features, resulting in insufficient feature mining and low utilization. In addition, purely data-driven FL naturally interpret the analyzed causality as propagation paths. In this paper, a knowledge-guided FL method with spatio-temporal interactive network is proposed for fault propagation path identification. First, we develop a knowledge-guided comprehensive graph construction module to provide the required adjacency matrix for subsequent model. After that, spatio-temporal soft attention prediction model is designed to mine spatio-temporal interactive information and the communication efficiency and model accuracy are further reconciled by residual-wise adaptive parameter aggregation scheme. Then, the causality is accurately characterized with the collaboration of mechanism knowledge and data, which makes the fault propagation path more explainable. Finally, in order to verify the effectiveness of the method, we carried out experiments on the simulation of Tennessee Eastman process and the real-world coal mill unit dataset.