Feature-Decoupled Multi-Task Graph Attention Network for Chemical Process Fault Classification and Localization

Abstract Fault diagnosis of chemical processes has become increasingly critical for ensuring operational safety and efficiency. However, it faces persistent challenges in accurate fault classification and localization, primarily due to feature entanglement across heterogeneous diagnostic tasks and topological misalignment between equipment configurations. To address these limitations, we propose a novel feature-decoupled multi-task graph attention network (FDMGAT), consisting of hierarchical feature decoupling, adaptive topology alignment, and dynamic multi-task optimization. FDMGAT separates domain-invariant shared features from task-specific private features via hierarchical feature decoupling, incorporates node mapping dictionaries and edge mask mechanisms for graph structural adaptation, and utilizes decay-weighted loss balancing to optimize task-specific heads. Experimental validations on two chemical processes demonstrate the superior performance of FDMGAT, achieving average accuracy of 93.38% and 93.00% for fault classification and localization respectively, which outperform the state-of-the-art and classical algorithms like MTL-CNN, M2FN, GAT, CNN, FD-KNN and FR-DT in the field of chemical process fault diagnosis.