Reinforcement learning driven adaptive graph construction for fault diagnosis of chemical processes

Faults in chemical processes industries can lead to significant economic losses, environmental damage, and safety risks, making effective fault diagnosis critical for ensuring operational reliability. Graph-based learning methods have shown great promise in capturing complex relationships in time-series data for fault diagnosis. However, traditional data-driven fault graph construction methods often suffer from limited expressive potential due to their reliance on single rules. To address this limitation, a node-distance-adaptive Graph Reinforcement Learning (GRL) method Graph Double Deep Q Network (GDDQN) is proposed, which leverages the decision-making capabilities of Reinforcement Learning (RL) to optimize graph construction strategies. Furthermore, to handle the noise and inherent similarities in fault time-series data, an Enhanced Graph Convolutional Network (EGCN) is introduced. By encouraging angular diversity in the weights, the EGCN model demonstrates superior feature extraction and generalization capabilities, effectively processing node features represented by time-series signals. Combining these advancements, we developed the GDDQN-EGCN algorithm, which effectively integrates adaptive graph construction with robust feature processing. Extensive experiments on two datasets from the Tennessee Eastman (TE) process validate the outstanding performance of the proposed method in fault diagnosis, showcasing its potential for real-world applications.