Knowledge-Guided Graph iTransformer Enhanced by Reinforcement Learning for Industrial Fault Diagnosis

Complex industrial processes are characterized by strong coupling, nonlinear interactions, and dynamic causal dependencies between variables, posing significant challenges for accurate fault diagnosis. Conventional data-driven methods often fail to effectively exploit structural prior knowledge, limiting their ability to model complex spatiotemporal relationships. To address this issue, this paper proposes a graph iTransformer-based fault diagnosis method enhanced by reinforcement learning (RL) optimization. First, a signed directed graph (SDG) is constructed to represent the causal topological relationships between process variables, and a graph embedding algorithm is employed to generate positional encodings, enabling the incorporation of structural prior knowledge into the time series modeling framework. Subsequently, a graph iTransformer model is developed by treating individual variable sequences as tokens, thereby enhancing the learning of multivariate dependencies and complex spatiotemporal features. Furthermore, a multi-dimensional discrete Q-network (MDDQN) is introduced to jointly optimize the iTransformer modules through adaptive collaborative parameter tuning. Experiments conducted on a three-phase flow process dataset demonstrate that the proposed method achieves superior fault diagnosis performance compared with existing approaches. The results validate the effectiveness of the proposed framework in improving fault diagnosis accuracy and generalization performance for complex industrial processes.