Game Attribution-Based Causal Learning Interpretable Networks for Intelligent Fault Diagnosis

In recent years, deep learning-based intelligent fault diagnosis methods have been widely applied and developed for the diagnostic and health management of mechanical systems. However, the black-box nature of these intelligent diagnostic models has significantly hindered their widespread application in risk-sensitive industrial fields. Few studies aim to ensure that models possess strong diagnostic performance while maintaining interpretability. Model interpretability is crucial as it helps engineers identify the root causes of faults and enhances the reliability of diagnostic systems. This interpretability is an indispensable component in the practical application of industrial fault diagnosis models. To address the current challenges, this paper proposes a Game Attribution-based Causal Learning Network (GA-CLN), which relies on causal relationships to establish interpretable intelligent diagnostic models that comply with physical laws. By employing game-theoretic attribution, the GA-CLN method learns causal fault features, mitigating the impact of irrelevant noise factors on diagnostic conclusions. This process preserves the causal features that remain invariant under physical laws and mathematical logic, thereby enabling generalized, interpretable diagnostics. The effectiveness of this method is demonstrated through experiments on diagnostic tasks for rotating machinery under varying speeds and loads.