Journal Paper

The integration of Artificial Intelligence (AI) and Machine Learning (ML) into industrial environments, particularly for optimising fault detection and diagnosis, has accelerated with Industry 4.0 and 5.0. However, the “black-box” nature of these methods hinders practical implementation, as trust, interpretability, and explainability are crucial for informed decision-making. Furthermore, impending regulatory frameworks like the EU AI Act make directly implementing opaque AI for critical industrial tasks infeasible. Explainable AI (XAI) offers a promising solution by enhancing ML model interpretability and auditability through human-understandable explanations. This review comprehensively analyses recent XAI advancements for industrial fault detection and diagnosis, presenting a novel taxonomy for XAI methods and discussing how XAI outputs are generated, conveyed to end-users, and evaluated. It then systematically reviews real-world industrial XAI implementations, highlighting their applications, methods, and output presentation approaches. Key identified trends include the dominance of post-hoc feature attribution methods, widespread use of SHAP and GradCAM, and a strong reliance on graphical explanation tools. Finally, it identifies current challenges and outlines future research directions to promote the development of interpretable, trustworthy, and auditable AI systems in industrial settings.