Explainable artificial intelligence in fault detection and diagnosis: a review of methods, applications, and implementation challenges

Industry 4.0 increasingly relies on AI methods for fault detection and diagnosis (FDD). However, advanced machine learning models lack transparency, reducing trust in safety-critical settings. This review examines eXplainable AI (XAI) methods adapted for industrial FDD. It also proposes a taxonomy spanning model-agnostic methods, model-specific approaches, and hybrid rule-based schemes. For each category, the paper explains how the methods reveal fault-related decision logic and examine their impact on diagnostic accuracy. The analysis shows that SHAP and feature-importance methods are the most widely used in FDD applications. Other methods (e.g., LIME) have seen limited adoption, partly due to scalability concerns. This study further examines limitations including high computational cost, restricted real-time performance, and scalability constraints. The findings indicate that although model-specific methods enhance interpretability, they continue to face challenges in scalability. The study also outlines key research questions related to evaluating explanation quality, integrating XAI into real-time FDD systems.