Abstract
Background: Predictive maintenance (PdM) in smart manufacturing leverages machine learning to forecast equipment failures, but the lack of interpretability hinders adoption. Explainable AI (XAI) techniques address this by providing insights into model decisions. Objective: This paper proposes a hybrid XAI framework combining SHAP and LIME for PdM in industrial machinery, aiming to enhance trust and operational efficiency. Methods: Using a real-world dataset from a semiconductor manufacturing plant, we trained a gradient boosting model for failure prediction. Post-hoc explanations were generated via SHAP for global feature importance and LIME for local instance explanations. The framework was evaluated on accuracy, fidelity, and user satisfaction through a survey of 15 maintenance engineers. Results: The model achieved 96% F1-score. SHAP consistently identified temperature and vibration as top predictors, while LIME provided coherent local explanations. User survey indicated a 40% improvement in trust and a 30% reduction in diagnostic time. Discussion: The hybrid approach balances global and local interpretability, outperforming single-method baselines. Limitations include computational overhead and sensitivity to hyperparameters. Conclusion: Integrating XAI into PdM systems enhances transparency and operational efficiency, with potential for broader industrial adoption. Future work should explore real-time explanations and integration with digital twins.