Abstract
The integration of Explainable Artificial Intelligence (XAI) into predictive maintenance (PdM) systems is critical for fostering trust, transparency, and actionable insights in smart manufacturing environments. Despite the proliferation of deep learning models achieving high accuracy in failure prediction, their black-box nature limits adoption by human operators and maintenance engineers. This study proposes a comprehensive framework that combines XAI techniques—specifically SHAP, LIME, and attention-based mechanisms—with deep learning architectures (LSTM and CNN) for PdM in industrial machinery. Using a real-world dataset from a semiconductor manufacturing process, we evaluate model performance and interpretability across multiple metrics, including prediction accuracy, F1-score, explanation fidelity, and user trust. Our results demonstrate that the hybrid XAI-PdM model achieves a precision of 0.94 and recall of 0.91, outperforming baseline models. Furthermore, explanation quality, measured by faithfulness and completeness, shows significant improvement over opaque models. A user study with 30 maintenance engineers indicates that XAI-enhanced predictions increase trust and decision-making speed by 35% compared to black-box models. The findings underscore the necessity of embedding interpretability into PdM systems to enable effective human-AI collaboration. This research contributes to the growing body of literature on transparent AI in Industry 4.0 and provides a practical roadmap for deploying XAI in real-world manufacturing settings.
Keywords
Explainable AI, Predictive Maintenance, Smart Manufacturing, Deep Learning, Industry 4.0, Human-AI Collaboration, SHAP, LIME