Reliable and interpretable predictive maintenance in digital twin: a hybrid modelling approach powered by XAI

As industrial environments transition toward Industry 4.0, the ability to transform sensor data into actionable maintenance intelligence remains a critical managerial challenge. This paper introduces a real-time digital twin framework for predictive maintenance, leveraging a hybrid tunable LSTM architecture and physics-informed analytics. Unlike static models, our approach utilizes dynamic hyperparameter optimization to address temporal degradation and sensor noise inherent in complex industrial systems. A key contribution is the integration of SHAP-based explainability, providing maintenance managers with interpretable diagnostics that justify intervention strategies. Results from cross-industry case studies indicate a 15–30% improvement in early fault detection lead times and a 41% reduction in unplanned downtime. By providing precise remaining useful life (RUL) estimates, the framework enables a shift from reactive to strategic condition-based maintenance, offering significant cost-saving implications for large-scale manufacturing operations.