Digital Twin-Driven Predictive Maintenance in Smart Manufacturing: Integrating Machine Learning for Intelligent Mechanical System Optimization

The significance of predictive maintenance in the context of smart manufacturing is rooted in the fact that unexpected machine downtimes impact productivity, maintenance planning, and asset utilization. This research proposes an explainable machine-learning approach to use operational measurements as a data-driven virtual condition layer to support operational decisions for digital-twin-based maintenance. The AI4I 2020 Predictive Maintenance dataset with 10 thousand observations was analysed with 6 non-leaky machine-condition predictors. Precision-recall area under the curve (PR-AUC) was used as the main evaluation metric to compare the models optimized on the development data set and tested on an independent validation data set: Logistic Regression, Random Forest, Extra Trees and HistGradientBoosting. HistGradientBoosting was chosen and the locked-test PR-AUC, ROC-AUC, balanced accuracy and failure recall were 0.8242, 0.9750, 0.9031, and 0.8235 respectively. The most important predictors identified by the permutation importance were torque, air temperature, and tool wear. One layer focused on validation, that showed maintenance risk, brought together 48 out of 51 failures for the locked test in the Critical band. The framework shows how explainable machine learning can be used to assist in failure-risk prioritization, condition-based maintenance and intelligent mechanical-system optimisation without losing the human element or ‘leaking' the target.