<![CDATA[Predictive maintenance models are often evaluated primarily as classifiers, although practical maintenance decisions depend on the relative consequences of preventive interventions and missed failures. This study develops a reliability-oriented predictive maintenance framework that integrates machine-learning failure probabilities with cost-sensitive maintenance thresholds. A reproducible simulation of 80,000 industrial-equipment cycles was constructed from the documented process envelopes and failure mechanisms of the AI4I 2020 benchmark, while the original UCI dataset was used only as a calibration reference. Six raw operating variables were supplemented with temperature difference, mechanical power, and a type-adjusted wear–torque index. Four classifiers—logistic regression, random forest, extra trees, and XGBoost—were evaluated using stratified training, validation, and independent test partitions. Random forest achieved the strongest test PR-AUC of 0.967, ROC-AUC of 0.989, F1 of 0.965, and MCC of 0.964. Incorporating physics-derived variables increased PR-AUC from 0.943 to 0.966 and reduced the Brier score from 0.0088 to 0.0035. Cost-sensitive thresholds reduced normalized maintenance costs by 46.3–88.9% relative to reactive maintenance across failure-to-preventive-maintenance cost ratios of 2:1 to 20:1. Moderate sensor perturbations preserved predictive performance, whereas 10% standard-deviation noise substantially reduced PR-AUC and recall. These findings demonstrate that predictive performance, engineering-informed feature design, sensor robustness, and maintenance economics should be evaluated jointly before industrial deployment. ]]>
Copyrights © 2026
