Cost-Sensitive and Explainable Machine Learning for Predictive Maintenance Decision Support in Industry 4.0

Abstract

Unplanned equipment failure remains one of the largest controllable cost drivers in discrete manufacturing, and the sensor infrastructure of Industry 4.0 has made data-driven predictive maintenance a practical alternative to reactive and calendar-based policies. Published benchmarks on this task, however, are dominated by threshold-free accuracy reporting: models are compared on metrics that are optimistic under severe class imbalance, synthetic resampling is applied as a default remedy, the predicted probability is converted into a maintenance action at an arbitrary cut-off of 0.50, and the resulting model is rarely interrogated for the reasoning behind its alarms. The purpose of this study is to develop and validate an integrated decision-support framework that connects classifier selection, imbalance handling, probability calibration, decision-theoretic thresholding and post-hoc explanation within a single reproducible pipeline. The framework is evaluated on the AI4I 2020 predictive maintenance benchmark, which contains 10,000 machining observations with a 3.39% failure rate. Four physics-informed descriptors are derived from the raw signals, eleven classifiers are screened under ten-fold stratified cross-validation with the area under the precision-recall curve as the primary criterion, seven resampling schemes are compared against an unresampled baseline, and differences are tested with the Friedman test followed by Holm-corrected Wilcoxon signed-rank comparisons. Gradient boosting achieves the best cross-validated performance (PR-AUC 0.9145) and a hold-out PR-AUC of 0.8987 with a Matthews correlation coefficient of 0.888; it is statistically indistinguishable from random forest and LightGBM but superior to the remaining eight learners. None of the seven resampling schemes improves ranking quality, and every one of them degrades precision substantially. The operational gain instead comes from the decision layer: isotonic calibration reduces the Brier score from 0.00726 to 0.00654, and combining calibrated probabilities with the cost-minimising threshold lowers expected misclassification cost by 37.2% relative to the default cut-off and by 87.0% relative to a run-to-failure policy. Shapley attributions identify tool wear, rotational speed, the process-air temperature difference and torque as the dominant drivers, reproducing the documented physical failure mechanisms of the benchmark and supporting engineer-facing justification of individual alarms.