Explainable machine learning for failure prediction in industrial assets: a cost-sensitive and reliability-aware decision support framework

Industrial failure-prediction models are often evaluated as classifiers, although maintenance decisions also require interpretable and cost-aware actions. This article presents a reliability-aware explainable machine-learning framework for predictive maintenance, RA-XPdM. Its contribution is the integration of established prediction, calibration, SHAP attribution and cost-sensitive thresholding methods with a maintenance-priority heuristic and comparative risk benchmarks. Evaluation uses an AI4I-aligned controlled simulation and a second synthetic benchmark, the public AI4I 2020 dataset; neither provides plant field validation. The reported public-benchmark test cost is 0.0665 normalized cost units per observation at the validation-selected threshold, compared with 0.4080 for a no-action policy, under illustrative costs. These are model-based decision losses, not measured industrial savings. Explanations describe an uncalibrated gradient-boosting model, and their reported stability is conditional on a fixed fitted model. The findings illustrate the feasibility of a transparent decision-support workflow on synthetic snapshot data. They do not establish advance failure prediction, explanation stability under model refitting or effectiveness in operational maintenance.