AI-driven predictive maintenance framework for public-service fleets: a cost-sensitive explainable prioritization and optimization approach

Purpose This article aims to develop a business-centred, cost-sensitive and explainable predictive maintenance decision framework for public-service fleets with limited maintenance capacity. Design/methodology/approach The prediction layer is validated on the public AI4I 2020 benchmark using discrimination, calibration and SHAP diagnostics, and is then embedded in a synthetic civil-defense fleet decision model using AFMPS, binary capacity constraints, expected-loss benchmarking, asset-level holdout, Monte Carlo uncertainty and sensitivity analysis of assumptions. Findings Gradient boosting produced calibrated benchmark risk estimates, and the AFMPS policy reduced expected failures, downtime and cost compared with reactive, interval, threshold and expected-loss policies in the synthetic fleet case. The results show how AI-generated risk, RUL urgency and critical alarms can shift maintenance from routine time-based cycles toward targeted, capacity-aware interventions. Practical implications The framework gives maintenance managers a reproducible procedure for ranking vehicles, allocating scarce workshop capacity, incorporating spare-parts cost exposure and monitoring critical-service availability. Originality/value The article links explainable failure-risk prediction, calibration, mission criticality, business-centred cost logic, expected-loss comparison and capacity-constrained optimization in a public-service fleet context.