A machine learning-driven approach to optimizing maintenance scheduling for industrial machinery

Abstract

This study develops an integrated framework combining supervised machine learning classification with mathematical optimization to predict equipment failures and minimise maintenance costs under prediction uncertainty. Six supervised classification algorithms are evaluated for fault prediction using sensor and operational data. Type I and Type II error rates derived from confusion matrices are incorporated directly as parameters within a Mixed Integer Linear Programming (MILP) objective function. The model minimises total maintenance costs while satisfying operational scheduling constraints. Sobol global sensitivity analysis and Monte Carlo simulation with 12,000 iterations assess parameter influence and model robustness under real-world cost variability. Random Forest was selected as the optimal model based on its superior cross-validation stability and balanced error rate profile. The framework produced a cost-effective schedule with zero emergency maintenance activities, confirming that embedding prediction uncertainty into the scheduling objective eliminates reactive interventions. Sobol analysis identified false negative cost as the most influential parameter. Monte Carlo risk measures provide quantified cost boundaries for maintenance budget planning. The framework provides evidence-based scheduling decisions and quantified cost boundaries for budget planning, to support the transition from reactive to proactive maintenance strategies. This study directly incorporates classification error rates into a cost-based mixed-integer linear programming (MILP) framework, ensuring prediction uncertainty propagates into scheduling decisions and bridging predictive analytics with operational planning.