Uncertainty-Aware Predictive Maintenance Scheduling: A Decision-Support Framework for Industrial Production Systems

Abstract

Preventive maintenance in a manufacturing plant is, at bottom, a resource-allocation problem: every intervention weighs the cost of replacing a component too early against the risk of an unplanned stoppage under the challenging realities of production calendars, crew availability, and machine access. Most facilities still rely on fixed calendar intervals that ignore what the equipment is signaling, paying for the difference in wasted component life and avoidable downtime. This paper describes a decision-support framework that connects data-driven prognostics to a production-constrained optimization model. A Gradient-Boosted Regression Tree (GBRT) supplies point estimates of Remaining Useful Life (RUL). A Quantile Regression Forest (QRF) turns these into calibrated prediction intervals with finite-sample coverage guarantees, and a Mixed-Integer Linear Program (MILP) converts the resulting per-machine failure probabilities into maintenance schedules that respect shift windows and crew capacity. The contribution lies less in the individual components than in how they are joined: uncertainty flows directly into the scheduler, so every intervention decision reflects both predictive risk and production feasibility. We evaluated the framework on a public milling benchmark and eight months of data from a twelve-machine packaging facility (ten independent runs each). Relative to the calendar-based policy currently used at the facility, total maintenance cost fell by 6.3% on the packaging line and 11.9% on the NASA milling dataset; unplanned downtime events fell by 7.2% and 44.4%; and unnecessary preventive replacements fell by 26.2% and 66.7%. Calibration error remained below 1.5 percentage points at every coverage level tested, and all scheduling instances were solved to certified optimality within seconds for fleets of up to 50 machines on standard hardware.