Application of Machine Learning in Predictive Maintenance Scheduling: An Industrial Case Study

Predictive maintenance (PdM) aims to anticipate equipment failures, thereby reducing unplanned downtime and maintenance costs. This paper presents an industrial case study on applying machine learning (ML) for PdM-driven maintenance scheduling in a high-throughput packaging facility. A stacked time-series pipeline (feature learning via LSTM and gradient boosting, plus conformal uncertainty) predicts remaining useful life (RUL) and short-horizon failure risk, which are then embedded in a rolling-horizon mixed-integer program (MIP) to co-optimize work-order timing, capacity, and production impacts. Against reactive and time-based baselines, the proposed approach reduced unplanned downtime by ۳۱.۴%, increased MTBF by ۲۲.۷%, and cut maintenance overtime by ۱۸.۹% over ۱۶ weeks, at comparable spare-parts consumption. We discuss model/optimizer interplay, uncertainty handling, and transferability, and situate findings in the ۲۰۱۹–۲۰۲۵ literature, highlighting a persistent gap: robust, data-efficient integration of probabilistic RUL with capacity-constrained multi-asset scheduling under real plant calendars. (Related surveys and recent RL/MILP advances support this approach and the identification of the gap.)Predictive maintenance (PdM) aims to anticipate equipment failures, thereby reducing unplanned downtime and maintenance costs. This paper presents an industrial case study on applying machine learning (ML) for PdM-driven maintenance scheduling in a high-throughput packaging facility. A stacked time-series pipeline (feature learning via LSTM and gradient boosting, plus conformal uncertainty) predicts remaining useful life (RUL) and short-horizon failure risk, which are then embedded in a rolling-horizon mixed-integer program (MIP) to co-optimize work-order timing, capacity, and production impacts. Against reactive and time-based baselines, the proposed approach reduced unplanned downtime by ۳۱.۴%, increased MTBF by ۲۲.۷%, and cut maintenance overtime by ۱۸.۹% over ۱۶ weeks, at comparable spare-parts consumption. We discuss model/optimizer interplay, uncertainty handling, and transferability, and situate findings in the ۲۰۱۹–۲۰۲۵ literature, highlighting a persistent gap: robust, data-efficient integration of probabilistic RUL with capacity-constrained multi-asset scheduling under real plant calendars. (Related surveys and recent RL/MILP advances support this approach and the identification of the gap.)

کلیدواژه ها:

Predictive Maintenance ، Machine Learning ، Remaining Useful Life ، Maintenance Scheduling ، Industrial Case Study