A hybrid predictive maintenance framework using machine learning for RAM study in mining assets

Purpose The purpose of this paper is to develop and validate a hybrid predictive maintenance framework that integrates machine learning–based failure prediction with reliability, availability, and maintainability (RAM) analysis to support proactive maintenance decision-making for asset-intensive mining equipment. Design/methodology/approach The study proposes a Hybrid RAM–predictive maintenance framework hybridRAM–PdM framework that combines short-horizon failure risk classification with quantile-based remaining useful life (RUL) estimation. Supervised tree-ensemble models are employed to predict failure risk at 24, 72, and 168 h, while quantile boosting is used to estimate RUL intervals. Model training is conducted strictly in chronological order to preserve operational realism. Predictive outputs are systematically mapped to RAM indicators, including mean time between failures mean time to repair and asset availability. The framework is validated using three months of real-world Computerized Maintenance Management System data from mining mobile equipment, including dump trucks and load–haul–dump machines. Findings Results indicate that the proposed framework provides useful near-term failure-risk warnings, actionable RUL intervals, and RAM-oriented readiness indicators for maintenance planning. The integration of predictive outputs with observed repair information supports asset-level prioritization and targeted maintenance interventions. However, because the case study is based on a three-month observation window and a limited number of assets, the results should be interpreted as preliminary operational evidence rather than definitive proof of long-term reliability or availability improvement. Research limitations/implications The study is based on a single industrial case study, five assets, and a three-month historical data window containing 25 observed failure events. Future research should evaluate the framework over longer operating periods, larger fleets, additional asset classes, and different industrial sectors. Further work should also incorporate cost-aware threshold optimization and formal validation of the relationship between ML-derived planning indicators and classical RAM quantities. Practical implications The framework enables maintenance practitioners to translate machine-learning predictions into intuitive alerts, RUL-based planning windows, and RAM-oriented readiness indicators. These outputs can support proactive inspection planning, maintenance scheduling, and repair-prioritization decisions, while final deployment should be calibrated using site-specific cost, safety, and production constraints. Originality/value This paper provides a novel contribution by explicitly linking predictive maintenance analytics with RAM performance measures, bridging the gap between data-driven prediction models and practical maintenance engineering decision-making.