Machine Learning–Driven Predictive Maintenance for Smart Beverage Manufacturing Systems

Abstract

This study proposed a machine learning-based predictive maintenance system in a beverage manufacturing process to enhance failure prediction accuracy and support improved maintenance planning. The approach focused on the beverage filling line, captured short-term temporal process dynamics using rolling-window feature extraction, and addressed the strong class imbalance between normal operation and breakdown events. Significant sensor variables were initially identified through correlation analysis, and the selected sensors and breakdown criteria were subsequently analyzed using various classification models. A 60-minute decision horizon was defined as a classification-based early-warning window derived from the time-to-failure (TTF) variable. Observations were labeled breakdown-proximate state when the next breakdown occurred within 60 minutes (TTF < 60 min) and as normal operation when the next breakdown occurred after 60 minutes (TTF > 60 min). Therefore, the model classifies a breakdown-proximate state rather than predicting the exact remaining time to failure. Model performance was evaluated using precision, recall, F1-score, ROC-AUC, and chronological walk-forward validation. Based on correlation analysis and process relevance, 15 sensors were selected as key features associated with process breakdown. The Random Forest model was the best performer, achieving an accuracy of 0.990, precision of 0.918, recall of 0.722, F1-score of 0.808, and ROC-AUC of 0.995 at the optimized threshold of 0.45. The recall result indicates acceptable detection of breakdown-proximate states, although 27.8% was missed as a false negative. Considering the balance between recall and precision, the precision exceeded 90%, showing that false alarms were maintained at a low level. These results demonstrate a practical trade-off between reducing missed breakdown warnings and avoiding excessive unnecessary inspections. Thus, the Random Forest is suitable for predicting breakdowns with accuracy, stability and robustness. Future work will extend the predictive maintenance results linked with failure-risk predictions to reliability metrics of time between failures (TBF) and mean time between failures (MTBF), improved alignment between maintenance actions and actual machine conditions.