Machine Learning-Based Time-to-Failure Prediction for a Beverage Packaging Labeling Machine

Abstract

Predictive maintenance applies machine learning (ML) to forecast industrial time-to-failure (TTF), but reliable evaluation is challenged by changing operating conditions, repeated failure events, and information leakage. This study develops a leakage-aware TTF prediction framework using four years of operational and maintenance data (2021–2024) from an industrial beverage-packaging labeling machine, including 1,560 operational records and 479 unplanned failures. Six ML algorithms—Random Forest (RF), Gradient Boosting, XGBoost, Support Vector Regression, Multi-Layer Perceptron, and CatBoost—were com-pared with five historical, temporal, persistence, and linear baselines. CatBoost-Cox, using a Cox survival objective, was also evaluated for relative failure-risk discrimination. Nested walk-forward validation with event-grouped partitioning and fold-specific preprocessing, feature selection, and hyperparameter tuning was employed. All ML models outperformed baselines in RMSE, MAE, and MedAE. RF achieved the lowest mean RMSE (1.8813 days), MAE (1.0121 days), and MedAE (0.2552 days), while CatBoost achieved the highest mean R² (0.6280). However, no significant RMSE differences remained after Holm correction.  CatBoost-Cox achieved a mean concordance index of 0.9155 (SD = 0.0336). Residual, temporal, computational, and sensitivity analyses revealed model-specific trade-offs, emphasizing leakage-aware temporal validation and application-specific model selection.