Ensemble Learning for Predictive Maintenance: A Comparative Analysis of Bagging, Boosting, and Stacking Strategies

Predictive maintenance (PdM) has become a crucial element of Industry 4.0, aiming to reduce equipment downtime and operational costs through data-driven failure prediction. Traditional machine learning models, while effective, face challenges such as imbalanced class distributions, sensor noise, and complex failure patterns, which limit their reliability. This study investigates the potential of ensemble learning techniques—XGBoost, Random Forest, and Logistic Regression—to enhance predictive accuracy and robustness in industrial PdM systems. Using the Kaggle Predictive Maintenance Dataset, we implement and compare three ensemble strategies: (1) Random Forest (bagging), (2) XGBoost (boosting), and (3) a stacked ensemble combining XGBoost, Random Forest, and Logistic Regression. Our experiments reveal that stacking achieves the highest performance (F1-score: 0.92), particularly in detecting rare failure classes. Additionally, we evaluate computational efficiency, interpretability trade-offs, and the impact of hyperparameter tuning on model generalizability. The results demonstrate that ensemble methods, especially hybrid approaches, significantly outperform traditional single-model frameworks, providing actionable insights for deploying scalable PdM solutions in industrial environments.