Stacking ensemble-based predictive maintenance framework on AI4I datasets in IIoT systems for forecasting machine failures

Purpose The current study proposes building a strong predictive maintenance (PdM) model for industrial systems through the use of ensemble machine learning models to forecast equipment failures before they happen. The research emphasizes enhancing the reliability during operation, minimizing downtime and offering a scalable solution to fill the gap between theoretical studies and real-world industrial applications in the Industry 5.0 era. Design/methodology/approach Two benchmark datasets: AI4I 2020 and its extended variant AI4I Predictive Maintenance Dataset with Irregularities (PMDI) are employed to verify the proposed approach. The research undertakes thorough data pre-processing, such as label encoding, feature scaling and Synthetic Minority Over-sampling Technique for handling class imbalance and data quality problems. Three heavy-hitting gradient boosting algorithms: XGBoost, LightGBM and CatBoost are employed as base learners in a stacking ensemble, with logistic regression as the meta-learner. The performance of the model is tested rigorously by Accuracy, Precision, Recall, F1-score and Receiver Operating Characteristic-Area Under the Curve (ROC-AUC). Findings Experimental results indicate that stacking ensemble outperforms standalone base models consistently and attains 99.16% accuracy for the AI4I 2020 dataset and 99.60% accuracy for the AI4I PMDI dataset with ROC-AUC values close to 0.99. These results validate the success of ensemble learning, especially stacking, in achieving high accuracy and reliability for industrial PdM applications. Research limitations/implications The present study is restricted to publicly available benchmark datasets and artificial industrial settings. Real-time deployment, integration with deep models like CNN, BiLSTM or Transformers, and addition of cost-sensitive measures or explainability tools like SHAP and LIME for increased interpretability and usability are potential future extensions. Originality/value This research combines three gradient boosting models into a stacked ensemble framework designed specifically for PdM in an innovative way. It adds a reproducible, understandable and highly performant PdM model with practical industrial benefit, research relevance and social impact that supports Industry 5.0-compatible and sustainable data-based maintenance practices.