A Comparative Evaluation of Interpolation and Generative Oversampling Techniques for Predictive Maintenance

Predictive maintenance (PdM) enhances industrial operational efficiency by facilitating timely detection of equipment failures using machine learning models developed from historical maintenance data.  Real-world industrial datasets frequently exhibit significant class imbalance, as failures are infrequent occurrences. This imbalance substantially diminishes predictive accuracy for the minority class (failures).  This study systematically evaluates three data augmentation techniques—Synthetic Minority Oversampling Technique (SMOTE), SMOTETomek, and Conditional Tabular Generative Adversarial Networks (CTGAN)—to address this challenge, utilising the AI4I 2020 Predictive Maintenance dataset.  A Random Forest classifier was trained on augmented data, with a comparison of augmentation methods conducted through various performance metrics, including precision, recall, F1-score, ROC-AUC, and PR-AUC.  The findings indicate that both SMOTE and SMOTETomek significantly enhance failure detection performance, with F1-scores and recall rates surpassing 0.99.  In contrast, CTGAN demonstrates marginally lower classification performance (F1-score ≈ 0.88) while effectively generating realistic synthetic samples that maintain the original data distributions and inter-variable relationships.  These results underscore the trade-offs between oversampling methods and generative models: SMOTE-based approaches optimise raw predictive accuracy for rare failures, whereas CTGAN demonstrates significant potential for improving model generalisation in complex industrial applications.