Imbalance-robust predictive maintenance using deep learning with virtual measurement modeling and explainable AI

Abstract

Predictive maintenance (PdM) has become a key enabler of intelligent industrial systems; however, its effectiveness is often constrained by measurement uncertainty, the absence of raw sensor signals, and severe class imbalance between normal and failure events. This paper proposes a measurement-aware predictive maintenance framework that utilizes operational variables as virtual measurement proxies, enabling measurement-oriented analysis without requiring raw sensor signals. Using the AI4I 2020 Predictive Maintenance Dataset, the proposed framework integrates a hybrid imbalance-handling strategy that combines Synthetic Minority Oversampling Technique (SMOTE) and cost-sensitive learning to improve failure detection under highly imbalanced conditions. Experimental results demonstrate that the proposed hybrid framework consistently outperforms the baseline, the SMOTE-only configuration, and the cost-sensitive configuration. Among the evaluated classifiers, LightGBM achieved the best overall performance with an F1-score of 0.9043, followed by DNN (0.8708) and XGBoost (0.8680). SHAP-based explainability analysis identified tool wear and torque as the most influential predictors of machine failure. At the same time, robustness experiments with simulated Gaussian measurement noise demonstrated stable performance across low and moderate noise levels. An ablation study further confirmed the complementary contributions of SMOTE and cost-sensitive learning. Overall, the proposed framework provides an effective and interpretable approach for predictive maintenance on the AI4I 2020 benchmark dataset, while further validation using real industrial datasets is required before practical deployment.