Journal of Industrial Engineering and Management Studies

Abstract Abstract
Unplanned pump failures inflict high operational costs, rendering traditional maintenance strategies ‎inefficient. While Machine Learning (ML) has advanced fault diagnosis, a critical gap remains in ‎simultaneously classifying operational states and predicting the "Time Remaining until a State Change" ‎‎(TRSC) using real-world, imbalanced data. This study addresses this necessity by developing an integrated ‎predictive maintenance framework for industrial centrifugal pumps. Leveraging a four-year vibration ‎dataset (2020–2024), we employ Random Forest (RF), XGBoost, and Multi-Layer Perceptron (MLP), ‎utilizing SMOTE and jittering augmentation to mitigate data scarcity and imbalance. The study makes two ‎primary contributions: (1) accurate classification of four operational states (Normal to Failure), and (2) ‎precise regression of TRSC to optimize spare parts logistics. Results indicate that ensemble models ‎‎(RF/XGBoost) achieve classification accuracies exceeding 92% and TRSC prediction with an RMSE ‎below 25 days, significantly outperforming MLP. Furthermore, SHAP analysis reveals that horizontal and ‎axial vibrations are the dominant precursors to failure. These findings offer a robust, interpretable tool for ‎shifting towards condition-based maintenance, ensuring reliability and cost efficiency.‎