Time-aware predictive maintenance of electrical submersible pumps using catboost ensemble learning and trend-based labeling

Predictive maintenance of Electrical Submersible Pumps (ESPs) is critical for minimizing unplanned shutdowns and improving the reliability of oil production systems. This study explores the application of ensemble machine learning algorithms to predict the operational status of ESPs using real-world SCADA sensor data from a field containing seven active wells. A single well, selected based on its comprehensive anomaly logs and documented failure history, was analyzed over a 105-day period. During this interval, approximately 87,000 multivariate sensor data points were collected, offering a rich dataset for modeling degradation patterns and predictive maintenance. A trend-based, domain-informed labeling strategy was implemented to classify ESP behavior into three operational states: stable, unstable, and failure. Key preprocessing steps included outlier removal using Isolation Forest and dimensionality reduction via Principal Component Analysis (PCA). The processed data were then used to train and evaluate two ensemble classifiers CatBoost and ExtraTrees alongside baseline models such as Support Vector Machines (SVM) and Multilayer Perceptrons (MLP). Among the models tested, CatBoost achieved the highest classification performance with an overall accuracy of 94.2%, and a recall of 96.5% for the failure class, outperforming other models particularly in handling imbalanced data. The model demonstrated strong generalization under cross-validation and robustness to noise and outliers. The novelty of this work lies in integrating rule-based labeling, PCA, and ensemble methods into a unified, interpretable, and scalable predictive framework suitable for field-wide ESP monitoring. Compared to existing approaches, this method improves early failure detection and reduces false alarms, making it more practical for real-time deployment in oilfield SCADA environments.