Application of Data-Driven Models for Real-Time Monitoring, ESP Pump Failure Prediction, an‎d Remaining Useful Life Estimation

12/9/2026 12:00:00 AM

Electric Submersible Pumps (ESPs), as one of the most widely used artificial lift methods in the oil an‎d gas industry, play a critical role in sustaining hydrocarbon production. However, due to harsh downhole operating conditions an‎d continuous exposure to pressure fluctuations, high temperatures, an‎d mechanical vibrations, these systems are highly susceptible to premature failures. Unplanned shutdowns caused by ESP failures not only increase maintenance an‎d workover costs but also lead to significant production losses an‎d operational disruptions. Therefore, the development of advanced real-time monitoring an‎d Remaining Useful Life (RUL) prediction methods is of paramount importance. In this study, a data-driven framework based on deep learning was developed using operational data collected from condition monitoring an‎d SCADA systems. The proposed framework is capable of learning normal operational patterns an‎d detecting deviations from healthy behavior. In the first stage, raw data were preprocessed through noise filtering, outlier detection, normalization, an‎d feature engineering to ensure data quality an‎d model robustness. Subsequently, two deep learning architectures—LSTM Autoencoder (LSTM-AE) an‎d Transformer Autoencoder (Transformer-AE)—were employed to reconstruct normal pump behavior an‎d identify anomalies. Reconstruction error was defined as the primary failure indicator. Data-driven fixed thresholds, derived from the statistical distribution of healthy-state data, were implemented to prevent error masking an‎d enable accurate classification of three operational states: normal, warning, an‎d failure. In addition, adaptive thresholds were utilized as a complementary analytical tool to capture sudden shocks an‎d localized deviations during visual inspections. To enhance prognostic capability, reconstruction error trend analysis combined with a multi-window approach was applied for Remaining Useful Life estimation, providing a more reliable prediction of time-to-failure. Quantitative eva‎luation demonstrated that the Transformer-AE model, with its cause-oriented detection capability, significantly outperformed the LSTM-AE model. In a rapid degradation scenario, Transformer-AE predicted an RUL of 4.7 days, whereas the LSTM model produced an unrealistic estimate of 102 days. In the sudden failure scenario, the Transformer-based model generated a higher risk score (58.9 versus 25.4), enabling more accurate identification of the system’s critical condition. The results indicate that integrating these data-driven approaches substantially improves early fault detection accuracy an‎d enhances the reliability of RUL estimation. This achievement contributes to reducing unplanned downtime, lowering maintenance costs, an‎d improving operational reliability an‎d production efficiency in oil an‎d gas systems. The proposed framework can serve as a foundation for intelligent asset management systems an‎d proactive decision-making strategies in upstream petroleum operations.

مدل‌هاي داده‌محور , مانيتورينگ بلادرنگ , پمپ‌هاي شناور الكتريكي , پيش‌بيني خرابي , عمر باقي‌مانده

Data-driven models , Real-time monitoring , Electric Submersible Pumps (ESP) , Failure Prediction , Remaining Useful Life (RUL)

Farzad Rezaei Ghotbabadi

Dr. Hamid Reza Jahanqiri /Dr. Seyed Mojtaba Hosseini Nasab