Machine Learning-Driven Failure Detection for Proactive Maintenance of Electric Submersible Pumps (ESPs)

Electric Submersible Pumps (ESPs) are critical in oilfield production, yet their exposure to harsh subsurface environments often leads to unexpected mechanical and electrical failures. Such failures cause costly downtime and reactive maintenance. Conventional strategies—whether scheduled or corrective—lack foresight to predict these events.
This thesis presents a machine learning–driven framework for predictive maintenance of ESPs using Long Short-Term Memory (LSTM) neural networks. The model is trained on multivariate time-series sensor data—temperature, pressure, vibration, and electrical load—to capture short- and long-term operational trends. Alongside LSTM, Transformer and TimeBERT models were explored to benchmark performance across modern sequence-learning approaches.
The proposed LSTM achieved 99.95% test accuracy and an F1-score of 0.8480, balancing precision and recall despite severe class imbalance (failure prevalence <2%). SHAP (SHapley Additive exPlanations) analysis was employed to highlight the most influential predictors of ESP failure, supporting domain validation and operational trust.
Compared to prior benchmarks—such as a PCA + XGBoost model with an F1 score of ~0.686—the LSTM-based approach delivers a scalable, proactive solution for failure detection, enabling timely interventions and extending equipment life in sensitive, high-risk oilfields.

Supervisor: Prof. Fawaz S. Al-Anzi

Co-supervisor: Prof. Ayed Salman

Convener: Prof. Mehmet Karaata

Examination Committee: Dr. Abdullah Alshaibani