Electrical Submersible Pump Failure Classification Using Machine Learning Models with Advanced Data Splitting Techniques for Field Application

Abstract

Electrical Submersible Pumps (ESPs) drive approximately 60% of global artificial lift production; however, their continuous exposure to harsh downhole environments can cause mechanical and electrical degradation. This unpredictability in run-life directly causes unplanned well downtime, deferred production, and high-cost workover interventions. Predicting specific ESP failure modes, rather than relying on binary failure-survival estimations, is critical for optimizing well performance. This study presents a rigorous, reproducible machine learning pipeline designed to proactively classify ten distinct ESP failure mechanisms using readily available well and production data. The dataset comprises a decade (2012–2023) of operational history from over 350 wells in Egypt's Western Desert. A strict data preprocessing framework—enforcing physical engineering limits and domain-specific imputation—transformed 2.147 million raw records into 933,535 validated, daily time-series observations across 37 input model features. To ensure real-world applicability and prevent data leakage, the methodology utilized advanced validation frameworks, including time-based sequential, well-based (group-aware), and rolling-window splitting techniques. Severe operational class imbalances (ratios up to 171:1) were mitigated using a weighted Stratified Synthetic Minority Over-Sampling Technique (SMOTE). Ten distinct algorithms, ranging from tree-based ensembles to deep neural networks, were benchmarked. The results demonstrate that using a One-vs-All (OvA) approach with a 1-Dimensional Convolutional Neural Network (1D-CNN) is most effective at isolating minority failure entries. The 1D-CNN OvA architecture achieved 94.85% diagnostic accuracy and a 0.946 Macro F1-Score, significantly outperforming standard multiclass deep learning and advanced ensembles like XGBoost (89.40%). To validate its real-world applicability, the model was evaluated in a strictly blind field test on a historically problematic well, Well KX. The algorithm correctly classified five consecutive ESP failures, accurately identifying distinct mechanical and electrical failure modes, including pump sticking, shaft breakage, motor short circuit, and cable short circuit. By pinpointing the specific impending failure mode, the framework enables operators to pre-schedule rigs and optimize replacement ESP design, fundamentally eliminating reactive unplanned downtime per event and extended workover delays.