Evaluation of Machine Learning Techniques for ESP Diagnosis Using a Synthetic Time Series Dataset

Electrical submersible pumps (ESPs) play a pivotal role in oil and gas extraction, elevating large volumes of fluids from subsurface reservoirs to the surface. The ESP is the primary method for artificially lifting oil in the Middle East to meet production demands. The reliable operation of these pumps is paramount for optimizing production and minimizing downtime. Despite the critical importance of ESPs, there has been a gap in the literature concerning the application of machine learning techniques exclusively focused on synthetic digital twin data for condition monitoring and fault diagnosis in the oil and gas sector. This study fills this gap by conducting a comprehensive evaluation of machine learning algorithms, specifically Gaussian Mixture Models (GMM) and Support Vector Machines (SVM), for ESP diagnostics. Results indicate that GMM outperforms K-means in anomaly detection, achieving an accuracy rate of 91%, while K-means reached a maximum of 49%. In multi-class event classification, SVM classifiers demonstrated high efficacy, with an F1-score reaching up to 94.35%. These methods were rigorously assessed through cross-validation, utilizing evaluation metrics such as accuracy, precision, recall, and F1-score. The study not only underscores the efficacy of GMM and SVM in ESP diagnostics but also introduces a novel approach to leveraging synthetic digital twin data for condition monitoring in the oil and gas industry.