An Innovative Machine Learning (ML)-Based Approach for ESP Predictive Monitoring Using Anomaly Detection and Survival Models

Abstract

Electrical Submersible Pumps (ESPs) are commonly used in high-flow wells for their compact design and reliability, but their replacement requires costly well interventions. Early failure prediction is thus crucial in reducing downtime and improving maintenance strategies. This study introduces a cutting-edge machine learning (ML)-based approach that monitors ESP performance and predicts failure risk, survival probability, and remaining useful life using real-time sensor data and ESP metadata, enabling proactive decision-making and operational efficiency. The process begins with sensor data cleanup and pre-processing, followed by ML model training to predict key operational parameters such as motor temperature and delta pressure. The predicted values are compared against actual sensor readings, and deviations beyond a predefined threshold trigger alarms. To enhance accuracy, anomalies from transient startup/shutdown events are filtered out. A time-dependent survival model then estimates ESP failure probability over time, incorporating factors such as pump vendor, field and reservoir properties, startup/shutdown frequencies, downhole sensor failures, and detected anomalies from anomaly detection models. Results are visualized on an interactive dashboard for real-time monitoring and decision support. The anomaly detection and survival models developed in this study provide an ML-based framework for real-time ESP monitoring and failure risk prediction. The predicted failure risk is visualized as dynamic survival curves, which update continuously in response to new anomalies or changes in operating conditions or sensor status. This enables operators to proactively identify wells with frequent alarms, prioritize ESP interventions based on predicted failure risks, and reduce unplanned shutdowns. The interactive dashboard presents a ranked list of ESPs predicted failure risk, alongside relevant sensor data trends. Two key features enhance usability, ChatESP, a GenAI-powered assistant that provides instant access to well history, sensor trends, and predictive insights, and a threaded discussion feature that enables users to log and track operational comments specific to ESPs and relevant time periods. This integrated, data-driven approach enhances ESP reliability, streamlines intervention scheduling, and improves well performance, ultimately reducing operational risks, production deferment and maintenance costs, empowering operators to make informed, proactive decisions using real-time field data. This approach marks a shift from reactive monitoring to predictive, context-aware decision support by combining anomaly detection and survival analysis models into a unified, risk-ranked view of ESP performance. Moreover, the inclusion of human-in-the-loop features, such as comment threads for operational insights and a GenAI assistant for instant guidance, facilitates continuous learning and collaboration. Consequently, this bridges the gap between data science and field operations, enabling smarter, timely, and more actionable decisions for ESP management.