Risk of Failure and Remaining Useful Life Prediction for Electrical Submersible Pumps

Predicting or anticipating faults and catastrophic failure of electrical submersible pumps (ESPs) is challenging because of limited measurements, complex failure modes, and dynamic conditions. Existing solutions in literature are limited in terms of performance, robustness, data requirements, operationalizability, etc. Additionally, there is no universal baseline human expert solution. We present a plug-and-play machine-learning (ML) and domain-driven prognostics health management (PHM) engine that is operationalizable and scalable, works with existing (typical) surveillance data, and provides explainable and actionable insights. The novel solution consists of a features engine that uses ML, ESP domain, physics, and signal processing to extract patterns and events at different time scales from multiple raw measurements. Several detectors contextualize the pump condition from different perspectives and time frames. The PHM engine outputs continuous predictions of risk of failure (RoF) and remaining useful life (RUL) at different confidence levels. An alarm configurator is used for prioritization and reporting analytics. The engine also contains a dedicated data quality module that identifies data issues and gauge failures and automatically adjusts internal dependencies to produce meaningful outputs. A baseline PHM engine was trained and tested on diverse groups of pumps from multiple locations around the globe. Without targeted fine-tuning, 80% of pumps in various test datasets were observed to fail during high-RoF periods predicted by the baseline PHM engine; this percentage is higher when the engine is automatically tuned for a meaningful population of wells based on criteria such as location and reservoir. Missed failure detections were often due to poor quality data or sudden catastrophic failures lacking any leading indicators in the available measurements. Specific case studies highlight the engine's comprehensive outputs and coverage. Explainable AI helps surveillance operations pinpoint the reasons for high RoF and take proactive actions where possible. For operations, this can be used to extend run life, adjust control methods or set points, and strategize interventions, among others. At the population level, the engine can be used for workover planning, inventory management, operations assessment, failure analytics, design improvements, and other business intelligence. The baseline PHM engine also helps determine the optimal data needed to capture different types of failures. The impact of availability of certain measurements and the time resolution has also been systematically studied. A baseline PHM engine for ESPs is deployable on the edge or in the cloud and offers different levels of automated tuning for adaptability and continuous improvement. The engine provides consistent and actionable insights with end-to-end explainability and automatically adjusts confidence based on data quality and availability. The modular architecture makes it easy to add and disable various components to facilitate upgrades (such as new sensor technology) and customized solutions for different ESP populations.