Real-Time Automated Event Detection Framework for Electrical Submersible Pumps

Abstract

In artificial lift systems, electrical submersible pumps (ESPs) are often closely monitored due to the relatively high costs of unintended shutdowns and failure. ESP surveillance can range from local alarms and trips using simple logic rules, to full-fledged 24/7 surveillance centers with dedicated experts monitoring the pump and well behavior. These experts are tasked with preventing mis-operation to maximize uptime and ESP run life. Although monitoring has shown tangible results in improving overall ESP operation, a major challenge is the ability to scale up the level of surveillance performance required and maximizing the benefits over a large fleet of assets in a cost-effective manner. With the digitalization trend, the amount of sensor data is increasing, and many systems are equipped with three to five alarm states applied to each measurement. In addition, tighter bounds are required to track the desired ESP operating region, but this often results in a very large volume of notifications and threshold alarms that need to be managed as operational strategies and well conditions evolve over time. A real-time automated event detection engine framework processes ESP downhole, surface, and electrical sensor data to detect and identify potentially harmful patterns as required by the real time surveillance workflow. The engine is fully automated and can operate without any human interaction, adapting itself to each of the wells it monitors. Therefore, it does not require manual fine-tuning of parameters for each of the wells as required by the traditional multi-level threshold alarms method or other approaches such as fuzzy logic weights to capture relative importance. The engine is a plug-and-play engine that can be deployed on the edge or in the cloud. Notably, the engine is robust to noise and capable of discriminating various types of sensor failures, instrumentation noise, and harmful ESP conditions such as low-flow conditions, while reducing false or unnecessary alarms to an acceptable minimum. Engineered with the optimum balance between domain understanding and machine learning, the engine exploits complex correlations between measurements to provide a single diagnosis/output to end users, instead of multiple alarms per measurement as commonly used today. This engine has been extensively tested both offline and in real-time conditions on active wells and has dramatically improved the efficiency of existing surveillance approaches.