Using Machine Learning-Based Predictive Models to Enable Preventative Maintenance and Prevent ESP Downtime

Abstract

Abstract This paper focuses on the use of artificial intelligence (AI) and machine learning (ML) algorithms to implement anomaly detection and shows how this concept can be extended to implement autonomous well surveillance. Today, critical equipment is monitored by implementing automation and control systems with built-in protection logic for safe operation of equipment and for shutdown of the system in the event that operation deviates outside of valid process conditions. These automation and control systems require constant surveillance by a human operator to verify that all processes are running normally. It is the human operator's responsibility to react to any alarm conditions that occur during operation. Often, the alarm trigger event occurs without early notification and the operator has a short period of time to react. This way of controlling operations requires skilled operators with a great deal of experience to monitor and control the system in an effective manner. It also limits the amount of time the operator can allocate toward optimization. Autonomous surveillance is the concept of training an AI system to provide early detection of abnormal behavior. In this way, the system can take over the task of constant surveillance of process operation, leaving the experienced human operator with additional capacity to focus his time on more productive actions. For example, the human operator can use a combination of process data and decision support information from the AI system to consider current operating conditions and implement a more optimized setpoint, which could increase production or extend equipment life expectancy. In this paper, an example is presented, which is based on monitoring electric submersible pumps (ESPs) using a deep learning neural network that was trained on historical data from the process control system historian. In addition to outlining the benefits that AI-assisted surveillance provides when compared to conventional methods, the paper provides a system architecture blueprint for implementing an autonomous monitoring application across different types of ESP fleets by connecting sensor data directly to a cloud-based monitoring system. The paper builds on the work of previous SPE papers by providing up-to-date results of a pilot project, where a predictive maintenance model has been running for 10 months. On the project, 30 ESPs ranging in power from as low as 200-kW to as high as 500-kW were deployed and monitored using an AI-supported predictive maintenance model. To date, the results have been extremely positive. In one case, an ESP interruption was predicted by the application 12 days before the actual failure occurred. Following several months of testing/use of the ML-based predictive maintenance solution during the pilot deployment, the ESP fleet operator concluded that the system can detect multiple kinds of anomalies in advance, even previously unknown ones. This capability is a significant distinction between the AI-based model and conventional models employed by other ESP diagnostic tools. Although these new, unknown types of complex ESP operational anomalies were difficult to interpret as to their root causes, they could still have led to ESP performance degradation and possible failure nonetheless, if not mitigated or remediated.