Integrating Domain Knowledge with Machine Learning to Optimize Electrical Submersible Pump Performance

A data-driven workflow was developed to monitor electrical submersible pump (ESP) health using an anomaly detection method with high-frequency sensor data. The workflow would help maximize the run life of ESPs while reducing the cost of maintenance. The new workflow contrasts with conventional field maintenance which is often reactive and incurs additional downtime in logistics and inventory management in diagnosing the issues and taking the recommended actions. In contrast, using machine learning (ML) concepts can save operating costs, especially in the case of the ESPs widely used for artificial lift. Many operators augment ESPs with high-frequency (HF) sensors to monitor their performance, but much of this information remains either unused or partially used in post-failure analysis. The application of ML concepts in understanding ESP operational behavior complements the existing domain practice. The workflow we describe in this paper begins with domain knowledge and exploratory statistical analysis to find the key performance indicators (KPIs) related to ESP failure. Feature engineering and advanced ML techniques are used to build and test healthy ESP models for each selected KPI. Multiple health signals are fused to improve the performance of anomaly detection using historical ESP failure data and pullout reports as benchmarks. In a test of the workflow, the model was trained on the data from a group of active producing wells with reported historical events, failures, and pullout reports. The data contained several well events and several reported failures. This information was used to fine-tune the alarm thresholds for the health indicators. The model was able to detect approximately 70% of failure events (true positive rate) in the data set. The false alarm rates for the configured model were approximately at 20% (false positive rate). The solution can be implemented in a dashboard to monitor ESP KPIs and show health alarms. These alarms can be further prioritized based on the failure probability and remaining useful life of the ESP. The health signal degradation patterns can be captured and learned to predict the remaining useful life of the ESPs, thus enabling operators to allocate and prioritize maintenance resources. In addition, the analysis of ESP pullout reports can provide insight into the relationship between health signals and root causes of the failure, which can be structured into a formal Bayesian network to provide automatic root cause interpretation The data-driven approach takes advantage of the vast amount of reservoir, production, and facilities data and provides insights into nonlinear multidimensional relationships between parameters to better understand and optimize field development and to adopt a proactive approach toward equipment maintenance.