Harnessing Machine Learning for Proactive Detection and Predictive Maintenance of Sticking Pumps in Rod Lift Oil Wells

Abstract

This study introduces a novel approach to assist operators in identifying wells with sticking pumps that are suitable candidates for treatment. This initiative is crucial for enhancing operational efficiency and mitigating downtime due to pump failures in oil wells produced with sucker rod pumps. The economic benefits of early intervention of a sticking pump are significant. Elimination of a sticking pump in the North Dakota Bakken can result in the avoidance of a ~$50,000 workover, in addition to the elimination of days of non-productive downtime while the well waits for a workover rig. Comparing the potential benefits to the ~$700 cost of a well treatment translates into a compelling economic rationale for early intervention of a sticking pump. A machine learning model was trained from historical maintenance logs and operational well data for over 800 horizontal sucker rod pump wells in the North Dakota Bakken shale. More than 1,000 maintenance logs were provided by the operator to identify wells that had failed with a stuck pump. The operational data was examined and features were derived that could be useful to indicate a sticking pump. The machine learning model was trained on three weeks of such data before recorded stuck pump failures. Each week a list of the top D well candidates most likely to fail as predicted by the model was provided to the operator for review. Preliminary analysis of the model's efficacy reveals it to be satisfactory, with further room for optimization. Evaluating the model's performance metrics against the operator-provided training dataset yielded a precision of 68.4% and a recall of 61.6%. These metrics indicate the model's true positive rate and its ability to correctly identify actual positives. Flush treatments were applied to a subset of identified sticking pumps after human expert review, resulting in significant improvements compared to a previous method. Among the 7 treated wells, a pump health enhancement rate of 71.42% and a production improvement rate of 57.14% were observed, highlighting the effectiveness of the model in human-in-the-loop predictive maintenance. Live field data is currently being aggregated and analyzed for model validation. This data will be instrumental in recalibrating the model through additional training iterations, aimed at refining its predictive accuracy and reducing false positives and negatives. We recommend continuous improvement of the training data to include a greater variety of cases, along with enhancements in the explainability of the model’s predictions. Additionally, there is a need for new tools to document treatment procedures and outcomes effectively to accelerate the feedback loop and improve model accuracy. Future exploration will focus on alternative objectives more closely related to treatment success, such as increased production or decreased friction. The study introduces a novel automated tool for detecting sticking pumps using machine learning techniques. It represents a scalable technology framework capable of enhancing operational efficiency and being adapted to predicting other types of equipment failures proactively. Moreover, the study outlines a practical, operator-focused workflow, emphasizing real-world applicability and seamless integration into existing operational processes.