Abstract
This paper describes a data-driven methodology to identify tubing leaks in oil production wells installed with electrical submersible pumps. Tubing leaks result in fluid recirculation within the well and lead to numerous operational problems. Whilst it is possible for a technician to manually review wells for tubing leaks, this task becomes unmanageable for hundreds of wells. The methodology and workflow presented overcomes this problem by automatically identifying tubing leaks using commonly metered real-time parameters along with daily well test data. Data was collected from several onshore ESP production wells with confirmed tubing leaks. After preprocessing and feature engineering, four machine learning models were built and evaluated using the F1-score metric, as the problem was formulated as a supervised classification problem. In addition, SHAP scores were used to further investigate how classifications were made from the best performing model and ensure conformance with the physical understanding of well performance during a tubing leak. All the machine learning models evaluated performed well, with the best model providing an F1-score of 94.9% on the test data set. SHAP scores confirmed correct relative feature importance which increased confidence in the selected model's predictive ability. This workflow has been made available within ChampionX's suite of digital surveillance and optimization products.
