Diagnóstico das condições de operação e falhas de sensor em poços operando por bombeio mecânico utilizando Machine Learning

Abstract

In oil fields with many wells operated by sucker-rod pumping, due to the lack of early diagnosis of operational conditions or sensor failures, several problems can go unnoticed. These problems, can lead to significant losses, such as increased downtime, higher operational expenses (OPEX), decreased efficiency and production loss. In practice, the identification and diagnosis of operational conditions are primarily performed based on surface and downhole dynamometer cards, using pre-established patterns and relying on human visual inspection at operation centers. This task consumes a lot of time and effort and requires expertise, as it can be influenced by subjective factors. However, in recent years, with the advancements in Machine Learning (ML) algorithms, several research studies in this field have achieved promising results in diagnosing operational conditions, demonstrating the potential of ML for this purpose. Nevertheless, there are still common doubts regarding the difficulty level of classifying dynamometer cards, the best algorithm, the most suitable shape descriptor, the optimal metrics, and the impact of imbalanced datasets. In pursuit of answers to these questions, this work utilized real data from 38 wells operating via sucker-rod pumping in the Mossoró region, RN, Brazil. Over 50,000 cards were classified by the author of the work and categorized into eight operating modes and two common sensor failures in this field. Seventy-three tests were conducted and divided into nine groups. Three algorithms, namely Decision Tree, Random Forest, and XGBoost, were tested with and without hyperparameters tuning, along with pipelines provided by Automated Machine Learning (AutoML). The descriptors used included Centroid, Fourier, and Wavelet, in addition to the load values from the dynamometer card. Both balanced and imbalanced datasets were tested. The results confirm the feasibility of applying ML for diagnosing operational conditions and sensor failures in sucker-rod pumping systems, as 80.82% of the tests achieved an accuracy above 90%, with a remarkable accuracy of 99.9% also being attained. Finally, a portion of the implementations suggested in this research was evaluated in a production environment. Dynamometer cards from over a hundred wells were assessed, yielding an accuracy exceeding 96%, which reinforces the effectiveness of the implementations in a real-world context. Overall, the results highlight the potential of ML as a promising approach for diagnosing operational conditions and sensor failures in sucker-rod pumping systems.