Avaliação de algoritmos de machine learning para diagnóstico de operações em poços de bombeio mecânico

Abstract

This work investigates the performance of machine learning algorithms of varying complexities for diagnosing failures in rod sucker pumping systems, proposing that lower computational complexity methods can be effective in detecting failure patterns. Three algorithms — Logistic Regression, Random Forest, and Neural Networks — were tested, using load data and centroid descriptors as the primary features, with promising accuracy results (the lowest being 97%). The diagnosis of operational failures in sucker rod pumping systems is essential to reduce maintenance costs, prevent production downtime, and, consequently, increase efficiency. Dynamometer cards are widely used to identify failures, but their manual interpretation requires both time and specialized knowledge, especially when operators monitor multiple wells simultaneously. This study provides a foundation for the practical implementation of automated diagnostic systems, highlighting the potential for accessible and efficient solutions in the oil industry.