One of the applications of artificial intelligence (AI) within the industry is machine learning (ML), a powerful tool to perform predictive maintenance on machinery, as it allows for predicting future failure scenarios through data analysis and avoiding accidents. Electric submersible pumps (ESPs) are critical in the oil industry, as they are used in the artificial lift technique for oil extraction, standing out for their extensive production volume at high depths. However, ESPs are prone to frequent failures. For this reason, this study aimed to compare the performance of three popular nondeterministic ML algorithms in Rstudio, Random Forest (RF), Extreme Gradient Boosting (XGB), and Adaptive Boosting (AB), to determine the best predictive maintenance strategies for an ESP. This paper uses an artificial and public domain dataset containing cycle reading information, machine operating variables such as voltage and pressure expressed in average and standard deviation, maintenance cycle information, and component errors. The results showed that when predicting machine failures with the test dataset, the RF algorithm presented a lower variance (2.31273e-8) and higher average accuracy (99.878%); XGB (variance 4.61843e-8 and average accuracy of 99.850%), and AB (variance 7.688043e-8 and average accuracy of 99.844%). An independent sample Welch’s t-test was performed, and a significant difference was found between the accuracy of the RF and that of the other algorithms. It is concluded that RF proved more suitable for predictive maintenance of an ESP with the dataset used.
