Sistema Iot con análisis predictivo para el monitoreo de la salud de sistemas de bombeo de campos petroleros

Abstract

The operation of electric submersible pumps (ESPs) in oil fields is impacted by failures, downtime, and the limited use of digital tools for condition monitoring. To address this challenge, an IoT system with predictive analytics was developed to monitor the health of ESPs, with the aim of supporting maintenance decision-making and reducing the occurrence of critical failures in continuous production environments. To develop the solution, the CRISP-DM methodology was applied, defining a predictive analytics workflow based on historical data. In this stage, operational records were integrated, structured, and refined using electrical, hydraulic, mechanical, and thermal variables, and a preprocessing pipeline was implemented—including validation, cleaning, and feature transformation—to ensure data consistency and quality. On this basis, a multi-label dataset with five classification targets was built, enabling the joint execution of diagnostic and prognostic tasks. Subsequently, ensemble models based on decision trees were compared, and a multi-label Random Forest approach was adopted due to its balance of performance, robustness, and generalization capability across multiple classes and operational variability. During controlled-environment testing, the system achieved an approximate overall weighted precision of 85%, with per-label accuracies of 82% for failure variable, 83% for subcomponent, 82% for severity, 83% for pump state, and 91% for 24-hour failure prediction. Finally, deployment was carried out using Docker, a FastAPI inference service, InfluxDB, and Grafana dashboards, enabling visualization of process variables and estimated operating status, and demonstrating the technical feasibility of the proposed solution for predictive maintenance.