Sistema de monitoreo en tiempo real con Edge Computing para equipos de bombeo petrolero en predicción de fallos

Abstract

Failures in oil pumping equipment cause unplanned downtime and high costs. To address this need, a prototype real-time monitoring system based on edge computing was developed, operating entirely on a Raspberry Pi 5 located next to the data source. The system captures pressure, vibration, temperature, flow rate, and power consumption, processes the information locally, and issues early warnings without relying on the cloud. The development followed the CRISP-DM methodology. The “Pump Sensor Data” dataset (Kaggle; AI4A 2020; CWRU Bearing; Microsoft Azure) was used for modeling, comparing four classic algorithms. Random Forest was selected for its performance and stability. The architecture integrates local storage in MariaDB and visualization in Grafana; the flow includes reading, normalization, status classification, and recording of results, with automatic notifications when the established operating threshold is exceeded. In tests with a test set, Random Forest achieved 92.54% accuracy and 92.25% F1-Score. The average time per prediction was 81.63 ms, adequate for the 3-second acquisition cycle and 5-second dashboard refresh. Running continuously on the Raspberry Pi, the system started in less than 5 seconds and exceeded 1,000 readings without interruption, validating its stable operation, low cost, and potential for field deployment for predictive maintenance.