Artificial Intelligence-Based Fault Detection in Photovoltaic Power Plants Using Drone-Acquired Thermal Imaging
Abstract
This work presents an intelligent technique to support maintenance in large-scale photovoltaic power plants, utilizing a semi-automated pipeline that combines manual data acquisition with automated AI-powered analysis. The solution was designed and validated to streamline the monitoring procedure at the Fluminense Federal University (UFF) photovoltaic plant in Prédio H Praia Vermelha Campus. A certified pilot operates a DJI Matrice 30 Thermal (M30T) drone to manually conduct flights and collect thermal images of the solar panels. After data transfer, an AI algorithm, based on the YOLOv8 architecture, automatically analyzes the thermal images to detect and classify anomalies, with a primary focus on hotspots. The trained model achieved a robust performance in hotspot detection, reaching a mean Average Precision (mAP@0.5) of 85.3%. This automated classification process significantly accelerates Operation and Maintenance (O&M) workflows, making the application a commercially viable tool for predictive maintenance in the PV industry. The final system issues alerts indicating the exact position (X, Y coordinates) and ID of the faulty panel, enabling targeted and efficient repair interventions.
Keywords
Large scale photovoltaic systems, predictive maintenance, Artificial Intelligence, fault detection, Thermography, YOLOv8, O&M