Machine Learning Based Damage Detection in Photovoltaic Arrays Using UAV-Acquired Infrared and Visual Imagery

• Published in: Conference proceedings (International Conference on Unmanned Aircraft Systems. Online) pp. 264 - 271
• Main Authors: Barrett, Aidan, Bratanov, Dmitry, Amarasingam, Narmilan, Sera, Dezso, Gonzalez, Felipe
• Format: Conference Proceeding
• Language: English
• Published: IEEE 04.06.2024
• Subjects: Australia; deep learning; drone; Maintenance; Manuals; Pipelines; PVarray imaging; solar energy; solar panel; Solar panels; Training; Visualization
• ISSN: 2575-7296
• DOI: 10.1109/ICUAS60882.2024.10556847

The rapid global expansion of solar panel installations necessitates more efficient and cost-effective methods for performance monitoring and maintenance. The New England Solar Project, Australia's largest solar installation is a prime example of the scale and complexity of modern solar farms, making it increasingly challenging to rely solely on manual ground-based inspections. This paper addresses the challenge by focusing on the integration of unmanned aerial systems (UAS) based imagery and deep learning (DL) techniques to develop a semi-automated pipeline for accurately identifying and classifying photovoltaic (PV) cell surface damage. The study leverages the YOLOv8 and Faster R-CNN models to achieve this goal. Drone based visual and infrared spectrum imagery collected from a solar installation site in Queensland, Australia, during October 2022 form the basis of the dataset, enabling the training and evaluation of these models. Three distinct damage classifications (Single-Cell, Multi-Cell, and Surface-Anomaly) were established with input from a subject matter expert to ensure accurate categorization of damage types. The research results indicate promising outcomes for classifying the distinct damage classes. The YOLOv8s-seg model achieved a mean average precision (mAP) of 87% to segment the solar panels. The YOLOv8m model, trained with a relatively small dataset, achieved a commendable mAP of 76% for solar panel damage detection. The Faster R-CNN model showed potential in detecting damage with high confidence, although a more comprehensive evaluation is needed. This research contributes to the broader goal of enhancing preventive maintenance practices, thereby reducing damage-related losses, and ensuring the long-term sustainability of solar installations.