Thermal Image-Based Solar PV Fault Detection Leveraging YOLOv5 Model Variants

Abstract

This paper analyzed the performance of deep learning model variant of YOLOv5 for fault detection in solar photovoltaic (PV) systems using infrared thermography images. The explosive growth of solar PV energy globally has posed challenges in maintaining these systems. Traditional maintenance methods rely on manual on-site inspection, which are time-consuming - particularly for large numbers of solar panels - and may cause delays in maintenance work. To overcome these issues, Artificial Intelligence (AI) is utilized to develop automated and real-time solar PV fault detection using infrared thermal images. This paper investigates variants of deep learning techniques, specifically the YOLO architecture, in solar PV fault detection across eight classes using YOLOv5s, YOLOv5m, and YOLOv5l models. The eight fault classes considered in this study are: cell, multi-cell, shading, string, substring, reverse polarity, module and junction box. The training results and evaluation metrics were carefully examined to identify the most effective model. The YOLOv5l model emerged as the top-performing option, demonstrating superior detection accuracy, localization accuracy, class prediction accuracy, precision, recall, and mean average precision (mAP). The overall accuracy percentages for the YOLOv5s and YOLOv5m models were 46.67% and 51.50%, respectively. YOLOv5l achieved 53.40% accuracy, showing a clear improvement compared to the YOLOv5s and YOLOv5m models. Hence, this study highlights the potential of YOLOv5l model for detecting solar PV faults across eight classes using infrared thermography images.