Comparative Detection of Photovoltaic Panel Faults Using U-Net And Yolov8 Models Based on Thermal Images

<p class="MsoNormal" style="margin-top:6.0pt;margin-right:0cm;margin-bottom:6.0pt;margin-left:0cm;text-align:justify;line-height:normal"><span lang="EN-US" style="font-family:&quot;Times New Roman&quot;,serif;mso-ansi-language:EN-US">Accurateand timely fault detection in photovoltaic (PV) panels is critical tomaintaining operational reliability and energy efficiency. In this study, twodeep learning-based approaches are comparatively evaluated for hotspot anomalydetection using thermal images: a classical U-Net-based segmentation model anda YOLOv8-based object detection model. Both models are trained and tested on areal-world thermal image dataset acquired from the Roboflow platform underidentical preprocessing conditions. The U-Net model, supported by the segmentanything model (SAM) for mask generation, demonstrated strong segmentationcapabilities with an average F1-score of 0.8611 and a maximum F1-score of0.9787, along with an average IoU of 0.7624 and a maximum IoU of 0.9582. On theother hand, the YOLOv8 model outperformed in detection efficiency, achieving amean average precision ([email protected]) of 0.8643, peaking at 0.9283, and exhibiting afaster training curve. The comparative analysis highlights that while U-Netexcels in fine-grained pixel-level segmentation, YOLOv8 is more suited forreal-time deployment scenarios. The findings underscore the complementarystrengths of the two approaches: U-Net is better suited for detailed,pixel-level segmentation tasks, while YOLOv8 offers faster and more efficientdetection suitable for real-time deployment. This comparative analysis providespractical insights into model selection strategies for PV fault monitoringsystems, based on application-specific requirements.<o:p></o:p></span>