Research Information System
12th International New York Conference Onevolving Trends in Interdisciplinary Research & Practices, New York, United States Of America, 21 - 24 August 2025, pp.336-348, (Full Text)
Accurate
and timely fault detection in photovoltaic (PV) panels is critical to
maintaining operational reliability and energy efficiency. In this study, two
deep learning-based approaches are comparatively evaluated for hotspot anomaly
detection using thermal images: a classical U-Net-based segmentation model and
a YOLOv8-based object detection model. Both models are trained and tested on a
real-world thermal image dataset acquired from the Roboflow platform under
identical preprocessing conditions. The U-Net model, supported by the segment
anything model (SAM) for mask generation, demonstrated strong segmentation
capabilities with an average F1-score of 0.8611 and a maximum F1-score of
0.9787, along with an average IoU of 0.7624 and a maximum IoU of 0.9582. On the
other hand, the YOLOv8 model outperformed in detection efficiency, achieving a
mean average precision (mAP@0.5) of 0.8643, peaking at 0.9283, and exhibiting a
faster training curve. The comparative analysis highlights that while U-Net
excels in fine-grained pixel-level segmentation, YOLOv8 is more suited for
real-time deployment scenarios. The findings underscore the complementary
strengths of the two approaches: U-Net is better suited for detailed,
pixel-level segmentation tasks, while YOLOv8 offers faster and more efficient
detection suitable for real-time deployment. This comparative analysis provides
practical insights into model selection strategies for PV fault monitoring
systems, based on application-specific requirements.