Towards Efficient Solar Panel Inspection: A YOLO-based Method for Hotspot Detection

• Main Authors: Ameerdin, Muhammad Irshat, Jamaluddin, Muhammad Herman, Shukor, Ahmad Zaki, Al Hakim Kamaruzaman, Luqman, Mohamad, Syazwani
• Format: Conference Proceeding
• Language: English
• Subjects: Accuracy; Deep learning; Hotspots; Refining; Roboflow; Solar energy; Solar Panel; System performance; Thermal sensors; Training; YOLO
• ISSN: 2836-4317
• DOI: 10.1109/ISCAIE61308.2024.10576312

Solar energy that captured by the photovoltaic (PV) cells has gained recognition as an important factor in the global search for sustainable and clean energy sources in recent years. One of the Sustainable Development Goals (SDG) that solar technology directly supports is Affordable and Clean Energy. It can help increase access to clean energy sources by improving the efficiency and dependability of solar panels through minimizing its defects. However, a variety of defects can shorten the lifespan and effectiveness of PV array, which are crucial components of solar energy systems. The study concentrates on detecting hotspots on solar panels, identifiable through thermal imaging technology. This project aims to develop a deep learning-based approach for defect detection of solar panels. The project unfolds with a primary goal, that is designing the integration of a thermal sensor and deep learning to detect and identify defects in PV panels. It follows with crafting a robust algorithm within the deep learning environment for effective defect detection and identification. Next, the algorithm's performance will be evaluated, emphasizing its reliability and accuracy in enhancing defect detection. The process begins with physically examining a solar panel, followed by using a drone-mounted thermal camera to capture thermal images. After obtaining enough data, the images undergo model generation by labelling and annotation process using Roboflow. The model is then tested and trained for defect detection using YOLOv8. Once the desired accuracy is reached, the dataset is formatted. A user-friendly graphical interface is developed for ease of interaction. Then, the system's performance is evaluated using a confusion matrix to gauge the effectiveness of the defect detection approach. The panel's defect will be confirmed with the manual inspection. Based on the early result obtained, the model's confidence level that has been acquired is 76%.