Photovoltaic hotspot fault detection based on detectron2 with faster R-CNN

Solar energy is widely recognised as one of the cleanest and most sustainable methods for harnessing sunlight into electricity. However, photovoltaic systems are susceptible to critical issues such as hotspots, often caused by various factors such as partial shading, short circuits or line-to-line faults. These issues can significantly diminish efficiency and lifespan, undermining the reliability and effectiveness of solar power generation. This study investigates the performance of various Faster R-CNN backbones within Detectron2 to determine the most effective architecture for detecting hotspots in photovoltaic panels through thermal imagery. After training each backbone individually, a systematic evaluation of the performance of each backbone architecture is undertaken, emphasising classification accuracy, loss convergence and performance metrics on the test set to get a complete insight into its efficiency. The model presented in this paper achieved an impressive ${9 8 \%}$ accuracy in foreground classification and an Average Precision at 50% Intersection over Union (AP50) of ${7 6. 5 8 \%}$. This approach demonstrates the potential of various Faster R-CNN backbones for real-time fault detection, supports proactive maintenance and enhances the efficiency and reliability of Photovoltaic systems.