An Efficient Hot Spot Detection Method with Small Sample Learning for Photovoltaic Panels

With the rapid development of photovoltaic power stations, various faults frequently occur during the maintenance of photovoltaic panels. The hot spot is one of the critical issues which is not easy to observe and has a tremendously harmful impact. Traditional graph target recognition training requires a large amount of data in practical applications. However, there are many issues with hot spot detection in the real world. For example, images with hot spots often have a small sample number, detention methods are low recognition efficiency and slow speed. This paper constructs a fast photovoltaic hot spot detection method, Cheetah, for photovoltaic hot spot recognition with small sample learning. In this paper, we first preprocess the collected small sample dataset, and use the CycleGAN model to enhance the small sample dataset. Then, aiming at the problem of slow recognition speed, an improved YOLOX model is proposed for hot spot detection. Finally, our experimental results show that Cheetah has higher accuracy and stronger generalization ability and robustness than the traditional YOLOX method on the test set.