APM²Det: A Photovoltaic Hot-spot Fault Detection Network Based on Angle Perception and Model Migration

Under aerial imaging conditions, hot-spot faults exhibit angle distortion and small-scale characteristics in the infrared images, posing a challenge for precise localization in traditional detection algorithms. Therefore, a photovoltaic hot-spot fault detection algorithm based on angle perception and model migration is proposed, comprising a photovoltaic panel segmentation network and a hot-spot fault detection network. In the segmentation network, based on the DeepLabv3+ framework, a feature alignment module is designed to capture fine-grained photovoltaic panel regions. In the detection network, a cross-adaptive frequency Transformer module is constructed to enhance the detection accuracy of small-scale hot-spot faults, while a cross-scale shift network strengthens interactions between multi-scale features. Additionally, a convex hull feature adaptive method is introduced, which employs a joint optimization strategy to achieve precise localization of hot-spot faults with varying angles. To validate the algorithm’s superiority, it is compared with 13 detection algorithms. Experimental results show that our algorithm achieves a detection accuracy as high as 92.3% and can accurately detect hot-spot faults with diverse angles in complex environments.