Defect Diagnosis of Visible Light Images of Photovoltaic Modules Based on Active Transfer Learning

Huiqing RaoLi Qiong^*Long ChenZiwei XuJiayue ShuPu Xu

• Nanchang Hangkong University, Nanchang, China

To address the challenges of missed detection and false detection of bird droppings and dust defects caused by data imbalance during actual photovoltaic power station inspections, and to enhance operational efficiency and power generation safety, this paper proposes an improved active transfer learning method. The regular grid lines on the surface of photovoltaic modules are easily misclassified as linear defects, while small-sample defects are difficult to identify under extreme lighting conditions, and existing methods exhibit weak generalization capabilities. To tackle these issues, this paper first introduces the SENet attention mechanism into the transfer learning framework, constructing an SE-ResNet architecture that suppresses interference from grid line textures through channel attention mechanisms, thereby enhancing the perception of genuine defects. Secondly, in the active learning module, a sample selection strategy combining image uncertainty calculation and hierarchical clustering is proposed, prioritizing the screening of informative and representative hard samples to efficiently improve the model's generalization performance in small-sample and extreme scenarios. Experimental results show that, compared to traditional active transfer learning models, the proposed method achieves significant improvements in recall for bird droppings and dust defects by 8.22% and 11.11%, respectively, on a self-built photovoltaic defect dataset, with an overall mAP increase of 9.98%. Both experimental and statistical analyses demonstrate that the method significantly outperforms baseline models.