中圖分類號:TP391.41 文獻標(biāo)志碼:A DOI: 10.16157/j.issn.0258-7998.245362 中文引用格式: 黨寧,李世峰,于坤義. 基于多尺度伸縮卷積與注意力機制的光伏組件缺陷分割算法[J]. 電子技術(shù)應(yīng)用,2025,51(4):66-71. 英文引用格式: Dang Ning,Li Shifeng,Yu Kunyi. Defect segmentation network of photovoltaic modules based on multi-scale convolution and attention mechanism[J]. Application of Electronic Technique,2025,51(4):66-71.
Defect segmentation network of photovoltaic modules based on multi-scale convolution and attention mechanism
Dang Ning,Li Shifeng,Yu Kunyi
State Power Investment Group Gansu Electric Power Co., Ltd.
Abstract: In the inspection process of photovoltaic system, unmanned aerial vehicles need to accurately and quickly identify the defects of photovoltaic modules. Therefore, a photovoltaic module defect segmentation network based on multi-scale convolution and attention mechanism is proposed. Firstly, a multi-scale convolution module is added to each Stage of the traditional U-Net network to segment the defects of photovoltaic modules, and the pixel accuracy rate reaches 98.61%. Compared with the traditional U-Net and FCN networks, the accuracy rate is increased by 0.32% and 1.17% respectively, and the algorithm consumes 0.054 s. Compared with the comparison, the segmentation algorithm improves by 0.006 s~0.013 s; Then the split defect mask and the original image are operated. Finally, the defects (heat spots, cracks, bird feces) of photovoltaic modules are detected and classified by lightweight network MobileNetV3, with an accuracy of 98.82%. Compared with SqueezeNet, ShuffleNet V2, and GhostNet, the average detection time is increased by 0.43%, 1.08%, and 0.8%, respectively. Compared with Squeezenet, ShuffleNet V2, and GhostNet, the average detection time is increased by 0.002 s~0.036 s. The experimental results show that the defect segmentation network based on multi-scale convolution and attention mechanism has high accuracy and recognition rate.