控制与信息技术Issue(5)：24-30,7.DOI:10.13889/j.issn.2096-5427.2025.05.200
基于深度卷积神经网络的太阳能电池电致发光图像缺陷分类实验
Experimental Study on Classification of Solar Cell Defects from Electro-luminescence Imaging Based on Deep Convolutional Neural Networks
摘要
Abstract
Electroluminescence(EL)imaging can sensitively reflect internal defects in solar cells,yet analyzing these images presents challenges due to high complexity,defect diversity,and large data volumes.To address the lack of criteria for model selection in industrial scenarios,this paper systematically evaluates the performance of five classic convolutional neural networks across three defect classification tasks,aiming to provide a theoretical foundation and methodological support for the intelligent detection and classification of cell defects.First,to mitigate severe class imbalance in the PVEL-AD datasets,stratified sampling was employed to construct 4-class and 7-class subsets,which,together with the original 2-class ELPV dataset,form a multi-level experimental framework.Second,using metrics such as balanced accuracy and F1-score,along with analyses of model runtime and convergence curves,a comprehensive evaluation was conducted on the performance of LeNet-5,AlexNet,VGG16,GoogLeNet,and ResNet18 in classification tasks.Experimental results demonstrated that ResNet18 significantly outperformed others in the 7-class task,achieving an F1-score of 93.70%,while its training time was only 47.5%that of VGG16,validating the superiority of residual structures in complex classification tasks.Through systematic experiments,this paper reveals the intrinsic relationship between model structures and classification performance,providing a decision-making basis for the efficient deployment of inspection systems for photovoltaic production lines.关键词
太阳能电池/电致发光图像/缺陷分类/卷积神经网络/深度学习Key words
solar cell/electroluminescence imaging/defect classification/convolutional neural network/deep learning分类
信息技术与安全科学引用本文复制引用
毕莹,张荣巧,常亚鑫,张拓..基于深度卷积神经网络的太阳能电池电致发光图像缺陷分类实验[J].控制与信息技术,2025,(5):24-30,7.基金项目
国家自然科学基金面上项目(62376253) （62376253）
