基于生成对抗网络的太阳能电池缺陷增强方法

Solar cell defect enhancement method based on generative adversarial network

为了解决太阳能电池样本不均衡问题,提出负样本引导生成对抗网络的太阳能电池缺陷样本增强方法.通过在生成对抗模型中引入大量负样本和增加负样本引导损失,促进模型对正样本特征的表达,提升生成样本的多样性;设计自适应的权值约束方法,平衡生成器和判别器的表达能力,提升生成样本的质量.实验结果表明,在太阳能电池电致发光(EL)缺陷数据集上,提出方法的生成质量和检测精度优于深度卷积生成对抗网络(DCGAN)、梯度惩罚Wasserstein距离生成对抗网络(WGAN-GP)和一阶导数生成对抗网络(FOGAN);该方法的F测度较DCGAN、WGAN-GP和FOGAN分别最高提升了10%、8%和5%,具有较好的数据增强性能.在带钢表面缺陷数据集及DAGM 2007公共数据集上,提出方法的性能优于DCGAN、WGAN-GP和FOGAN,具有一定的泛化能力.

A solar cells defect sample enhancement method which is negative sample-guided generative adversarial network was proposed in order to solve the problem of sample imbalance of solar cells.The representation ability of positive samples features was promoted and the diversity of generated samples was improved by introducing many negative samples and mixing negative sample guidance loss in the generative adversarial network.An adaptive weight constraint method was designed to balance the representation ability of generators and discriminators,and the quality of generated samples was improved.The experimental results show that the proposed method outperforms deep convolutional generative adversarial network(DCGAN),Wasserstein generative adversarial network-gradient penalty(WGAN-GP)and first-order generative adversarial network(FOGAN)in generation quality and detection accuracy on electroluminescence(EL)defect data sets of solar cells.F-measure of the method was 10%,8%and 5%higher than DCGAN,WGAN-GP and FOGAN respectively,which showed better data enhancement performance.The performance of the proposed method is better than DCGAN,WGAN-GP and FOGAN on strip steel surface defect dataset and DAGM 2007 public dataset,which shows certain generalization ability.