基于生成对抗网络的小样本激光焊接缺陷数据集生成  

Defect generation of small sample laser welding based on generative adversarial network

作　　者：肖思哲 刘振国[1] 闫志鸿[2] 李敏 黄及远 XIAO Sizhe;LIU Zhenguo;YAN Zhihong;LI Min;HUANG Jiyuan(Beijing Research Institute of Automation for Machinery Industry,Beijing,100120,China;Faculty of Materials and Manufacturing,Beijing University of Technology,Beijing,100124,China)

机构地区：[1]北京机械工业自动化研究所,北京100120 [2]北京工业大学材料与制造学部,北京100124

出　　处：《焊接学报》2022年第10期43-48,I0005,I0006,共8页Transactions of The China Welding Institution

基　　金：国家自然科学基金项目(51975015)。

摘　　要：为了提升深度学习模型在不均衡小样本激光焊接表面缺陷数据集上的性能,优化了小样本数据量输入下的对抗生成网络(Generative Adversarial Network,GAN)模型.通过对比激光焊接缺陷和其他用于测试对抗生成网络的公共数据集在特征复杂度上的区别,设计了一种全新的OCM (one class mixup)模块,并将其引入至针对有限样本的stylegan2-ada中,以提升GAN的性能,加快其收敛.试验结果表明,在分类模型上,通过OCM-stylegan2-ada生成的数据集,比原始数据集性能提升40%,比使用mixup和stylegan2-ada增强后的数据集性能上提升20%,同时生成的焊接缺陷图片质量大幅提升.To improve the performance of deep learning models on an unbalanced dataset of small samples of laser welded surface defects,an adversarial generative network(GAN)model using small datasets as input is optimized.By comparing the difference in feature complexity between laser welding defects and other public datasets used to test adversarial generative networks,a new OCM(one class mixup) module is designed and introduced into the stylegan2-ada for a limited number of samples to improve the performance of the adversarial generative network and accelerate its convergence.The results show that the dataset generated by OCM-stylegan2-ada improves the performance of the classification model by 40% over the original dataset and by 20% over the dataset enhanced with mixup and stylegan2-ada.Also the quality of the visually generated images of weld defects is greatly improved.

关 键 词：焊接缺陷 对抗生成网络 小样本 不均衡数据集 

分 类 号：TG441[金属学及工艺—焊接]