基于深度学习的带钢焊缝杯突试验结果分类  

Classification of Cupping Test Results for Strip Welds Based on Deep Learning

作　　者：张勇[1] ZHANG Yong(Maanshan Iron and Steel Co.,Ltd.,Maanshan,Anhui 243000,China)

机构地区：[1]马鞍山钢铁股份有限公司,安徽马鞍山243000

出　　处：《冶金动力》2024年第1期82-85,共4页Metallurgical Power

摘　　要：为了实现对带钢焊缝月牙边杯突试验结果的自动分类,设计了一种基于轻量级网络的分类方法,首先采用数据增强扩充数据集样本数量,接着引入Grad-CAM算法对试验模型的中间层以热力图的形式进行可视化,最后结合MobileNet V3网络可视化中间层设计了冻结特征提取部分的迁移学习训练方法,并对比测试了4种轻量型网络,试验结果表明基于迁移学习的MobileNet V3网络具有较好的缺陷分类能力。In order to realize the automatic classification of the cupping test results of crescent edge of strip weld,a classification method based on lightweight network is designed.Firstly,data enhancement is used to expand the number of samples in the dataset,then the Grad-CAM algorithm is introduced to visualize the intermediate layer of the test model in the form of heat maps.Finally,a migration learning training method for the freezing feature extraction part is designed in conjunction with the visualization of the intermediate layer of the MobileNet V3 network,and the four types of lightweight networks are tested in comparison,the experimental results show that the MobileNet V3 network based on migration learning has a better ability to classify the defects.

关 键 词：杯突试验 MobileNet V3 Grad-CAM 迁移学习 

分 类 号：TP27[自动化与计算机技术—检测技术与自动化装置]