基于机器学习的不锈钢薄板MIG焊焊穿缺陷识别  

Identification of MIG Welding Burn-through Defects in Stainless Steel Sheet Based on Machine Learning

作　　者：王杰[1] 张志芬[1] 秦锐 温广瑞[1] WANG Jie;ZHANG Zhifen;QIN Rui;WEN Guangrui(iHarbour Academy of Frontier Equipment,Xi'an Jiaotong University,Xi'an 710100,China)

机构地区：[1]西安交通大学高端装备研究院,陕西西安710100

出　　处：《电焊机》2023年第9期70-77,共8页Electric Welding Machine

基　　金：国家重点研发计划(2020YFB1710002)。

摘　　要：随着焊接工艺在工业环境中的广泛应用,对焊接缺陷的研究受到越来越多的关注,焊穿缺陷作为焊接过程中最严重的缺陷之一,必须采取相应的措施对其进行监测以及控制。提出了一种采用被动视觉传感技术,利用工业CCD相机结合减光、滤光系统消除焊接过程中强弧光干扰,获得MIG焊熔池图像的方法。设计正交试验并从图像的多维特征出发,提取熔池图像4个种类对应的10个维度特征作为机器学习的输入。在进行焊穿缺陷的标签标定时,基于所提出的特征提取对正交试验中获得的熔池图像数据进行预处理,通过反向搜索实现对无效数据的剔除和有效数据的正确归类,同时发现了数据在空间分布的规律。最后探究了不同维度输入特征下三种分类器的识别效果,结果表明:所选择的特征在三种分类器上均获得了不低于97%的分类精度,为后续开发基于视觉在线焊穿缺陷检测系统提供了技术支撑和理论依据。With the wide application of welding technology in industrial environment,more and more scholars pay attention to the study of welding defects.As one of the most serious defects in the welding process,the corresponding measures must be taken to monitor and control the burn-through defects.In this paper,a passive visual sensing technique is proposed to eliminate the strong arc light interference in the welding process by using industrial CCD camera combined with dimming and filtering system,then obtain MIG welding pool image.Based on the multi-dimensional features of the image,ten dimensional features corresponding to the four categories of molten pool image are extracted as the input of machine learning.During the label calibration of normal and burn-through defects,the image data of molten pool obtained from orthogonal test were preprocessed based on the proposed feature extraction,the invalid data were eliminated and the valid data were correctly classified by reverse search.Meanwhile,the spatial distribution of data was found.Finally,the recognition effects of the three classifiers under the input features of different dimensions were explored,and the results showed that the selected features obtained classification accuracy of no less than 97%on the three classifiers,which provided technical support and theoretical basis for the subsequent development of visual online burn-through defects detection system.

关 键 词：MIG焊 视觉传感 熔池图像处理 机器学习 缺陷识别 

分 类 号：TG409[金属学及工艺—焊接]