基于机器视觉的焊点质量检测方法研究  

Solder Joint Quality Inspection Method Based On Machine Vision

作　　者：孙丽[1] 张瀚文 白雨轩 王品烁 SUN Li;ZHANG Hanwen;BAI Yuxuan;WANG Pingshuo(School of Mechanical Engineering,Dalian Jiaotong University,Dalian 116028,China;School of Economics and Management,Dalian Jiaotong University,Dalian 116028,China)

机构地区：[1]大连交通大学机械工程学院,辽宁大连116028 [2]大连交通大学经济管理学院,辽宁大连116028

出　　处：《大连交通大学学报》2025年第1期93-100,共8页Journal of Dalian Jiaotong University

基　　金：辽宁省交通科技研究计划资助项目(202149);辽宁省自然科学基金资助项目(2021-KF-15-02)。

摘　　要：针对L公司楼宇机组控制面板制作过程中,各种元器件PCB板在焊接时人工对焊点质量进行缺陷检测的效率低、误差大等问题,采用机器视觉方法实现快速识别。首先,对采集到的图像数据进行预处理,提出改进的小波阈值去噪方法和改进NGO(Northern Goshawk Optimization)优化的OTSU多阈值分割方法;其次,分别采用HOG、LBP、GLCM 3种特征和SVM、KNN、Tree 3种模型共12种分类情况对焊点图像进行描述,用于更好地将焊点图像的信息体现出来;最后,将CNN_SVM与传统的CNN及SVM模型进行对比,CNN_SVM对焊点图像分类的准确率为98.3%,与CNN及SVM对比分别提高了2.5%和4.6%。同时构建了L公司焊点数据集,试验结果证明,同人工对比,单个焊点检测时间约减少了0.9 s。In view of the problems of low efficiency and significant error in manually detecting defects in the quality of solder joints during the production process of building units of Company L,machine vision methods are used to achieve rapid identification.Firstly,the collected image data is preprocessed,and an improved wavelet threshold denoising method and an improved OTSU multi-threshold segmentation method optimized by NGO(Northern Goshawk Optimization)are proposed.Then,the three characteristics of HOG,LBP,GLCM and SVM,KNN and Tree are used to describe the solder joint image in twelve categories,which are used better to reflect the information of the solder joint image.On this basis,the CNN_SVM traditional CNN and SVM models are compared.The experiment results show that the accuracy of image classification of CNN_SVM solder joints is 98.3%,which is improved by 2.5%and 4.6%compared with CNN and SVM respectively.The Solder Joint Data Set of Company L was constructed,and the experiment results prove that compared with manual inspection,a single solder joint is improved by about 0.9s.

关 键 词：机器视觉 焊点质量 分类识别 卷积神经网络 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]