基于深度学习的焊缝定位与缺陷识别

基于深度学习的焊缝定位与缺陷识别 Weld location and defect identification based on deep learning 薛龙 XUE Long 1 first-author 曹楷顺 CAO Kaishun 1 黄军芬 HUANG Junfen 1 corresp huangjunfen@bipt.edu.cn 黄继强 HUANG Jiqiang 1 邹勇 ZOU Yong 1 曹莹瑜 CAO Yingyu 1 北京石油化工学院机械工程学院 北京石油化工学院机械工程学院 School of Mechanical Engineering,Beijing Institute of Petrochemical Technology School of Mechanical Engineering,Beijing Institute of Petrochemical Technology 针对大型构件缺陷焊缝的自动定位及缺陷识别是实现焊缝打磨、补焊等自动化操作的必要条件。大型构件焊缝及焊缝缺陷图像具有形状多样、灰度分布随机等特点,加大了图像处理的难度。提出一种基于深度学习的焊缝定位及缺陷识别方法,通过深度学习目标检测方法确定焊缝位置并识别焊瘤及不合格缺陷,通过深度学习语义分割方法识别气孔及凹坑缺陷。选取FPN网络结构创建和训练焊缝定位及缺陷识别模型,并通过增加样本数量完成模型优化,焊缝定位识别准确率达到95%,焊瘤识别准确率达到98%,气孔与凹坑两类缺陷的识别准确率约为91.8%。 Automatic location and defect identification of weld defects for large-scale components are the necessary condition to realize the automatic operations of weld grinding and repairing. Due to the characteristics of large-scale component weld and weld defect images, such as shape diversity and random gray distribution, the difficulty of image processing is increased. A method of weld location and defect recognition based on deep learning was proposed. The weld position was determined and the weld bead and unqualified defects were identified through the deep learning target detection method. The gas pore and pit defects were identified by the deep learning semantic segmentation method. The weld location and defect identification models were created and trained based on the FPN network structure, and the model optimization was completed by increasing the number of samples. The accuracy rate of weld location is 95%, the identification accuracy rate of weld bead is 98%, and the identification accuracy rate of gas pore and pit is about 91.8%. 焊缝定位 缺陷识别 深度学习方法 weld location defect identification deep learning method TP18;TG441.7 国家重点研发计划项目（2017YFB1303300） 2022-03-08 9 51