基于深度学习的激光熔覆层表面气孔识别研究  

Research on surface porosity recognition of laser cladding layer based on deep learning

作　　者：崔陆军[1,2] 刘亚轩 郭士锐 李海洋[1,2] CUI Lujun;LIU Yaxuan;GUo Shirui;Li Haiyang(Zhongyuan University of Technology,School of Mechanical&Electronic Engineering,Zhengzhou 450007,China;Zhengzhou Key Laboratory of Laser Additive Manufacturing Technology,Mechanical Industry Key Laboratory of Optical Sensing and Testing Technology,Zhengzhou 450o00,China)

机构地区：[1]中原工学院机电学院,河南郑州450007 [2]郑州市激光增材制造技术重点实验室,机械工业光学传感与测试技术重点实验室,河南郑州450000

出　　处：《光学技术》2023年第6期673-679,共7页Optical Technique

基　　金：机械工业光学传感与测试技术重点实验室(2022SA-04-15);河南省自然科学基金项目(202300410514);河南省重点研发与推广专项(科技攻关)项目(232102220051);河南省水下智能装备重点实验室开放基金(YZC-2206-B0030-01-060);河南省高等教育教学改革研究与实践项目(学位与研究生教育)项目(2021SJGLX143Y);河南省研究生教育改革与质量提升工程项目(YJS2022AL057);安徽理工大学矿山智能装备与技术安徽省重点实验室开放基金项目(KSZN202002003);中原工学院科研团队发展项目“激光增材制造技术团队”(K2021TD002);中原工学院优势学科实力提升计划资助“学科骨干教师支持计划”项目(GG202220)与“骨干学科发展计划”项目(FZ202204);中原工学院研究生校企联合课程专项经费资助建设项目(LH202301);中原工学院基本科研业务费专项资金项目(K2019QN006)。

摘　　要：为了解决熔覆层表面气孔识别技术中耗时且准确度不足的问题,文章利用深度学习技术中的语义分割网络提出了基于U-net神经网络识别熔覆层表面气孔的2BNC-Unet神经网络。通过引入Batch Normalization层以及串联注意力机制(CBAM)合理部署在神经网络中,选取交并比(IoU)与Dice系数作为网络的评价指标。研究结果表明:在测试集中,2BNC-Unet网络的交并比与Dice系数分别为86.96%、86.42%,相比U-net神经网络分别提高了7.65%、4.73%。同时为了验证该网络的性能,选用SegNet、2BNC-Unet与U-net神经网络进行对比实验,结果表明2BNC-Unet的分割效果不仅优于SegNet和U-net网络,而且熔覆层表面的气孔细节能够被完整地分割。在深度学习技术中2BNC-Unet的分割速度和准确度都有了显著地提高,气孔的分割为熔覆层的性能分析提供了帮助。In order to solve the problems of time-consuming processes and insufficient accuracy in the surface porosity recognition technology of the cladding layer,A 2BNC-Unet neural network based on the U-Net neural network is proposed.The goal is to identify pores on the cladding layers surface using semantic segmentation in deep learning technology.By introducing the Batch Normalization layer and the Convolutional Block Attention Module(CBAM)into the neural network in a reasonable manner,the Intersection over Union(IoU)and Dice coefficient were selected as evaluation indicators for the network.The results show that,in the test set,the intersection over union and Dice coefficient of the 2BNCUnet network are 86.96%and 86.42%,respectively,which are 7.65%and 4.73%higher than those of the U-Net neural network.Additionally,to verify the performance of the network,comparative experiments were conducted using Seg-Net,2BNC-Unet,and U-Net neural networks.The results demonstrate that the segmentation effect of 2BNC-Unet is not only better than that of SegNet and U-Net networks but also capable of completely segmenting the pore details on the cladding layers surface.In the realm of deep learning technology,the segmentation speed and accuracy of 2BNC-Unet have been significantly improved,providing assistance in the performance analysis of cladding layers through pore segmentation.

关 键 词：激光熔覆 语义分割 熔覆层气孔 深度学习 串行注意力机制 

分 类 号：TN249[电子电信—物理电子学] TP302.1[自动化与计算机技术—计算机系统结构]