航空学报 > 2022, Vol. 43 Issue (S1): 726928-726928   doi: 10.7527/S1000-6893.2022.26928

基于复合卷积层神经网络结构的焊缝缺陷分类技术

刘欢, 刘骁佳, 王宇斐, 王宁, 曹立俊   

  1. 1. 上海航天精密机械研究所,上海 201600
  • 收稿日期:2022-01-10 修回日期:2022-01-25 发布日期:2022-02-28
  • 通讯作者: 刘骁佳,E-mail:lxj9039@126.com E-mail:lxj9039@126.com
  • 基金资助:
    上海市浦江人才计划(20PJ1405000)

Weld defect classification technology based on compound convolution neural network structure

LIU Huan, LIU Xiaojia, WANG Yufei, WANG Ning, CAO Lijun   

  1. 1. Shanghai Aerospace Precision Machinery Research Institute, Shanghai 201600, China
  • Received:2022-01-10 Revised:2022-01-25 Published:2022-02-28
  • Supported by:
    Shanghai Pujiang Program (20PJ1405000)

摘要: X射线检测方法被广泛应用于焊缝内部缺陷的检测过程,但检测过程主要依靠人工查看X射线图像对产品质量做出评级,此种检测方式效率低且需要操作人员丰富的检测经验,难以满足批量检测需求,亟需借助卷积神经网络的方法提升其检测效率。因此提出一种复合卷积层网络结构的CC-ResNet模型,将不同尺度的卷积核复合为一层卷积,替换单个卷积核结构,并对模型训练过程中的损失函数进行优化。使用车间产品焊缝X射线图像建立数据集,通过对CC-ResNet网络模型进行实验可得,其平均召回率和平均准确率为98.52%和95.23%,与其他网络模型相比提高了缺陷检测能力,为车间进行焊缝缺陷的智能化和批量化检测提供了算法基础。

关键词: 焊缝, X射线, 缺陷检测, ResNet, 复合卷积

Abstract: The X-ray detection method is widely used in the detection of weld internal defects, but in the detection process, product quality is rated mainly by manually viewing X-ray images. This detection method is inefficient, and requires rich experience of operators, which is difficult to meet the needs of batch detection. It is thus urgent to improve the detection efficiency with the help of the convolution neural network. This paper proposes a CC-ResNet model with composite convolution layer network structure, which combines convolution cores of different scales into one layer of convolution, replaces the original single convolution core structure, and optimizes the loss function in the process of model training. A data set is established by using the weld X-ray images of workshop products. Experiment on CC-ResNet network model shows that the average recall rate and accuracy are 98.52% and 95.23%, respectively. Compared with other network models, the model proposed can improve the defect detection ability, providing an algorithm basis for intelligent and batch detection of weld defects in the workshop.

Key words: weld, X-ray, defect detection, ResNet, compound convolution

中图分类号: