Scholarlink

<p>This study presents a methodology based on deep learning aimed at the automation of weld defect identification in non-temporally constrained underwater welds on AH36 steel. Defect detection is framed as a binary image classification task where defective and non-defective images are classified, and lack of fusion (LOF) is considered a single category under the defective class. The methodology applied two convolutional neural network (CNN) architectures, residual neural network (ResNet-18) and visual geometry group (VGG-16), by utilizing experimental data. Experimental evaluations suggest that the best performance in terms of defect detection accuracy and robustness represents ResNet-18, with the vision of VGG-16 representation spent within challenging low-contrast underwater conditions. This superiority is attributed to residual connections in ResNet-18, which facilitate deeper feature propagation and mitigate vanishing gradient issues. On the other hand, VGG-16’s deeper, more parameter-heavy architecture requires longer training times and overfits on the limited dataset.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CNN-Based Weld Defect Detection in Underwater Welding of AH36 Steel

  • 期刊论文

  • 发表日期：

    2026年2月13日

查看全文
  • Ganta Venkateswara Rao,
  • Saurav Suman

摘要

This study presents a methodology based on deep learning aimed at the automation of weld defect identification in non-temporally constrained underwater welds on AH36 steel. Defect detection is framed as a binary image classification task where defective and non-defective images are classified, and lack of fusion (LOF) is considered a single category under the defective class. The methodology applied two convolutional neural network (CNN) architectures, residual neural network (ResNet-18) and visual geometry group (VGG-16), by utilizing experimental data. Experimental evaluations suggest that the best performance in terms of defect detection accuracy and robustness represents ResNet-18, with the vision of VGG-16 representation spent within challenging low-contrast underwater conditions. This superiority is attributed to residual connections in ResNet-18, which facilitate deeper feature propagation and mitigate vanishing gradient issues. On the other hand, VGG-16’s deeper, more parameter-heavy architecture requires longer training times and overfits on the limited dataset.