Scholarlink

<p>Keyhole Tungsten Inert Gas (K-TIG) welding has become more and more common due to its advantages for medium and thick plate welding. The segmentation of the weld pool and keyhole in K-TIG weld pool monitoring poses significant challenges due to the need for accuracy and real-time performance. We propose an enhanced PSPNet-based real-time weld pool segmentation technique to monitor weld pool fluctuations in order to overcome these difficulties. Initially, MobileNetV2 was selected as the backbone network in order to increase detection speed. The ParNet and Criss-Cross Attention modules were then added to improve detection accuracy. Finally, Dice loss was used to reduce the difference between positive and negative samples. The enhanced model obtained a mean Intersection over Union of 94.75% on the validation set. It successfully achieved a balance between accuracy and processing speed, achieving a real-time performance of 80.36 FPS, representing a 106.4% improvement in real-time performance over the original PSPNet. We used this model to extract data like area and aspect ratio, which allowed us to monitor and analyze weld pool changes in real time. This approach enhances welding quality monitoring and provides a strong basis for introducing closed-loop control in K-TIG welding.</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

Real-time monitoring of keyhole tungsten inert gas weld pool feature variation using an improved PSPNet

  • 期刊论文

  • 发表日期：

    2025年6月3日

查看全文
  • Shuwan Cui,
  • Junhao Yang,
  • Hao Li,
  • Xiaozhen Liu,
  • Xuan Zhou

摘要

Keyhole Tungsten Inert Gas (K-TIG) welding has become more and more common due to its advantages for medium and thick plate welding. The segmentation of the weld pool and keyhole in K-TIG weld pool monitoring poses significant challenges due to the need for accuracy and real-time performance. We propose an enhanced PSPNet-based real-time weld pool segmentation technique to monitor weld pool fluctuations in order to overcome these difficulties. Initially, MobileNetV2 was selected as the backbone network in order to increase detection speed. The ParNet and Criss-Cross Attention modules were then added to improve detection accuracy. Finally, Dice loss was used to reduce the difference between positive and negative samples. The enhanced model obtained a mean Intersection over Union of 94.75% on the validation set. It successfully achieved a balance between accuracy and processing speed, achieving a real-time performance of 80.36 FPS, representing a 106.4% improvement in real-time performance over the original PSPNet. We used this model to extract data like area and aspect ratio, which allowed us to monitor and analyze weld pool changes in real time. This approach enhances welding quality monitoring and provides a strong basis for introducing closed-loop control in K-TIG welding.