Scholarlink

<p>Real-time monitoring and defect detection in Gas Metal Arc Welding can significantly reduce post-welding repairs, leading to fewer production delays and increasing customer satisfaction. This study proposes real-time deep learning-based uni-modal and multi-modal defect detection. Welding images and sound data were collected from an industrial collaborative welding robot in R &amp; D setting to detect common defects including lack of penetration, lack of fusion, undercut, cold lap, and porosity in fillet joints. Our proposed multi-modal models leverage complementary information from both modalities and improve the defect detection performance by up to 58.62% compared to uni-modal models. Specifically, our proposed multi-modal model achieves a F1 score of 0.88 to detect porosity. An improvement in F1 score in multi-modal defect detection indicates a higher accurate defect detection rate and fewer false alarms. This enhancement enables early-stage defect identification, allowing timely corrective interventions and reducing the need for extensive welding corrections or replacing the welded item. Additionally, fewer false alarms prevent unnecessary corrective actions, saving both time and resources in production workflows.</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

Multi-modal data improves real-time defect classification deep learning models for fillet joints in gas metal arc welding

  • 期刊论文

  • 发表日期：

    2025年6月5日

查看全文
  • Mobina Mobaraki,
  • Soodeh Ahani,
  • Ringo Gonzalez,
  • Klaske van Heusden,
  • Guy A. Dumont

摘要

Real-time monitoring and defect detection in Gas Metal Arc Welding can significantly reduce post-welding repairs, leading to fewer production delays and increasing customer satisfaction. This study proposes real-time deep learning-based uni-modal and multi-modal defect detection. Welding images and sound data were collected from an industrial collaborative welding robot in R & D setting to detect common defects including lack of penetration, lack of fusion, undercut, cold lap, and porosity in fillet joints. Our proposed multi-modal models leverage complementary information from both modalities and improve the defect detection performance by up to 58.62% compared to uni-modal models. Specifically, our proposed multi-modal model achieves a F1 score of 0.88 to detect porosity. An improvement in F1 score in multi-modal defect detection indicates a higher accurate defect detection rate and fewer false alarms. This enhancement enables early-stage defect identification, allowing timely corrective interventions and reducing the need for extensive welding corrections or replacing the welded item. Additionally, fewer false alarms prevent unnecessary corrective actions, saving both time and resources in production workflows.