Scholarlink

<p>The occurrence of magnetic arc blow during the arc welding process significantly affects weld quality. To achieve automatic real-time recognition of magnetic arc blow, this paper proposes an extraction and quantification model (EQM) based on YOLOv8n-improvement, an improved version of the YOLOv8n deep learning model. Firstly, the FasterNetBlock and GhostConv modules are introduced to lighten the C2f module of YOLOv8n. The improved version is utilized to extract the coordinate information of the arc and tip of tungsten electrode, which maintains a mean average precision at IoU of 50% (MAP50) of 0.995 while reducing model parameters by 23.6%, decreasing floating point operations (FLOPs) by 12.5%, shrinking the model size by 22.1%. Subsequently, this paper introduces a method for representing arc energy density based on grayscale value and designs a quantitative index for the degree of magnetic arc blow (ABI), combining the coordinate information of the arc and tip of tungsten electrode. Experimental results demonstrate that the model can produce monotonic quantitative results for arcs of different shapes. Additionally, when deployed on the real-time recognition end, the model attains a processing rate of 68.2 frames per second (FPS) and maintains excellent recognition performance for arc forms that are not included in the dataset, demonstrating good generalization capabilities.</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

A magnetic arc blow extraction and quantification model based on YOLOv8n-improvement

  • 期刊论文

  • 发表日期：

    2025年7月23日

查看全文
  • Zhenmin Wang,
  • Baojun Zuo,
  • Danhuan Zhou,
  • Ying Dong,
  • Weiji Su,
  • Haipeng Liao,
  • Qin Zhang

摘要

The occurrence of magnetic arc blow during the arc welding process significantly affects weld quality. To achieve automatic real-time recognition of magnetic arc blow, this paper proposes an extraction and quantification model (EQM) based on YOLOv8n-improvement, an improved version of the YOLOv8n deep learning model. Firstly, the FasterNetBlock and GhostConv modules are introduced to lighten the C2f module of YOLOv8n. The improved version is utilized to extract the coordinate information of the arc and tip of tungsten electrode, which maintains a mean average precision at IoU of 50% (MAP50) of 0.995 while reducing model parameters by 23.6%, decreasing floating point operations (FLOPs) by 12.5%, shrinking the model size by 22.1%. Subsequently, this paper introduces a method for representing arc energy density based on grayscale value and designs a quantitative index for the degree of magnetic arc blow (ABI), combining the coordinate information of the arc and tip of tungsten electrode. Experimental results demonstrate that the model can produce monotonic quantitative results for arcs of different shapes. Additionally, when deployed on the real-time recognition end, the model attains a processing rate of 68.2 frames per second (FPS) and maintains excellent recognition performance for arc forms that are not included in the dataset, demonstrating good generalization capabilities.