Scholarlink

<p>Online monitoring is crucial in laser beam welding (LBW) to ensure product quality. Despite the recent widespread application of deep learning in visual quality monitoring, most methods rely on single-vision-task (SVT) models, which may inadequately depict the welding process in certain scenarios. In this work, we develop an innovative vision-based approach that employs a multi-vision-task convolutional neural network (CNN) to monitor the LBW process of liquid rocket engine nozzles. We constructed a multi-label image dataset from experimental LBW images for CNN training and testing. To comprehensively depict the LBW process, a novel welding-segmentation-recognition network (WSRNet) was proposed. WSRNet adopts an architecture of a shared backbone and two branches for simultaneous image segmentation and recognition. It achieved outstanding performance with an Intersection over Union of 89.61% and an accuracy of 93.64% in weld pool zone segmentation and welding defect recognition tasks, respectively. Moreover, the inference speed reached 171 frames per second. These metrics are comparable or superior to those of typical SVT CNNs in welding. The effectiveness of the proposed monitoring method is validated via online monitoring experiments.</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

An online monitoring approach for laser beam welding process based on multi-vision-task learning

  • 期刊论文

  • 发表日期：

    2025年12月9日

查看全文
  • Yifeng Zhou,
  • Boce Xue,
  • Keming Guan,
  • Haijin Yue,
  • Hefei Zou,
  • Baohua Chang,
  • Dong Du

摘要

Online monitoring is crucial in laser beam welding (LBW) to ensure product quality. Despite the recent widespread application of deep learning in visual quality monitoring, most methods rely on single-vision-task (SVT) models, which may inadequately depict the welding process in certain scenarios. In this work, we develop an innovative vision-based approach that employs a multi-vision-task convolutional neural network (CNN) to monitor the LBW process of liquid rocket engine nozzles. We constructed a multi-label image dataset from experimental LBW images for CNN training and testing. To comprehensively depict the LBW process, a novel welding-segmentation-recognition network (WSRNet) was proposed. WSRNet adopts an architecture of a shared backbone and two branches for simultaneous image segmentation and recognition. It achieved outstanding performance with an Intersection over Union of 89.61% and an accuracy of 93.64% in weld pool zone segmentation and welding defect recognition tasks, respectively. Moreover, the inference speed reached 171 frames per second. These metrics are comparable or superior to those of typical SVT CNNs in welding. The effectiveness of the proposed monitoring method is validated via online monitoring experiments.