Scholarlink

<p>Weld expulsion is one of the most common welding defects during the resistance spot welding (RSW) process. It is desired that&#xa0;the expulsion intensity&#xa0;to be inspected online via in-process sensing signals and machine learning methods so as to control and eventually eliminate weld expulsion in production. However, conventional machine learning methods struggle with out-of-distribution (OOD) data. Their performance would significantly deteriorate when there is a deviation between the distribution of test data and training data. In this study, by incorporating a specially designed autoencoder and physical constraints, a new approach using physics-informed neural networks (PINN) successfully integrates domain knowledge from welding physics to enhance the generalization performance. The results showed that the new method exhibits improved generalization capability to OOD data, allowing accurate prediction of weld expulsion intensity even under abnormal welding conditions such as electrode wear. Compared to traditional methods, the new approach achieves a 60% increase in accuracy, making it suitable for addressing the issue of lacking labeled data and uncertainty disturbances of welding conditions in mass production. This study provides new ideas for the application of PINN in monitoring and control of the welding process.</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

Improving out-of-distribution generalization for online weld expulsion inspection using physics-informed neural networks

  • 期刊论文

  • 发表日期：

    2025年2月9日

查看全文
  • Yu-Jun Xia,
  • Qiang Song,
  • BenGang Yi,
  • TianLe Lyu,
  • ZhiQiang Sun,
  • YongBing Li

摘要

Weld expulsion is one of the most common welding defects during the resistance spot welding (RSW) process. It is desired that the expulsion intensity to be inspected online via in-process sensing signals and machine learning methods so as to control and eventually eliminate weld expulsion in production. However, conventional machine learning methods struggle with out-of-distribution (OOD) data. Their performance would significantly deteriorate when there is a deviation between the distribution of test data and training data. In this study, by incorporating a specially designed autoencoder and physical constraints, a new approach using physics-informed neural networks (PINN) successfully integrates domain knowledge from welding physics to enhance the generalization performance. The results showed that the new method exhibits improved generalization capability to OOD data, allowing accurate prediction of weld expulsion intensity even under abnormal welding conditions such as electrode wear. Compared to traditional methods, the new approach achieves a 60% increase in accuracy, making it suitable for addressing the issue of lacking labeled data and uncertainty disturbances of welding conditions in mass production. This study provides new ideas for the application of PINN in monitoring and control of the welding process.