Scholarlink

<p>Assembly inaccuracies and thermal distortions during welding highlight the need for reliable, real-time monitoring of penetration depth. Although deep learning methods offer promising capabilities, the scarcity of labeled training data hinders their deployment in industrial environments. To address these limitations, this work proposes a novel physics-informed neural network architecture designed for real-time prediction of weld penetration depth using limited datasets. The model leverages molten pool image sequences to quantitatively infer variations in several thermal-related variables, including penetration depth variation, effective thermal diffusivity, thermal conductivity, laser energy absorptivity, and heat source distribution radius. These inferred parameters are incorporated into an analytical heat transfer model to maintain essential physical consistency, thereby enabling accurate predictions of penetration depth. The hybrid architecture extracts spatial features and models temporal dynamics, explicitly representing keyhole evolution while implicitly capturing molten pool fluid behavior. Physics-aware activation functions constrain thermal-related parameters within physically meaningful ranges, thereby stabilizing convergence and improving training efficiency. Experimental evaluations demonstrate a mean absolute error of no more than 0.1136&#xa0;mm, even when the proposed model is trained on only 10% of the dataset. Moreover, the model reliably predicts extreme, previously unseen gap conditions, outperforming other baseline methods. With only 0.55&#xa0;M parameters and less than 6.67&#xa0;ms inference latency, the model achieves superior accuracy, enabling robust real-time deployment in industrial welding applications. The model implementation is available at <a href="https://github.com/talosfabs/PI-CNN-LSTM-LBW">https://github.com/talosfabs/PI-CNN-LSTM-LBW</a>.</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 penetration depth prediction via physics-informed learning from molten pool surface morphology in laser filler wire welding

  • 期刊论文

  • 发表日期：

    2025年12月5日

查看全文
  • Rundong Lu,
  • Ming Lou,
  • Yujun Xia,
  • Yongbing Li

摘要

Assembly inaccuracies and thermal distortions during welding highlight the need for reliable, real-time monitoring of penetration depth. Although deep learning methods offer promising capabilities, the scarcity of labeled training data hinders their deployment in industrial environments. To address these limitations, this work proposes a novel physics-informed neural network architecture designed for real-time prediction of weld penetration depth using limited datasets. The model leverages molten pool image sequences to quantitatively infer variations in several thermal-related variables, including penetration depth variation, effective thermal diffusivity, thermal conductivity, laser energy absorptivity, and heat source distribution radius. These inferred parameters are incorporated into an analytical heat transfer model to maintain essential physical consistency, thereby enabling accurate predictions of penetration depth. The hybrid architecture extracts spatial features and models temporal dynamics, explicitly representing keyhole evolution while implicitly capturing molten pool fluid behavior. Physics-aware activation functions constrain thermal-related parameters within physically meaningful ranges, thereby stabilizing convergence and improving training efficiency. Experimental evaluations demonstrate a mean absolute error of no more than 0.1136 mm, even when the proposed model is trained on only 10% of the dataset. Moreover, the model reliably predicts extreme, previously unseen gap conditions, outperforming other baseline methods. With only 0.55 M parameters and less than 6.67 ms inference latency, the model achieves superior accuracy, enabling robust real-time deployment in industrial welding applications. The model implementation is available at https://github.com/talosfabs/PI-CNN-LSTM-LBW.