Scholarlink

<p>This paper explored the application of fully connected neural networks and transfer learning techniques in laser forming, explicitly focusing on predicting scanning paths in multi-stage forming processes. Due to the time-consuming data collection, this research built imaginary data to replace experimental data with training in neural network efficiency. The research developed a machine learning model trained on imaginary datasets that can predict the scanning paths of experimental datasets. However, the test results on experimental data indicated that the imaginary data generation method led to a performance that did not meet expectations. Therefore, this paper also compared the differences between the imaginary and the experimental data. Furthermore, this research adopted transfer learning to address the issues of overfitting and low prediction accuracy encountered by fully connected neural networks at higher forming stages. It explored the fine-tuning and final-layer methods. The results suggest that the fine-tuning method can effectively enhance the model’s generalization capability, achieving better performance in complex forming stages. Compared to the fully connected network, it reduces nearly 35% of training iterations and improves test accuracy by 3–5 mm.</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

Enhanced multi-stage laser forming path prediction through transfer learning and imaginary data validation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年4月21日

查看全文
  • Ping-Hsien Chou,
  • Keiji Yamada,
  • Yean-Ren Hwang,
  • Eisuke Sentoku,
  • Ryutaro Tanaka,
  • Katsuhiko Sekiya

摘要

This paper explored the application of fully connected neural networks and transfer learning techniques in laser forming, explicitly focusing on predicting scanning paths in multi-stage forming processes. Due to the time-consuming data collection, this research built imaginary data to replace experimental data with training in neural network efficiency. The research developed a machine learning model trained on imaginary datasets that can predict the scanning paths of experimental datasets. However, the test results on experimental data indicated that the imaginary data generation method led to a performance that did not meet expectations. Therefore, this paper also compared the differences between the imaginary and the experimental data. Furthermore, this research adopted transfer learning to address the issues of overfitting and low prediction accuracy encountered by fully connected neural networks at higher forming stages. It explored the fine-tuning and final-layer methods. The results suggest that the fine-tuning method can effectively enhance the model’s generalization capability, achieving better performance in complex forming stages. Compared to the fully connected network, it reduces nearly 35% of training iterations and improves test accuracy by 3–5 mm.