China Mechanical Engineering ›› 2026, Vol. 37 ›› Issue (6): 1486-1496.DOI: 10.3969/j.issn.1004-132X.2026.06.021

Previous Articles    

Springback Prediction in Plate Forming Based on Constitutive Model Embedded Deep Neural Network

ZHANG Long1,2(), LIANG Qiyu1,2(), WEI Xiao3, WAN Chuang2   

  1. 1.Key Laboratory of High Performance Ship Technology,Ministry of Education,Wuhan University of Technology,Wuhan,430063
    2.School of Naval Architecture,Ocean and Energy Power Engineering,Wuhan University of Technology,Wuhan,430063
    3.China Ship Development and Design Center,Wuhan,430064
  • Received:2025-05-27 Online:2026-06-25 Published:2026-07-17
  • Contact: LIANG Qiyu

基于本构嵌入深度学习模型的板材多点柔性成形回弹预测

张龙1,2(), 梁棋钰1,2(), 魏骁3, 万闯2   

  1. 1.武汉理工大学高性能船舶技术教育部重点实验室, 武汉, 430063
    2.武汉理工大学船海与能源动力工程学院, 武汉, 430063
    3.中国舰船研究设计中心, 武汉, 430064
  • 通讯作者: 梁棋钰
  • 作者简介:张龙,男,2000年生,硕士研究生。研究方向为板材成形及回弹预测、深度学习。E-mail: zhanglong202503@163.com
    梁棋钰*(通信作者),女,1993年生,副教授、博士。研究方向为塑性成形、深度学习。授权发明专利4项。发表论文10余篇。E-mail: qyliang@whut.edu.cn
  • 基金资助:
    国家自然科学基金(52201378)

Abstract:

Several sheet metal forming experiments were conducted on the ship three-dimensional computational numerical control bending machine, and corresponding numerical method was validated by these experimental results. Considering actual production needs and processing accuracy, the Latin hypercube sampling method was employed to generate sampling points by intervals for model training, with the corresponding dataset obtained through numerical simulations and experimental validations. A constitutive equation-embedded deep learning neural network(CE-DNN) model was developed by optimizing the visual geometry group(VGG) network architecture through the integration of material-informed convolutional layers, thereby establishing a multi-parameter coupled learning system that synthesized material properties, thickness, and geometric features. The performance of the proposed model under data-constrained scenarios was quantitatively evaluated through data extrapolation and training set reduction strategies. Results demonstrate that the proposed model exhibites certain robustness when the training data for material, thickness, and shape parameters are reduced, while maintaining generalization capability in springback extrapolation prediction.

Key words: sheet metal forming, springback prediction, deep learning, constitutive relationship

摘要:

针对船舶三维数控弯板机这一船体外板成形设备,开展了板材回弹实验及仿真方法验证。考虑实际生产需求和加工精度,采用拉丁超立方采样方法间隔生成用于模型训练的工况数据,并采用验证后的仿真方法构建了模型训练集及预测集。提出了一种基于本构嵌入的深度学习模型(CE-DNN),在视觉几何组(VGG)网络中嵌入材料卷积层,构建融合材料参数、厚度及形状特征的多参数耦合学习框架。采用数据外推和训练集缩减策略量化评估该模型在数据受限场景下的性能,结果表明,该模型在材料、厚度和形状参数的训练数据量减少情况下具有一定的鲁棒性,同时在回弹数据外推预测方面具有一定的泛化性能。

关键词: 板材成形, 回弹预测, 深度学习, 本构关系

CLC Number: