China Mechanical Engineering ›› 2026, Vol. 37 ›› Issue (5): 1199-1209.DOI: 10.3969/j.issn.1004-132X.2026.05.020

Previous Articles    

Physics-guided Neural Network Model for Predicting Rolling Forces in Aluminum Strip Cold Rolling

ZHANG Ji1(), YUAN Haibo1, WANG Zhixuan1, ZHU Sihua1, BAI Zhenhua1,2()   

  1. 1.National Engineering Research Center for Equipment and Technology of Cold Strip Rolling,Yanshan University,Qinhuangdao,Hebei,066004
    2.Shenzhen Research Institute of Yanshan University,Shenzhen,Guangdong,518000
  • Received:2025-05-31 Online:2026-05-25 Published:2026-06-09
  • Contact: BAI Zhenhua

物理引导的神经网络模型预测铝带冷轧轧制力

张冀1(), 袁海博1, 王智璇1, 朱思华1, 白振华1,2()   

  1. 1.燕山大学国家冷轧板带装备及工艺工程技术研究中心, 秦皇岛, 066004
    2.燕山大学深圳研究院, 深圳, 518000
  • 通讯作者: 白振华
  • 作者简介:张冀,男,1999年生,博士研究生。研究方向为钢铁智能制造工艺。E-mail:1332011776@qq.com
    白振华*(通信作者),男,1975年生,教授、博士研究生导师。研究方向为轧钢设备及工艺。E-mail:bai_zhenhua@aliyun.com
  • 基金资助:
    河北省自然科学基金(E2024203125);河北省省级科技计划创新应用场景专项(252Q0303D);河北省省级科技计划数字产业创新发展专项(252F0303D);唐山市科技计划(24140203C)

Abstract:

Accurate prediction of rolling forces in aluminum strip cold rolling was essential for process optimization and ensuring product quality. Traditional physical models often struggled to accurately capture the complex interactions among various factors, while purely data-driven models typically lacked physical constraints and exhibited limited generalization ability. In order to develop a rolling force prediction model for aluminum strip cold rolling that combined both physical interpretability and high prediction accuracy, taking 5182 aluminum strip as the research object, a yield strength model was fitted. After preprocessing a large volume of real production data, the traditional Hill rolling force model was optimized by using the PSO-Nelder-Mead algorithm, which significantly improved the predictive capability. Then, a PGNN model was proposed. This model was based on a convolutional neural network, integrating a physical constraint term derived from the optimized mathematical model into the loss function,which enabled the model to follow the physical laws while simultaneously being data-driven. Bayesian optimization was employed to optimize the model's hyperparameters. Finally, the established model was verified by using actual production data. The results show that the PGNN model exhibits excellent prediction performance on the validation set. The prediction accuracy is significantly higher than that of the mathematical models both before and after optimization, and shows strong generalization ability. Furthermore, the iterative analysis of the loss function further confirms the effectiveness of the physical constraints.

Key words: aluminum strip cold rolling, rolling force prediction, physics-guided neural network(PGNN), data-driven, physical constraint

摘要:

铝带冷轧过程中轧制力的精确预测是优化工艺和保障产品质量的关键,传统的物理模型难以精确捕捉复杂因素间的相互作用,而纯数据驱动模型缺乏物理约束且泛化能力受限。为构建一种兼具物理可解释性和高预测精度的铝带冷轧轧制力预测模型,以5182铝带为研究对象并拟合了屈服强度模型,对大量实际生产数据进行预处理后,采用PSO-Nelder-Mead算法对传统Hill轧制力模型进行了优化,提高了该模型的预测能力。在此基础上,提出了一种物理引导的神经网络(PGNN)模型。该模型以卷积神经网络为基础,在损失函数中融入了基于优化数学模型的物理约束项,使得模型在数据驱动的同时遵循物理规律,并使用贝叶斯优化对模型超参数进行寻优。最后将所建模型使用实际生产数据进行验证。研究结果表明,PGNN模型在验证集上展现出良好的预测性能,预测精度显著高于优化前后的数学模型且泛化能力良好,损失函数的迭代分析也辅证了物理约束的有效性。

关键词: 铝带冷轧, 轧制力预测, 物理引导神经网络, 数据驱动, 物理约束

CLC Number: