航空学报 > 2025, Vol. 46 Issue (5): 531450-531450   doi: 10.7527/S1000-6893.2024.31450

固体力学与飞行器总体设计

动载荷识别物理嵌入式神经网络模型与方法

杨智春1,2(), 杨特1,2   

  1. 1.西北工业大学 航空学院,西安 710072 2.结构强度与完整性全国重点实验室,西安 710072
    3.飞行器流体物理全国重点实验室,绵阳 621000
  • 收稿日期:2024-10-28 修回日期:2024-11-07 接受日期:2024-11-27 出版日期:2024-12-17 发布日期:2024-12-10
  • 通讯作者: 杨智春 E-mail:yangzc@nwpu.edu.cn
  • 基金资助:
    航空科学基金(20220015053002);飞行器流体物理全国重点实验室青苗基金(2024-APF-KFQMJJ-10)

Physical embedded neural network model and method for dynamic load identification

Zhichun YANG1,2(), Te YANG1,2   

  1. 1.School of Aeronautics,Northwestern Polytechnical University,Xi’an 710072,China
    2.National Key Laboratory of Strength and Structural Integrity,Xi’an 710072,China
    3.National Key Laboratory of Aerospace Physics in Fluids,Mianyang 621000,China
  • Received:2024-10-28 Revised:2024-11-07 Accepted:2024-11-27 Online:2024-12-17 Published:2024-12-10
  • Contact: Zhichun YANG E-mail:yangzc@nwpu.edu.cn
  • Supported by:
    Aeronautical Science Foundation of China(20220015053002);Research Fund of National Key Laboratory of Aerospace Physics in Fluids(2024-APF-KFQMJJ-10)

摘要:

针对传统动载荷识别方法中频响函数矩阵求逆运算导致的不适定性问题,以及深度学习方法缺乏物理可解释性的局限,提出了一种全新的物理嵌入式神经网络(PENN)动载荷识别模型与方法。通过将结构动力学参数(如模态质量、模态刚度、模态阻尼等)直接嵌入神经网络中,构建出具有物理可解释性的PENN动载荷识别模型。PENN模型能够以正向计算过程直接识别动载荷的功率谱密度,避免了传统方法的频响函数矩阵求逆运算,并能够对其内部的物理参数进行自适应修正,保证了在先验物理参数不准确时仍能实现动载荷的高精度识别。详细阐述了方法机理、PENN模型构建规则、参数设定及训练流程,并对多种工况下的动载荷进行了数值仿真与实验验证,结果表明,本方法在动力学系统先验参数不准确和仅有1组训练样本的情况下,识别动载荷的皮尔逊相关系数均不低于95%,展现出较好的鲁棒性和工程应用潜力。

关键词: 动载荷识别, 物理嵌入式神经网络, 可解释神经网络, 模态分析, 参数校正, 凸优化

Abstract:

To address the ill-posedness caused by the inversion of the frequency response function matrix in traditional dynamic load identification methods, as well as the lack of physical interpretability in deep learning methods, a novel Physical Embedded Neural Network (PENN) for dynamic load identification is proposed. By embedding structural dynamic parameters, such as modal mass, modal stiffness, and modal damping, directly into the neural network, the PENN model is constructed to offer physical interpretability. The PENN model can directly identify the power spectral density of dynamic loads through a forward computational process, avoiding the need for inverting the frequency response function matrix as in traditional methods. Additionally, this model can adaptively adjust internal physical parameters, ensuring high-precision identification of dynamic loads even when prior physical parameters are inaccurate. The paper provides a detailed explanation of the method’s mechanism, the construction rules of the PENN model, parameter settings, and the training process. Numerical simulations and experimental validations were conducted under various conditions The results show that even when the prior dynamic system parameters are inaccurate and only one training sample is available, the Pearson correlation coefficient of dynamic load identification was consistently above 95%, demonstrating strong robustness and potential for engineering applications.

Key words: dynamic load identification, physical embedded neural network, interpretable neural network, modal analysis, parameter correction, convex optimization

中图分类号: