基于物理信息神经网络的飞机气动参数辨识方法  

A physics informed neural network based method for aircraft aerodynamic parameter identification

作　　者：付军泉 钟伯文 钟运琴 刘赟[1,2] FU Junquan;ZHONG Bowen;ZHONG Yunqin;LIU Yun(School of Aircraft Engineering,Nanchang Hangkong University,Nanchang 330063,China;Jiangxi Key Laboratory for Aircraft Design and Aerodynamic Simulation,Nanchang 330063,China;Key Laboratory of Big Data Mining and Knowledge Management,University of Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]南昌航空大学飞行器工程学院,南昌330063 [2]江西省飞行器设计与气动仿真重点实验室,南昌330063 [3]中国科学院大学中国科学院大数据挖掘与知识管理重点实验室,北京100190

出　　处：《空气动力学学报》2023年第9期30-37,共8页Acta Aerodynamica Sinica

基　　金：国家自然科学基金(12262023);江西省重点研发计划项目(20223BBE51001)。

摘　　要：在飞机设计与研制过程中,通过气动参数辨识建立可靠的飞行动力学模型非常重要。传统的气动参数辨识工程算法,诸如极大似然法,需要给出合理的飞行动力学模型以及待辨识参数的初值。基于传统神经网络的气动参数辨识可以避免飞行动力学建模过程,这种方法需要通过增量法、导数法间接地从神经网络提取气动参数。本文提出了一种基于物理信息神经网络的飞机气动参数辨识方法,可将含待辨识参数的飞行动力学模型作为正则项加入损失函数,直接辨识得到气动参数。该方法可以显著减少建模数据需求,也能提高建模精度。飞行仿真数据验证结果表明,该方法的无噪声、含2%噪声仿真数据,纵向飞行状态空间模型辨识最大相对误差分别为1.80%、4.64%,表明了基于物理信息神经网络的飞机气动参数辨识方法具有可行性,并对含噪声的飞行数据具有泛化性。It is of great importance to establish a reliable flight dynamic model based on aerodynamic parameter identification in the process of aircraft design.Traditional engineering methods for aerodynamic parameter identification,such as the maximum likelihood method etc.,require a reasonable dynamic model and initial values for the parameters to be identified.The identification of aerodynamic parameters based on traditional neural networks has no requirement for the flight dynamic model and parameter initialization.In this study,an aerodynamic parameter identification method based on physics informed neural network(PINN)is proposed,which can directly identify aerodynamic parameters by incorporating the flight dynamic model as a regularization term into the loss function.This method can significantly reduce the required training data and improve the modeling accuracy.Verification results for the flight simulation data show that the maximum relative errors of the PINN based method on the longitudinal flight state space model identification with no noise and 2%noise are 1.80%and 4.64%,respectively.It suggests that the PINN based aerodynamic parameter identification method is feasible and has the generalization capacity to noisy flight data.

关 键 词：深度学习 基于物理信息的神经网络 飞行动力学 状态空间 气动参数辨识 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] V212.1[自动化与计算机技术—控制科学与工程]