三种机器学习非定常气动力建模方法对比分析  

Comparative analysis of three machine learning methods for unsteady aerodynamic modeling

作　　者：李康丽 史志伟[1] 付军泉 童晟翔 张伟麟 LI Kangli;SHI Zhiwei;FU Junquan;TONG Shengxiang;ZHANG Weilin(Key Laboratory of Unsteady Aerodynamics and Flow Control,Ministry of Industry and Information Technology,NUAA,Nanjing 210016,China;School of Aircraft Engineering,Nanchang Hangkong University,Nanchang 330000,China)

机构地区：[1]南京航空航天大学非定常空气动力学与流动控制工业和信息化部重点实验室,江苏南京210016 [2]南昌航空大学飞行器工程学院,江西南昌330000

出　　处：《飞行力学》2023年第3期33-39,共7页Flight Dynamics

摘　　要：针对大迎角下非定常气动力建模问题,以F-16战斗机为研究对象对随机森林方法、深度神经网络方法和物理神经网络方法三种模型的大迎角下非定常气动力建模效果开展研究。首先,介绍了纯数据驱动的随机森林方法和深度神经网络方法两种机器学习方法;然后,在深度神经网络方法的基础上,提出了一种将状态空间模型与神经网络相结合的物理神经网络方法;最后,对三种方法的内/外插性能进行对比。仿真结果表明,三种建模方法均具有很高的内插精度,而物理神经网络建模方法由于在训练过程中增加了物理信息约束,外插性能优于其他两种方法。Aiming at the problem of the unsteady aerodynamic modeling at high angles of attack,the F-16 fighter aircraft is taken as the research object to study the unsteady aerodynamic modeling effect of three models(random forest,deep neural network and physics-informed neural network methods).Firstly,two pure data-driven machine learning methods(random forest and deep neural network methods)are introduced.Then,on the basis of the deep neural network method,a physics-informed neural network method combining the state-space model with the neural network is proposed.Finally,the internal and external performance of the three methods are compared.The simulation results show that all three methods have a high interpolation accuracy,and the physics-informed neural network modeling method has a better extrapolation performance due to the extra physical information constraints during the training process.

关 键 词：非定常气动力模型 随机森林 深度神经网络 物理神经网络 内/外插性能 

分 类 号：V211[航空宇航科学与技术—航空宇航推进理论与工程] TP181[自动化与计算机技术—控制理论与控制工程]