欢迎访问《兵工学报》官方网站,今天是

兵工学报 ›› 2024, Vol. 45 ›› Issue (7): 2197-2208.doi: 10.12382/bgxb.2023.0127

• • 上一篇    下一篇

结合系统辨识和迁移学习的高速旋转弹气动力建模方法

季稳1,2, 李春娜1,2,*(), 贾续毅1,2, 王刚3, 龚春林1,2   

  1. 1 西北工业大学 航天学院, 陕西 西安 710072
    2 陕西省空天飞行器设计重点实验室, 陕西 西安 710072
    3 西北工业大学 航空学院, 陕西 西安 710072
  • 收稿日期:2023-02-26 上线日期:2023-07-13
  • 通讯作者:
  • 基金资助:
    国家自然科学基金项目(U2141254)

A High-spinning Projectile Aerodynamic Modeling Method Combining System Identification and Transfer Learning

JI Wen1,2, LI Chunna1,2,*(), JIA Xuyi1,2, WANG Gang3, GONG Chunlin1,2   

  1. 1 School of Astronautics, Northwestern Polytechnical University, Xi’an 710072, Shaanxi, China
    2 Shaanxi Aerospace Flight Vehicle Design Key Laboratory, Xi’an 710072, Shaanxi, China
    3 School of Aeronautics, Northwestern Polytechnical University, Xi’an 710072, Shaanxi, China
  • Received:2023-02-26 Online:2023-07-13

摘要:

计算流体动力学与刚体动力学(Computational Fluid Dynamics and Rigid Body Dynamics,CFD/RBD)耦合仿真是旋转弹飞行性能评估的常用方法之一,但由于需要进行大量CFD计算,该方法效率较低。建立一个高效、精确且泛化能力强的气动力模型并以之替代耦合仿真中的CFD模块,可以大幅度提升仿真效率。针对前述旋转弹气动力建模问题,提出一种结合系统辨识和迁移学习的建模方法。给定旋转弹运动初始条件并采用CFD/RBD耦合仿真获得样本,采用自回归滑动平均方法建立原始气动力模型,同时采用长短时记忆网络建立状态预测模型。当初始条件变化不大时,原始气动力模型仍然适用;当初始条件发生较大改变时,利用迁移学习将状态预测模型迁移到该初始条件下,并预测相应初始条件下的状态参数,基于预测得到的状态参数,采用自回归滑动平均方法建立气动力模型。算例结果表明:所提方法适用于初始转速和俯仰角变化较大时对旋转弹气动力的精确建模;与直接以CFD/RBD耦合仿真结果为样本、采用自回归滑动平均方法建模相比,在精度相同时建模时间缩短了一半。

关键词: 高速旋转弹, 气动力建模, 自回归滑动平均, 长短时记忆网络, 迁移学习

Abstract:

Computational fluid dynamics andrigid body dynamics(CFD/RBD) coupling simulation is a common method for evaluating the flight performance of spinning projectiles, which is quite inefficient due to huge amount of CFD simulations. An efficient and accurate aerodynamic model with strong generalization ability is established to significantly improve the simulation efficiency by replacing the CFD module in coupling simulation. An aerodynamic modeling method combining system identification and transfer learning is proposed to address the above-mentioned problem. First, the samples of spinning projectile are obtained by CFD/RBD coupling simulations under the given initial conditions. Then the autoregressive moving average method is used to build an original aerodynamic model, and the long short-term memory network is utilized to build a state prediction model. In condition of small variation of initial state, the original aerodynamic model remains valid; however in case of large variation of initial state, the state prediction model is transferred to the corresponding initial state, and then an aerodynamic model is built by using the autoregressive moving average method based on the predicted state parameters. The results show that the proposed method is suitable for accurate aerodynamic modeling of high-spinning projectile under the large variations of initial angular velocity and pitch angle. In comparison with the autoregressive moving average method based on direct CFD/RBD coupling simulations, the modeling efficiency of the proposed method is doubled.

Key words: high-spinning projectile, aerodynamic modeling, autoregressive moving average, long short-term memory network, transfer learning

中图分类号: