条件生成对抗网络的翼型反设计方法  

An Inverse Design Method of Airfoil Using Conditional Generative Adversarial Network

作　　者：吴明雨 陈志华[1] 邱志明 吴威涛 WU Mingyu;CHEN Zhihua;QIU Zhiming;WU Weitao(Key Laboratory of Transient Physics,Nanjing University of Science and Technology,Nanjing 210094,China;Naval Research Academy,Beijing 102442,China;School of Mechanical Engineering,Nanjing University of Science and Technology,Nanjing 210094,China)

机构地区：[1]南京理工大学瞬态物理重点实验室,南京210094 [2]海军研究院,北京102442 [3]南京理工大学机械工程学院,南京210094

出　　处：《宇航学报》2023年第10期1512-1521,共10页Journal of Astronautics

基　　金：江苏省自然科学基金(BK20201302)。

摘　　要：针对变体飞行器实时控制翼型形状的需求,提出了基于深度学习的翼型反设计方法,利用多层感知机搭建了由生成器与判别器组成的条件生成对抗网络。生成器从带有随机噪声的气动参数中提取内在特征,习得特征到翼型的映射关系;判别器则将生成器产生的翼型或真实翼型与前述气动参数混合作为输入,输出该翼型为符合指定气动条件的真实翼型的概率。为了优化网络模型,研究并分析了噪声尺寸、超参数及网络结构对模型收敛性能的影响。训练好的网络模型即可根据给定的期望气动参数,快速生成配套的翼型。测试结果表明预测翼型与真实翼型的均方根误差的平均值为0.17%,耗时仅为23 ms,大大提高了设计精度与效率;并且在有噪声干扰情况下依旧保持良好的设计性能,增强了翼型设计模型的鲁棒性。研究成果可以应用于变体飞行器自适应在线最优气动构型控制。Aiming at the real-time control of airfoil shape for morphing vehicle,a deep learning-based airfoil inverse design method is proposed,which utilizes multilayer perceptron to build a conditional generative adversarial network consisting of a generator and a discriminator.The generator extracts intrinsic features from aerodynamic parameters with random noise,and learns the mapping between the features and the airfoil shape.The synthesized or real airfoils are mixed with the aforementioned aerodynamic parameters as inputs to the discriminator,which then outputs the probability that the airfoil is a true airfoil shape that meets the specified aerodynamic conditions.To optimize the network model,the effects of noise size,model hyperparameters and network structure on the convergence performance of the model are also investigated and analysed.The trained network model can rapidly generate matching airfoil profiles based on the given desired aerodynamic parameters.Test results show the average of root mean square error between the predicted and real airfoil is 0.17%,and the prediction time costs only 23 ms,which greatly improves the design accuracy and efficiency.The design performance remains good in the presence of noise interference,which enhances the robustness of the airfoil design model.The research results can be applied to adaptive online optimal aerodynamic configuration control of morphing vehicle.

关 键 词：变体飞行器 条件生成对抗网络 翼型反设计 多层感知 深度学习 

分 类 号：V211.3[航空宇航科学与技术—航空宇航推进理论与工程]