基于深度学习的翼型反设计方法  

Inverse design method of airfoil based on deep learning

作　　者：何磊[1] 钱炜祺[1] 刘滔 张显才 董康生 HE Lei;QIAN Weiqi;LIU Tao;ZHANG Xiancai;DONG Kangsheng(Computational Aerodynamics Institute,China Aerodynamics Research and Development Center,Mianyang Sichuan 621000,China)

机构地区：[1]中国空气动力研究与发展中心计算空气动力研究所,四川绵阳621000

出　　处：《航空动力学报》2020年第9期1909-1917,共9页Journal of Aerospace Power

基　　金：国家自然科学基金(11802325);中国空气动力研究与发展中心基础和前沿技术研究基金(FFTRF20172015)。

摘　　要：建立了一种基于深度学习的翼型反设计方法,将翼型曲线及其对应的压力分布图像作为训练学习对象,建立其内在联系的模型,实现通过卷积神经网络提取压力分布图像的特征,计算获得翼型曲线。该方法直接将压力分布图像作为模型输入,更加直观简洁,同时避免了传统方法中耗时的数值计算过程。模型测试中,6 000组压力分布图像和翼型曲线用于模型训练,另外561组用于模型验证,验证耗时仅6.7 s,预测的翼型曲线与CFD计算结果的平均相对误差为0.55%。对比实验中,通过对压力分布曲线添加噪声、改变输出层尺寸等方式,进一步验证和分析了预测模型性能。结果表明该翼型反设计方法具有较高预测精度和较强鲁棒性,能在保证精度的情况下降低计算时间,提高设计效率。An inverse design method of airfoil based on deep learning was proposed, and the pressure distribution image and airfoil curve were taken as the learning objects to build the prediction model. It can obtain the airfoil curve by using convolutional neural network to extract geometric features from pressure distribution image, while avoiding the time-consuming process of traditional numerical calculation method. In the test case, 6 000 samples of pressure distribution images and airfoil curves were used to train the prediction model, and the other 561 samples were used for validation with validation time only 6.7 s. The average relative error between the predicted and CFD results was 0.55%. In the comparison test, noise was added to the pressure distribution curve and the output layer size was changed to further validate the performance of the model. The results show that the proposed inverse design method of airfoil has strong robustness and prediction accuracy. The method can greatly reduce the computation time and improve the design efficiency of airfoil with satisfactory design accuracy.

关 键 词：反设计方法 翼型曲线 压力分布 卷积神经网络 深度学习 预测模型 

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]