面向吸气控制设计阻力评估的机器学习建模策略  

Machine Learning Modeling Strategies for Drag Assessment in Suction Flow Control Design

作　　者：韩艺 崔晓春[1] 杨龙 张铁军[1] HAN Yi;CUI Xiaochun;YANG Long;ZHANG Tiejun(Aero Science Key Lab of High Reynolds Aerodynamic Force at High speed,AVIC Aerodynamics Research Institute,Shenyang Liaoning 110034,China)

机构地区：[1]中国航空工业空气动力研究院高速高雷诺数气动力航空科技重点实验室,辽宁沈阳110034

出　　处：《海军航空大学学报》2024年第6期649-659,共11页Journal of Naval Aviation University

基　　金：航空科学基金(2023M071027001)。

摘　　要：为获得高精度吸气控制阻力评估建模策略,文章选用基于径向基核函数和马特恩核函数的两种高斯过程回归、基于径向基核函数和Sigmoid核函数的两种支持向量回归、基于径向基核函数的核岭回归5种机器学习模型,以NACA0012翼型为案例,对比了不同调参评估指标、建模变量、采样方法、机器学习模型的建模效果。建模变量研究结果表明:与阻力系数直接建模相比,吸气控制减阻差量建模可将回归预测的平均绝对误差降低24%~62%;将无吸气条件下翼型的阻力系数作为建模输入变量,可将回归预测的平均绝对误差降低43%~86%。总体结果表明,面向吸气控制设计阻力评估的机器学习建模具备可行性。To develop a high-precision modeling strategy for drag assessment in suction flow control,five machine learning models are selected:two Gaussian Process Regression(GPR)models based on Radial Basis Function(RBF)kernel and Matérn kernel,two Support Vector Regression(SVR)models based on RBF kernel and Sigmoid kernel,and one Kernel Ridge Regression(KRR)model based on RBF kernel.Using the NACA0012 airfoil as a case study,the modeling effects of different hyperparameter evaluation metrics,modeling variables,sampling methods,and machine learning models are compared.The study on modeling variables shows that compared to direct modeling of the drag coefficient,modeling the drag reduction increment in suction flow control can reduce the mean absolute error of regression prediction by 24%to 62%.Using the drag coefficient of the airfoil under non-suction conditions as an input variable can reduce the mean absolute error of regression prediction by 43%to 86%.The overall results indicate that machine learning modeling for drag assessment in suction flow control design is feasible.

关 键 词：吸气控制技术 流动控制 机器学习 气动特性建模 建模策略 

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]