采用知识迁移加速的智能气动设计优化方法  

Knowledge Transfer-Accelerated Intelligent Aerodynamic Design Optimization

作　　者：郭振东[1] 李存晰 宋立明[1] 李军[1] 丰镇平[1] GUO Zhendong;LI Cunxi;SONG Liming;LI Jun;FENG Zhenping(School of Energy and Power Engineering,Xi’an Jiaotong University,Xi’an 710049,China)

机构地区：[1]西安交通大学能源与动力工程学院,西安710049

出　　处：《西安交通大学学报》2023年第10期53-63,共11页Journal of Xi'an Jiaotong University

基　　金：国家科技重大专项资助项目(2019-Ⅱ-0008-0028);秦创原引用高层次创新创业人才项目(QCYRCXM-2022-210)。

摘　　要：为缩短精细气动形状设计优化所需的最少性能评估次数与任务周期内所能容许的最大性能评估次数之间的差距,基于机器学习领域迁移学习理念,开展了采用知识迁移加速的智能气动设计优化方法研究。首先,搭建了翼型变分自编码器模型,利用其解码器实现了气动形状的智能参数化,同时借助其编码器将已完成任务样本统一至目标任务参数化空间;其次结合单保真度和多保真度代理模型,建立了贝叶斯迁移优化算法;然后,将翼型变分自编码器模型与贝叶斯迁移优化算法相结合,搭建了智能气动形状迁移优化框架;最后,通过任务相关性分析对知识迁移加速优化过程的机理进行了讨论。研究结果表明:通过开展翼型设计优化,智能气动形状迁移优化框架所获得的最优解中位数相较于无知识迁移的变分自编码器优化方法,性能提升了4.8%,比其他各参比方法提升了19.9%以上,验证了该知识迁移策略的有效性。To address the gap between the minimum number of evaluations required for refined aerodynamic shape optimization tasks and the maximum number of evaluations constrained by the mission period,a study on intelligent aerodynamic design optimization accelerated by knowledge transfer was conducted based on the concept of transfer learning in the field of machine learning.Firstly,an airfoil variational autoencoder(VAE)model was built.The model’s decoder enabled the intelligent parameterization of aerodynamic shapes,while its encoder unified samples from source tasks into the parameterization space of the target task.Secondly,the Bayesian transfer optimization algorithm was established,incorporating both single and multi-fidelity surrogate models.Then,an intelligent transfer optimization framework for aerodynamic shapes was established by integrating the airfoil VAE model and Bayesian transfer optimization algorithm.Last,task correlation analysis was made to discuss the mechanism of the optimization process accelerated by knowledge transfer.The results from airfoil optimization test demonstrate that such a framework improves the performance of median of the optimal solution by 4.8%compared to using VAE optimization lack of knowledge transfer,and by more than 19.9%compared to other methods,confirming the effectiveness of knowledge transfer strategy.

关 键 词：气动形状设计优化 知识迁移 变分自编码器 贝叶斯优化 多保真度代理模型 

分 类 号：TK421.4[动力工程及工程热物理—动力机械及工程]