基于知识迁移的高效涡轮叶栅智能气动优化方法  

Knowledge Transfer Based Aerodynamic Optimization Method for Efficient Turbine Cascade Design

作　　者：李存晰 宋立明[1] 郭振东[1] 李军[1] LI Cunxi;SONG Liming;GUO Zhendong;LI Jun(Institute of Turbomachinery,Xi'an Jiaotong University,Xi'an 710049)

机构地区：[1]西安交通大学叶轮机械研究所,西安710049

出　　处：《机械工程学报》2025年第4期333-343,共11页Journal of Mechanical Engineering

基　　金：国家自然科学基金(52306048);国家科技重大专项(2019-II-0008-0028)资助项目。

摘　　要：为实现涡轮叶栅气动形状的高效优化,基于人工智能领域知识迁移理念,开展基于知识迁移的高效涡轮叶栅智能气动优化方法研究。首先,搭建叶型变分自动编码器模型,利用其解码器实现气动形状参数化;其次,为提高已完成任务样本重参数化精度,提出基于样本加权的模型重训练策略,并设计出基于梯度信息的叶型重参数化算法,将已完成任务样本编码至目标任务空间;然后,实现基于多保真度模型的贝叶斯迁移优化算法,以有效利用迁移源任务信息加速目标任务优化进程;最后,搭建智能涡轮叶栅气动迁移优化框架。通过在GE-E3低压涡轮算例中开展叶型设计优化表明,在计算成本相同的情况,知识迁移策略所获得的最优解相对参考设计总压损失减少0.66%,相对无知识迁移策略所获得的最优解总压损失减少0.18%;而在获得相似性能优化解的情况下,知识迁移策略所需计算成本相对传统无知识迁移方法降低50%以上。由此验证了所提出框架的性能优势。For the efficient aerodynamic shape optimization(ASO)of turbine blades,inspired by the concept of knowledge transfer from artificial intelligence,research on knowledge transfer based intelligent aerodynamic shape optimization methods for turbine blades is conducted.Firstly,the variational autoencoder(VAE)model is built,with its decoder to realize the intelligent parameterization of turbine blades.Secondly,for the improvement of the reparameterization accuracy of samples from related completed tasks,a sample-weighted model retraining strategy is proposed for encoding samples to a common design space.Besides,a gradient-based reparameterization algorithm is also designed.Moreover,the Bayesian transfer optimization algorithm is constructed based on multi-fidelity surrogates to take full advantage of transfer source samples and accelerate the optimization progress of target tasks.Eventually,an intelligent transfer ASO framework for turbine blades is established.Transfer optimization for low pressure turbine blade of GE-E3 is carried out using the proposed framework.With the same computational cost,the total pressure loss coefficient of the optimal blade is reduced by 0.66%compared to benchmark design,and 0.18%lower than legacy approach without knowledge transfer strategy.With reaching the same level of optimized performance,the computational cost of the optimization with proposed framework can be reduced by at least 50%compared to legacy approach.Thereby,the effectiveness of proposed knowledge transfer accelerated ASO framework for turbine blades is well demonstrated.

关 键 词：知识迁移 贝叶斯优化 多保真度代理模型 涡轮气动形状优化 

分 类 号：TK14[动力工程及工程热物理—热能工程]