基于小样本迁移代理模型的高速永磁电机应力优化  

Stress Optimization of High-speed Permanent Magnet Motor Based on Few-shot Transfer Learning Surrogate Model

作　　者：谢冰川 张岳 刘光伟[1] 张凤阁[1] 李永胜 XIE Bingchuan;ZHANG Yue;LIU Guangwei;ZHANG Fengge;LI Yongsheng(School of Electrical Engineering,Shenyang University of Technology,Shenyang 110870,Liaoning Province,China;School of Electrical Engineering,Shandong University,Jinan 250100,Shandong Province,China;Shandong Tianrui Heavy Industry Co.,LTD.,Weifang 261000,Shandong Province,China)

机构地区：[1]沈阳工业大学电气工程学院,辽宁省沈阳市110870 [2]山东大学电气工程学院,山东省济南市250100 [3]山东天瑞重工有限公司,山东省潍坊市261000

出　　处：《中国电机工程学报》2024年第13期5362-5373,I0029,共13页Proceedings of the CSEE

基　　金：国家重点研发计划项目(2021YFE0108600);国家自然科学基金项目(52077121,51920105011)。

摘　　要：针对目前电机代理模型样本空间计算效率低的问题,该文提出迁移径向基函数神经网络代理模型。该模型的样本信息源于足量可快速获取或先验积累的低可信度数据和少量高可信度的数据。基于高速永磁电机应力优化问题,研究该模型对小样本的学习能力、对电机相似拓扑结构的学习能力以及基函数对迁移模型精度的影响,证明该模型在工程优化中的效率优势。最后展望迁移学习技术在电机优化中的发展趋势。This paper proposes a transfer learning radial basis function neural network surrogate model to improve the computational efficiency of the sample space of the traditional motor surrogate model.The sample data of this model originate from enough low-fidelity data that can be quickly obtained or prior cumulative and few high-fidelity data.Based on the stress optimization problem of the high-speed permanent magnet motor,the learning ability of the model with few-shot samples,the learning ability to similar topological structures,and the influence of the radial basis function of the model are studied,and this model is proved efficient in the engineering optimization field.Finally,the development trend of transfer learning technology in motor optimization is prospected.

关 键 词：径向基函数神经网络 电机优化 高速永磁电机 应力优化 代理模型 迁移学习 迁移模型 工程优化 

分 类 号：TM352[电气工程—电机]