【作者】 张哲； 杨武； 陆卫兵；

【Author】 Zhe Zhang;Wu Yang;Wei-Bing Lu;College of Software,Southeast University;School of Information Science and Engineering,Southeast University;

【机构】 东南大学软件学院(苏州)； 东南大学信息科学与工程学院,毫米波国家重点实验室；

【摘要】 子全域(SED)基函数方法是求解大规模有限周期结构(LFPSs)电磁特性的一种有效方法,然而,当阵列规模变大时,该方法中SED基函数拓展系数的计算依然非常耗时。文献[1]采用人工神经网络(ANNs)快速预测SED基函数拓展系数,从而加速SED基函数方法,但该研究未考虑不同电磁波入射角的影响,不能很好满足实际工程需要。本文在该工作基础上,进一步研究针对不同角度电磁波入射下大规模有限周期结构散射特性计算的SED基函数加速方法。对因数据量急剧增加,导致模型训练时出现精度差、梯度爆炸等问题,采用残差神经网络(ResNet)有效解决,基于MindSpore框架搭建模型进行实验,结果验证了本方法的准确性和有效性。更多还原

【Abstract】 The Sub-Entire-Domain(SED) basis function method is an effective method to solve the electromagnetic characteristics of large-scale finite periodic structures(LFPSs).However,when the array size becomes larger,the calculation of the expansion coefficient of the SED basis function in this method is still very time-consuming.In literature [1],artificial neural networks(ANNs) were used to rapidly predict the expansion coefficient of the SED basis function,thus accelerating the SED basis function method.However,this study did not consider the impact of different incident angles of electromagnetic waves,and could not well meet the actual engineering needs.On the basis of this work,this paper further studies the SED basis function acceleration method for the calculation of scattering characteristics of large-scale finite periodic structures under different angles of electromagnetic wave incidence.For the problems of poor accuracy and gradient explosion in model training caused by the sharp increase of data volume,the residual neural network(ResNet) was used to effectively solve the problem,and the model was built based on MindSpore framework for experiment,and the results verified the accuracy and effectiveness of this method.更多还原