基于神经网络的RF MEMS移相器面向设计参数建模

Design Parameters Modeling for RF MEMS Phase Shifter Based on Artifical Neural Network

以毫米波段MEMS移相器为研究对象，提出一种直接面向设计参数的建模方法。该方法直接选取分布式RFMEMS移相器的关键设计参数作为建模目标，通过HFSS仿真获得人工神经网络建模的样本数据，并使用三种神经网络对移相器的S参数及设计参数进行建模。实验结果表明，与HFSS仿真数据相比，面向设计参数直接建模的方法对于移相器中心频率f0、10dB带宽B和插入损耗最小值（33-37GHz）建模误差绝对值均值分别在0．094～0．171GHz、0．085-0．159GHz和0．040-0．048dB之间。相比于基于S参数间接建模方法的准确率均至少提高了50％。

A modeling arithmetic based on design parameters for MEMS phase shifter working in millimeter wave is presented. Three crucial design parameters were selected as modeling objects. The HFSS software was used to abtain the data for training and valuing the three different neural networks which modeled the S and design parameters of the phase shifter. The experiments show that compared with HFSS simulation results, the modeling method presented in paper for the center frequence, wideband and minimum insertion loss （33-37 GHz） of the phase shifter had mean absolute error of 0. 094-0. 171 GHz, 0. 085-0. 159 GHz and 0. 040-0. 048 dB respectively. The accuracy of model has been improved at least 50% comparing to the method based on S parameters.