Modeling Optimization of UWB-MIMO Antenna Based on Novel Residual Network

To solve the problem of slow multi-parameter optimization in radio frequency (RF) device modeling, this paper designed a new network structure based on residual network (ResNet) to optimize the modeling process of Ultra Wide Band (UWB) Multiple-input Multiple-output (MIMO) antenna. In this paper, the new ResNet structure will first connect the input at the fully connected (FC) layer, to broaden the input parameters when entering the ResNet layer. The one-dimensional convolution check data is processed according to the data type and finally connected to the three-layer fully connected network. The geometric parameters of the UWB-MIMO antenna are modeled by using the new ResNet in this paper, and the features are extracted from the 8 geometric parameters of the antenna to predict the $S_{11}$ value of the antenna. Experiments show that the network has obvious advantages in fitting accuracy and fitting speed compared with convolutional neural networks (CNN), deep multilayer perceptron (DMLP), and shallow multilayer perceptron (MLP). To some extent, the new ResNet solves the problem of insufficient global feature extraction of CNN.