Antenna Notch Structure Optimization Using Deep Neural Networks

Abstract

To address the stressful and time-consuming problem with the current notched antenna modelling optimization tools, an improved deep multilayer perceptron (DMLP) neural network framework is designed. The method introduces an attention mechanism (Attn) layer to improve the interpretability of the model, uses the leaky ReLU activation function to prevent the gradient from vanishing, and optimizes the structure of the DMLP model using an improved particle swarm algorithm (PSO) to improve the model prediction accuracy. Then, the notch structure geometric parameters of the designed double-notch ultra-wideband (UWB) antenna serve as input to predict the return loss S11 of the antenna. The experimental results show that the method reduces the root mean square error of prediction for S11 by 73.01% compared to the traditional MLP and 64.14% compared to the unimproved DMLP, which provides a solution for modelling notched UWB antennas and helps to optimize the design of this type of antenna.