Optimization of Antenna Performance based on VAE-BPNN-PCA

Abstract

In this paper, a machine learning method combining variational auto-encoder (VAE) and back propagation network (BPNN) is proposed to predict the performance of a double-T monopole antenna, which is the return loss $(S_{11})$ in frequency 1.5-6GHz. Then the trained machine learning model can take the place of the traditional electromagnetic simulation software CST, to predict $S_{11}$ in a few seconds. Genetic algorithm (GA) is used to optimize $S_{11}$ in frequency band 2.4-3.0GHz and 5.15-5.6GHz. This method shows an accurate and efficient prediction performance for that the average prediction errors of training and testing samples for all frequency points are 1.08% and 1.07%, respectively. The optimized antenna performance based on VAE-BPNN-GA is almost in accordance with that acquired by CST, that is, the average prediction error for all frequency points is 1.20%, while occupying less computation time. VAE-BPNN-GA is also compared with PCA-BPNN-GA, and proved to have higher prediction accuracy.