Antenna Array Beamforming Based on Deep Learning Neural Network Architectures

Abstract

The implementation of antenna array beamforming using several neural network (NN) architectures is compared in this paper. Gated recurrent unit, feed-forward NN, convolutional NN, and long short-term memory architectures have been used for the beamforming process. This comparative study is carried out using various metrics, such as the root mean square error, and the computational time for each NN. In addition, the mean absolute divergences of the antenna array main lobe and nulls directions from their respective desired directions have also been used to assess the performance of each beamformer. The neural networks are trained using the simulation results of a 16-element microstrip patch antenna array. It is demonstrated that deep learning-based beamformers are capable of computing optimum antenna array weights in real time and in environments that change with time.