Two-Stage Hybrid Precoding for Minimizing Residuals Using Convolutional Neural Network

Abstract

This letter presents a two-stage (Ts) hybrid precoding algorithm based on convolutional neural network (CNN) with minimum hybrid precoding residual, named as Ts-CNN algorithm. Firstly, using the optimal precoding matrix as a label, a new CNN architecture is constructed, which can be trained to learn how to predict the vectorized hybrid precoder. Furthermore, a two-stage hybrid precoding scheme is proposed. In the first stage, the proposed CNN learns how to minimize the residual between the hybrid precoder and the optimal precoder when it only has the estimated channel as its input. Specifically, two special network layers are designed to satisfy the constraint conditions of the output. In the second stage, the estimated channel is fed into the well-trained CNN, correspondingly, the analog precoding matrix and the digital precoding matrix are output. Simulation results and complexity analysis show that under certain conditions, the proposed method offers better spectral efficiency than the other related methods with acceptable complexity.