Deep Learning for Beam Training in Millimeter Wave Massive MIMO Systems
Chenhao Qi, Yujie Wang, Geoffrey Ye Li
Abstract
This paper investigates deep learning for beam training in millimeter wave (mmWave) massive multiple-input multiple-output (MIMO) systems. Different from the existing work, we explore how the deep learning can be applied for beam training. In particular, a deep neural network (DNN) is used to deal with the nonlinear and nonmonotonic properties of channel power leakage in mmWave communications. Accordingly, we propose two DNN-based beam training (DBT) schemes. The first scheme, named original DBT (ODBT), uses a DNN to predict the beam combination best matching the strongest channel path of the mmWave channel based on the probability vector. The other scheme, named enhanced DBT (EDBT), performs additional beam training tests after obtaining the probability vector. Simulation results show that the proposed schemes can achieve satisfactory performance in terms of successful rate and achievable rate with substantially reduced beam training overhead and improved signal coverage.