Near-Field Beam Training Based on Deep Learning for Extremely Large-Scale MIMO

Extremely large-scale multiple-input multiple-output (XL-MIMO) is considered as a key technology for future wireless communications. For near-field beam training in XL-MIMO, the distance information is another important factor aside of the angular information. The most recent work trains the deep neural network (DNN) by performing beam training based on far-field codebook that only contains angular information, where the optimal angle and distance are independently predicted. In this letter, the DNN is trained by performing beam training based on near-field codebook that contains both angular and distance information, where the optimal angle and distance are jointly predicted with improved performance. We first propose a deep learning-based near-field beam training (DNBT) scheme. To further improve the beam training performance, a DNBT with supplementary codewords (DNBT-SC) scheme is proposed, where the supplementary codewords are selected to perform beam training based on the probability vector acquired in the DNBT scheme. Simulation results show that under the same training overhead, the proposed schemes can achieve better performance than the existing schemes.