MGGAN-Based Hybrid Beamforming Design for Massive MIMO Systems Against Rank-Deficient Channels

Abstract

The design of hybrid beamforming (HBF) is one of the key issues in millimeter wave (mmWave) and terahertz (THz) massive multi-input multi-output (MIMO) communications. In particular, only rank-deficient channel state information (CSI) can be acquired in hybrid MIMO architecture, which enhances design challenges. However, this practical problem of rank-deficient channels has been avoided by most of the existing studies on HBF. In this letter, we propose a new deep learning (DL) network model called Multi-Generator Generative Adversarial Network (MGGAN) to accomplish the design of HBF against rank-deficient channels. Specifically, MGGAN includes three generators for recovering the rank-deficient channels, achieving the analog and the digital beamforming matrices, respectively. In addition, a spectrum efficiency (SE) module is introduced to take system SE as the optimization goal of the network. Simulation results show the superiority of our proposed MGGAN model in SE performance over several DL network architectures and classical HBF algorithms. Moreover, it is more robust against the rank-deficiency of hybrid architecture MIMO channels.