Deep Learning for Massive MIMO: Channel Completion for TDD Downlink

In a realistic fifth generation (5G) massive multiple-input multiple-output (MIMO) system, hardware constraints often pose challenges towards network design that are not sufficiently considered in the literature. In this work, we consider a time division duplex (TDD) network where user equipments (UEs) are equipped with N> 1 antennas for receiving in the downlink (DL) but only with a single antenna for transmitting in the uplink (UL). Thus it is not possible to learn the complete downlink channel in a single timeslot from the uplink utilizing channel reciprocity. In this paper, we propose a novel solution based on deep learning with auxiliary input of the estimated single antenna channel in the uplink to accomplish the downlink channel completion for full rank transmission from the base station (BS). We use synthetic data for deep learning training and testing provided by the stochastic quasi-deterministic radio channel generator (QuaDRiGa). Evaluation results show that our work outperforms existing deep learning based algorithms and can provide highly effective recovered channels even with complex channel data and low compression ratio.

I. Introduction

Massive multiple-input multiple-output (MIMO) is the core technology in fifth generation (5G) to fulfill the promise of a 10 times spectral efficiency increase compared to fourth generation (4G). While the paper from Marzetta [1] is often cited as the beginning of massive MIMO, using more and more antennas for data transmission is a continuous development. There are publications from before that also provide insights and results about the scaling of antennas [2], [3], [4], [5]. Therefore, it is safe to state, that massive MIMO, or the idea to use large scale antenna systems at the base station (BS), is a logical and consequent continuation in the development of MIMO research.