Joint CSI Acquisition Based on Deep Learning for FDD Massive MIMO Systems

Abstract

AbstractTo fully utilize the spatial multiplexing gains and the array gains of massive multiple-input multiple-output (MIMO), the channel state information (CSI) must be acquired at the transmitter side. However, conventional solutions involve overwhelming overhead both for downlink channel training and uplink feedback in frequency-division duplexing (FDD) massive MIMO systems. In this paper, we propose a joint CSI acquisition solution based on deep learning (DL) to achieve the goal of reducing the overhead. Particularly, we consider a millimeter-wave (mmWave) massive MIMO system based on lens antenna array, in which dispatching users directly feedback the pilot observation to the base station (BS), and then joint CSI can be recovered at the BS. We further take advantage of the beamspace channel to transform the joint CSI recovery problem at the BS into a sparse signal recovery problem, and then we propose a learned-approximate message passing (LAMP) network based on deep neural network (DNN) to obtain CSI. Simulation results demonstrate that the proposed scheme can provide accurate CSI with lower overhead compared with the existing conventional CSI acquisition schemes.KeywordsMassive MIMOCSIFDDDeep learningmmWave