Situation-Aware Channel Covariance Prediction for Deep Learning Aided Massive MIMO Systems

Designing efficient massive MIMO systems operating in a frequency-division duplexing (FDD) mode is one of the main intriguing research directions in the last decade. One of the main design challenges is reducing the huge training overhead incurred from acquiring the downlink channel knowledge at the base station. This challenge is even more prominent when serving highly-mobile users with high levels of location uncertainty. In this paper, we propose a novel situation-aware channel covariance prediction solution for downlink beamforming design. The proposed solution acquires imperfect knowledge of uplink and downlink channels and user location in the learning phase. In the operation phase, the proposed solution acquires only uplink channel estimates to predict a denoised location, which is then used to predict the downlink channel covariance matrix, for downlink beamforming design. Simulation results show the pro-posed solution achieves robust performance against uncertainty in the location information and imperfection in the downlink channel knowledge, both acquired in the learning phase, which makes it promising for supporting highly-mobile applications.