Machine Learning Assisted DL CSI Estimation for High-Mobility Multi-Antenna Users with Partial UL CSI Availability in TDD Massive MIMO Systems

Abstract

Estimation of downlink (DL) channel state information (CSI) is necessary in massive multiple-input multiple-output (MIMO) systems to enable precoding and in general achieve high spectral efficiency. However, CSI estimation (both uplink (UL) and DL) is challenging in an environment with highly mobile users due to rapidly-varying fading. The estimation becomes even more challenging when the UL CSI knowledge is incomplete due to system constraints. In this work, we combine two machine learning techniques to tackle this twofold problem: 1) predicting DL CSI from earlier UL CSI estimates, and 2) estimating full UL CSI from its incomplete form. For the first sub-problem, we apply long short-term memory (LSTM) to capture the spatiotemporal correlation between CSI at different time instances and UE positions. For the second sub-problem, we use a conditional generative adversarial network (CGAN) to estimate the full UL CSI from its incomplete version. We study the normalized mean squared error performance of the proposed CGAN-LSTM method and compare the achieved spectral efficiency of the system with what is maximally achievable with full CSI knowledge.