Time-Frequency-Space Joint Extrapolation for 6G Near-Field Non-Stationary Massive MIMO Channels

Abstract

Massive multi-input multi-output (MIMO) systems confront the challenges of significant pilot overhead required for multi-domain channel state information (CSI) acquisition, especially in near-field non-stationary 6G environments with high-speed user mobility. To address this issue, we design a novel network architecture for multi-user time-frequency-space joint channel extrapolation that integrates a temporal graph convolutional network (TGCN) and a dual-discriminator generative adversarial network (2DGAN). By leveraging geographic topology to capture time-frequency-space correlations within the channels and multi-user interactions, the designed network reconstructs full spatial CSI from partial spatial CSI and extrapolates future subcarrier allocation states on the basis of historical allocations. Consequently, this approach promotes efficient information sharing to reduce CSI acquisition overhead. Simulation results show that our proposal performs nearly ideally, surpasses existing methods, and significantly improves the accuracy and robustness of channel extrapolation.