Masked Autoencoder-Aided Channel Extrapolation with SRS Hopping for 5G-A Systems

Abstract

Accurate channel state information (CSI) acquisition is vital for reliable transmission in modern wireless systems. In practical fifth-generation New Radio (NR) time-division duplex (TDD) massive multiple-input multiple-output (MIMO) systems, the time-varying CSI acquisition is highly challenging due to the hopping uplink pilot patterns, which has not been fully investigated in the existing works. This paper proposes a masked autoencoder (MAE)-aided channel extrapolation framework tailored to this protocol-constrained setting, which captures the intrinsic correlations of high-dimensional CSI via diverse masked reconstruction tasks. Then, a Sounding Reference Signal (SRS)-hopping-aware mask is employed to perform channel extrapolation from partial known CSI. Next, an improved CrossMAE architecture is proposed to reduce decoder inference complexity and fully integrate multilayer encoder features. The same framework is further extended to direct precoding matrix extrapolation with a spectral-efficiencyoriented loss. Simulation results verify that the proposed method outperforms existing baselines in extrapolation accuracy and spectral efficiency gain with lower pilot overhead.