Knowledge-Driven Deep Learning Based Channel Estimation for Terahertz Extra-Large MIMO Systems

Accurate channel state information (CSI) is critical in wideband terahertz (THz) extra-large multiple-input multiple-output (XL-MIMO) systems. However, the existing channel estimation methods suffer from the beam squint effect. In this paper, we propose a knowledge-driven channel estimation network called generalized multiple measurement vector learned approximate message passing (GMMV-LAMP) to jointly estimate the channels, whereby the approximately same physical angles shared by different subcarriers can be exploited. In particular, we propose a frequency-dependent discrete Fourier transform (DFT) wideband redundant dictionary (WRD), so that the support shift caused by the beam squint effect is eliminated. Simulation results demonstrate that the proposed approach can effectively achieve channel estimation with only several layers at the cost of limited pilots.