Deep learning-based channel extrapolation and multiuser beamforming for RIS-aided terahertz massive MIMO systems over hybrid-field channels

Abstract

The reconfigurable intelligent surface (RIS) is a promising novel technology for Terahertz (THz) massive multiple-input multiple-output (MIMO) communication systems. However, the acquirement of the high-dimensional channel state information (CSI) and the efficient active/passive beamforming for RIS are challenging due to its cascaded channel structure and its lack of signal processing units. To address this, we propose a deep learning (DL)-based physical signal processing scheme for RIS-aided THz massive MIMO systems over hybrid far-near field channels, where a channel estimation scheme with low pilot overhead and a robust beamforming scheme are conceived. Specifically, we first propose an end-to-end DL-based channel estimation framework, which consists of pilot design, CSI feedback, sub-channel estimation, and channel extrapolation. Specifically, we first only activate partial RIS elements and estimate a sub-sampling RIS channel, and then utilize a DL-based extrapolation network to reconstruct the full-dimensional CSI. Moreover, to maximize the sum rate under imperfect CSI, a DL-based scheme is developed to simultaneously design the hybrid active beamforming at the BS and passive beamforming at the RIS. Simulation results show that our proposed channel extrapolation scheme has better CSI reconstruction performance than conventional schemes while greatly reducing pilot overhead and our proposed beamforming scheme has superior performance over conventional schemes in terms of robustness to imperfect CSI.

More information

Identifiers

Catalogue record

Export record