Deep CNN-Based Spherical-Wave Channel Estimation for Terahertz Ultra-Massive MIMO Systems
Yuhang Chen, Chong Han
Abstract
The Terahertz (0.1-10 THz) band is envisioned to meet the demanding 100 Gbps data rates for 6G wireless communications. Aiming at combating the distance limitation problem in THz communications, the ultra-massive MIMO with array-of-subarrays (AoSA) hybrid beamforming is adopted as a promising technology. On the downside, an enlarged channel dimension gives rise to the high-complexity of channel estimation (CE). Furthermore, the distinctive spherical-wave propagation characteristic restricts the applicability of most existing CE techniques. To address these challenges, a deep convolutional neural network (DCNN)-based spherical-wave CE method for THz AoSA communication systems is developed in this paper. A fifteen-layer DCNN structure is designed, within which the spherical-wave channel parameters including the azimuth and elevation angles, the amplitude of the channel gain, and the phase shift matrix are carefully processed to be training labels. After leveraging supervised learning with a self-defined loss function from the labeled data, the DCNN is trained offline and deployed online to conduct CE. Extensive simulation results demonstrate that compared to existing on-grid and off-grid methods, the proposed DCNN converges fast with reduced complexity. Moreover, the proposed DCNN algorithm can achieve outstanding accuracy with the normalized mean-square-error (NMSE) of -13.8 dB at SNR=10 dB.