Feature Attention Based Blind Denoising Network for mmWave Beamspace Channel Estimation

Abstract

In millimeter-wave (mmWave) MIMO systems, when the number of radio frequency (RF) chains are limited, estimation of the beamspace channel can become compelling. Also, as the number of RF chains decreases, pilot overhead increases to make channel estimation reliable, eventually reducing the spectral efficiency. In this paper, we propose a channel estimation method which combines compressive sensing (CS) method of GM-LAMP that assumes beamspace channel elements follows the Gaussian mixture distribution a priori, with a novel denoising network based on sparse feature attention for the estimation. According to performance analysis and simulation results, the GM-LAMP combined with feature attention based denoising neural network outperforms state-of-the-art compressed sensing-based algorithms. Furthermore, the proposed method also outperforms previous LAMP-based neural networks with comparable processing time, albeit using less pilot transmission.