Deep Learning Compressed Sensing-Based Beamspace Channel Estimation in mmWave Massive MIMO Systems

Abstract

Channel estimation attaches great importance in millimeter wave (mmWave) massive multiple input multiple output (MIMO) systems. This letter proposes a two-step orthogonal matching pursuit (OMP) method to estimation channel state information (CSI) based on deep learning compressed sensing. Specifically, in the first-step OMP, a composite convolution kernel function (CKF) is designed for coarsely estimating angles of arrival/departure (AoAs/AoDs) from correlation matrix. In the second-step OMP, a Squeeze-and-Excitation Residual network (SE-Resnet) with Noise2Void (N2V) learning strategy is presented to denoise correlation matrix and finely estimate AoAs/AoDs. The proposed method can work without labeled data. Simulation shows that the two-step OMP significantly outperforms state-of-the-art mmWave channel estimation methods. Moreover, it works robustly in low signal-to-noise ratio (SNR) regimes with a small number of training frames.