Multi-Resolution Compressive Sensing With Deep Unfolding For Low-Pilot Channel Estimation In Massive MIMO Systems

Abstract

To realize high spectral efficiency, massive multiple-input multiple-output (MIMO) systems demand precise channel state information (CSI). However, existing channel estimation approaches come with considerable pilot overhead and computational burden. To overcome these issues, we present MR- DUCS, Multi-Resolution Deep Unfolded Compressive Sensing, a novel pilot-efficient massive MIMO channel estimation framework based on multi-resolution sparse representation and model-driven deep unfolding. We first exploit multi-resolution dictionaries to learn inherent angular-domain sparsity of massive MIMO channels with multiple spatial resolutions. This significantly reduces the dictionary mismatch and enhances the sparse support recovery. Then we unfold the iterative CS process into learnable network layers, which enables learning-based optimization of the sparse recovery parameters. We further design an adaptable module to select or update the proper resolution conditioned on the channel reconstruction and residual information. Through extensive simulation under various SNRs, pilot configurations and antenna sizes, we show MR- DUCS can deliver more accurate channel estimation using much fewer pilot symbols compared with least squares (LS), minimum mean squared error (MMSE), orthogonal matching pursuit (OMP), and vanilla deep learning-based method.