Model-Driven Deep Learning-Based Sparse Channel Representation and Recovery for Wideband mmWave Massive MIMO Systems
Jianqiao Chen, Fanyang Meng, Nan Ma, Xiaodong Xu, Ping Zhang
Abstract
In this paper, we exploit a novel model-driven deep learning (MDDL)-based scheme for efficient wideband millimeter-wave (mmWave) massive multiple-input multiple-output (MIMO) channel estimation where the neural networks for sparse channel representation and recovery are respectively designed. For the former, we propose an <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">angular-resampling network</i> to determine the sampling intervals of grids adaptive to the true angles of arrival (AoAs) of paths, based on which an effective dictionary for sparse channel representation in angle-domain can be constructed. To this end, the neural network consisting of three modules trained with a newly-designed angular Gaussian-mixture distribution loss function is developed. For the latter, we propose an inverse-free variational Bayesian learning (IF-VBL) driven <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">deep-unfolding network</i> for sparse channel recovery. Specifically, the IF-VBL method by maximizing a general relaxed evidence lower bound (ELBO) is first developed, which is then unfolded into a layer-wise architecture where some a-priori parameters are learned. Simulation results verify the superiority of the proposed MDDL-based channel estimation scheme with significantly improved convergence and performance over counterparts.