Trainable Proximal Gradient Descent Based Channel Estimation for mmWave Massive MIMO Systems

Abstract: In this letter, we address the problem of millimeter-Wave channel estimation in massive MIMO communication systems. Leveraging the sparsity of the mmWave channel in the beamspace, we formulate the estimation problem as a sparse signal recovery problem. To this end, we propose a deep learning based trainable proximal gradient descent network (TPGD-Net). The TPGD-Net unfolds the iterative proximal gradient descent (PGD) algorithm into a layer-wise network, with the gradient descent step size set as a trainable parameter. Additionally, we replace the proximal operator in the PGD algorithm with a neural network that exploits data-driven prior channel information to perform the proximal operation implicitly. To further enhance the transfer of feature information across layers, we introduce the cross-layer feature attention fusion module into the TPGD-Net. Our simulation results on the Saleh-Valenzuela channel model and the DeepMIMO dataset demonstrate the superior performance of TPGD-Net compared to state-of-the-art mmWave channel estimators.