Compressive Channel Estimation Based on the Deep Denoising Network in an IRS-Enhanced Massive MIMO System

Abstract

Integrating large intelligent reflecting surfaces (IRS) into a millimeter-wave (mmWave) massive multi-input-multi-output (MIMO) technique has been a promising approach to enhance the performance of the wireless communication system with the channel state information (CSI). Most existing work assume that ideal channel estimation can be obtained, but the proposed high-dimensional cascaded MIMO channels and passive reflectors pose a great challenge to these methods. To address the abovementioned problems, we proposed a new method for the reduction of training overhead in IRS with a partial ON/OFF model and an optimizing strategy for pilot design approach. The energy consumption of large-scale antenna arrays and the pilot overhead in the training phase of signal transmission are greatly reduced. Besides, we proposed an improved deep residual shrinkage denoising network, which possesses better denoising performance with a soft thresholding model. The channel data can be denoised by deep learning methods, which greatly improve the accuracy of channel estimation. Simulation results demonstrate that the superiority of the proposed network over prior solutions.