A Deep Compressed Sensing Approach to Channel Estimation in Reconfigurable Intelligent Surface–Aided MU‐MIMO Networks

ABSTRACT The preservation of high data rates and reliable operation of reconfigurable intelligent surface (RIS)–assisted wireless communication systems requires accurate channel state information (CSI), especially within the high‐frequency bands of millimeter wave (mmWave). But the problem is that CSI acquisition in the direction of large‐scale antenna arrays and RIS elements as the passive elements is highly fickle because of high dimensionality and the extra need for vast amounts of pilot overhead and a subsequent decision complexity. To address this challenge, we propose a novel coherence‐optimized training pilot–based channel estimation with a deep compressed sensing–enabled ResUNet encoder–decoder architecture having skip connections. We take advantage of having the double structure of the sparsity inherent in angular cascaded channels in multiuser MIMO networks so as to recover the sparse signal efficiently with minimal pilot overhead. In contrast to traditional compressed sensing or classical deep learning architecture, the proposed ResUNet architecture achieves better convergence and estimation performance due to the talents of residual learning and attention blocks that allow the preservation of essential spatial information structure across the network layers. Moreover, we look at an optimum sensing matrix with a mutual coherence criterion to improve the recovery in a greater measure through a wide range of sparsity diligence and signal‐to‐noise ratio (SNR). Large‐scale simulations show that our approach generates a better normalized mean squared error (NMSE) compared with state‐of‐the‐art schemes like double‐structured orthogonal matching pursuit (DS‐OMP), JDCNet, and classic CNN‐based estimators. It has been shown that performance is validated at different SNRs, pilot lengths, and different sparsity ratios, which indicate robustness and scalability of the proposed architecture. All these findings reinforce the fact that the combination of pilot optimization and deep compressed sensing can be used to deliver impactful and efficient CSI acquisitions in RIS‐aided mmWave MIMO systems.