A Deep Learning-based Efficient Beamspace Estimation Approach in Millimeter-Wave Massive MIMO Systems

In millimeter Wave (mmWave) massive MIMO technology, a group of lens antennas can be used to significantly reduce the number of radio frequency (RF) chains required for wireless communication. However, the estimation for channels is difficult for having fewer RF chains than antennas whereas the estimation of the beamspace (BC) can be described as a sparse signal recovery problem using the limited number of BCs. This problem can be solved using the approximate message passing (AMP) iterative algorithm and the associated learned AMP which is known as LAMP can be implemented through a deep neural network (DNN). In this study, we present a prior-aided GM-LAMP-based beamspace-oriented estimation of channels strategy to improve the performance of the channel estimation that is accomplished by substituting the freshly generated Gaussian mixture (GM) shrinkage function for the initial shrinking component in the LAMP network. Through simulations, the GM-LAMP network’s efficiency is confirmed using both a theoretical channel model and a channel dataset obtained through a ray tracing model. The findings show where the present GM-LAMP networks outperform the earlier techniques in terms of channel estimate accuracy.