Channel Estimation in Massive MIMO considering Hardware Non-Linearities using Deep Learning

Abstract

Deep learning is getting more popular in areas like computer vision and natural lan guage processing as it is difficult to describe real-world images and languages using rigorous mathematical models. Although deep learning algorithms are easy to imple ment, it is difficult to generate robust algorithms for real systems that include non linearity. This dissertation discusses the combined effects of non-linearities present in the hardware of base station (BS) and user equipment (UE) on single cell multiple input multiple output (MIMO) uplink performance in the practical Rican fading environment. The effective channel and distortion characteristics based on Bussgang decomposition are derived from the analytical method. Two deep feed forward neural networks are trained to estimate the effective channel and distortion variance at each BS antenna used for signal detection. The performance of the proposed method is compared with the Bayesian Estimator (Linear Minimum Mean Square Error (LMMSE)) for distortion aware and unaware scenario. The proposed deep learning based estimator uses attenu ation characteristics to improve the quality of the estimate, while the LMMSE method treats distortion as noise. Efficiency for non-linearity of order three or higher with deep learning-based estimators is significantly high which can be shown using the data gen erated by derived effective channels considering both the BS and UE non-linearities of general order, and hence deep learning based provides better estimate of the channel.