A Deep Learning based Approach for 5G NR CSI Estimation

Accurate estimation of channel state information (CSI) and sharing the same to transmitter is crucial to MIMO systems for efficient link adaptation. Although, performance of conventional CSI algorithms is satisfactory in 4G scenarios, scaling in bandwidth or transmission ranks typically results in degradation of accuracy. To address this behavior, we propose a deep learning (DL) based estimation framework for 5G New Radio (NR) focusing on both accuracy and complexity. The proposed model consists of two stages, a shared first stage for feature extraction followed by a second stage for feature combining. Through simulations, we show that the proposed model can improve the spectral efficiency (SE) achieved by up to 20.5% due to signal to noise ratio (SNR) gain of 1. 5dB compared to conventional approaches. Further, we achieve the gains at a complexity 11% lesser than its conventional counterparts, due to the common feature generation stage used.