CSI Compression With Kernel-Based Sparsity Learning for FDD Massive MIMO Systems

In frequency division duplex massive MIMO systems, the base station (gNB) needs to acquire the downlink channel state information (CSI) from the user equipment (UE) at regular intervals. The CSI feedback overhead increases exponentially as the number of antennas in the system increases. To reduce the overhead, in this paper, we propose a kernel-based sparsity learning (KSL) technique for predicting the channel coefficients. To meet the accuracy of the predicted channel coefficients, we design an error quantization codebook by employing a histogram-based non-uniform quantizer. At the UE, we develop a model that uses only a few channel coefficients and the error codeword index to predict the remaining channel coefficients at a given time instant. This model is then reported to the gNB to recover the CSI matrix. The performance of the developed model using existing regression techniques is also analyzed. The experiments are performed with both the real-time dataset and synthetic benchmark dataset, and the results are reported in this paper. The experimental results using real-time data reveal that the KSL reduces the prediction time by approximately 50% compared with other regression techniques, without any compromise in reconstruction accuracy. The experiments on synthetic data show that the proposed model is better than the baseline on reconstruction performance. Specifically at the CR =1/4, the normalized mean square error of the baseline CsiNet is 40% and 67% higher than the proposed method for indoor and outdoor scenarios, respectively.