Feature Quantization Convolutional Neural Networks for CSI Feedback in Massive MIMO Systems

Haohui Jia, Na Chen, Hui Gao, Minoru Okada

Abstract: In frequency division duplex (FDD) mode, the convolutional auto encoder (AE) compressed downlink channel state information (CSI) feedback can occupy less band resource and enable fast base station (BS) configuration for massive multiple-input multiple-out (MIMO) systems. In this paper, we propose a novel discrete quantization network (DQNet), which can quantify feature vectors with a discrete latent embedding space, to reduce feedback overhead while maintaining abundant feature information. Specifically, the discrete latent space reduce the dimension of CSI features by measuring the Ebullience distance and converting the CSI features into a small-size and discrete index information. Due to the benefit of DQNet, we apply sensing network (SN) in both encoder and decoder side, to enhance the CSI features by spanning several convolutional operations. According to experimental results, the proposed DQNet can improve the reliability in outdoor scenario with less training overhead and reduce model complexity.

External IDs:dblp:conf/hpcc/JiaCGO22