An Effective Network With Discrete Latent Representation Designed for Massive MIMO CSI Feedback

Abstract

The efficacy of massive multiple-input multiple-output techniques heavily relies on the accuracy of the downlink channel state information (CSI) in frequency division duplexing systems. Many works focus on CSI compression and quantization to enhance the CSI reconstruction accuracy with lower overhead of downlink pilots and uplink feedback. In this letter, an advanced network named Conformer is first introduced for CSI compression, which combines self-attention mechanisms and convolution to efficiently extract both global and detailed CSI features. In order to further reduce the feedback overhead, we also propose a vector quantization scheme based on the discrete latent representation of the vector quantised-variational autoencoder (VQ-VAE), namely VQCFB. Integrating Conformer blocks with VQCFB, the proposed encoder-quantizer-decoder framework achieves high-quality CSI reconstruction with low feedback overhead, outperforming previous state-of-the-art networks.