Adaptive Compression of Massive MIMO Channel State Information with Deep Learning

TL;DR

This work addresses CSI compression for massive MIMO in 6G by proposing a deep autoencoder-based approach that operates on a complex-to-real representation of the channel state information. A two-layer autoencoder learns a latent CSI representation whose size decreases as the compression ratio \(\kappa\) increases, enabling adaptive compression decisions via a simple selector or classifier. The approach is evaluated under CDL-C/CDL-E channel models with varying URA sizes, showing that the adaptive scheme outperforms static compression and nears uncompressed performance at high SNR, while maintaining a run-time that is effectively independent of the compression ratio. This demonstrates the practicality of AI-native CSI compression to reduce feedback/pilot overhead and support efficient neural receivers in 6G massive MIMO deployments.

Abstract

This paper proposes the use of deep autoencoders to compress the channel information in a \review{massive} multiple input and multiple output (MIMO) system. Although autoencoders perform lossy compression, they still have adequate usefulness when applied to massive MIMO system channel state information (CSI) compression. To demonstrate their impact on the CSI, we measure the performance of the system under two different channel models for different compression ratios. We disclose a few practical considerations in using autoencoders for this propose. We show through simulation that the run-time complexity of this deep autoencoder is irrelative to the compression ratio and thus an adaptive compression rate is feasible with an optimal compression ratio depending on the channel model and the signal to noise ratio.

Figures (5)

• Figure 1: An overview of channel compression onto the estimated channel $\mathbf{\hat{H}}$ with channel compression $\mathsf{C}_\kappa$, quantization $Q$, and decompression $\mathsf{C}_\kappa^{-1}$ functions (in gray). Precoder $\mathbf{F}$ and combiner $\mathbf{G}$ are computed from the reconstructed channel $\mathbf{\hat{\hat{H}}}$. Dashed (thick) borders represent functions that are computed at the user (base station) side.
• Figure 2: An autoencoder has three different components: an encoder, a latent layer, and a decoder.
• Figure 3: Timing diagram showing two training-inference time patterns: duty cycle (top) and staggering (bottom).
• Figure 4: The CDL-E channel state information matrix compression rate $\mathbf{H}$of a small ($4\times 4$) uniform rectangular antenna for the first receive antenna and a given user under different scenarios: uncompressed, compressed, and reconstructed at transmit SNR $\rho = 30$ dB.
• Figure :