Core Concepts
The proposed DiReNet network effectively disentangles dual-polarized channel state information (CSI) into polarization-shared and polarization-specific representations, enabling efficient compression and recovery of CSI while reducing information redundancy.
Abstract
The key highlights and insights of the content are:
Dual-polarized antennas are widely used in massive MIMO systems, where the CSI in vertical and horizontal polarization directions exhibit high correlation.
The proposed DiReNet network disentangles the dual-polarized CSI into three components: polarization-shared information, vertical polarization-specific information, and horizontal polarization-specific information. This disentanglement reduces the information redundancy caused by the polarization correlation.
DiReNet incorporates a convolutional attention mechanism to selectively focus on the most essential information for accurate CSI recovery. It also employs separate fully-connected networks to further reduce the number of network parameters.
Experimental results show that DiReNet outperforms existing deep learning-based CSI compression methods by 1.5~3 dB in NMSE performance, while reducing the total number of trainable parameters by nearly one third.
The network design leverages the inherent characteristics of dual-polarized CSI, demonstrating the effectiveness of model-driven approaches that exploit domain knowledge.
Stats
The following sentences contain key metrics or important figures used to support the author's key logics:
The proposed method achieves a performance gain of 1.5~3 dB under different compression ratios, while also reducing the total number of trainable parameters by nearly 1/3 compared to existing state-of-the-art (SOTA) methods.