Improving the Performance of R17 Type-II Codebook with Deep Learning

Ke Ma, Yiliang Sang, Ming Yang, Lian Jin, Chang Tian, Zhaocheng WangarXiv 2023 · DOI · arXiv

Abstract

The Type-II codebook in Release 17 (R17) exploits the angular-delay-domain partial reciprocity between uplink and downlink channels to select part of angular-delay-domain ports for measuring and feeding back the downlink channel state information (CSI), where the performance of existing deep learning enhanced CSI feedback methods is limited due to the deficiency of sparse structures. To address this issue, we propose two new perspectives of adopting deep learning to improve the R17 Type-II codebook. Firstly, considering the low signal-to-noise ratio of uplink channels, deep learning is utilized to accurately select the dominant angular-delay-domain ports, where the focal loss is harnessed to solve the class imbalance problem. Secondly, we propose to adopt deep learning to reconstruct the downlink CSI based on the feedback of the R17 Type-II codebook at the base station, where the information of sparse structures can be effectively leveraged. Besides, a weighted shortcut module is designed to facilitate the accurate reconstruction. Simulation results demonstrate that our proposed methods could improve the sum rate performance compared with its traditional R17 Type-II codebook and deep learning benchmarks.

Introduction

In massive MIMO systems, accurate Downlink Channel State Information (CSI) is essential for effective beamforming. In Frequency Division Duplex (FDD) systems, the user equipment (UE) measures the channel and feeds this information back to the base station (BS). However, the feedback overhead scales with the number of antennas, consuming significant bandwidth. The 3GPP Release 17 (R17) Type-II codebook addresses this by leveraging partial reciprocity—the similarity between uplink and downlink channels in the angular-delay domain. The BS uses uplink signals to select specific "ports" for the UE to measure.

Despite its efficiency, the R17 Type-II codebook faces two primary challenges:

  1. Noisy Port Selection: Low uplink Signal-to-Noise Ratio (SNR) often leads to suboptimal port selection.
  2. Quantization Errors: Feedback compression introduces errors that traditional deep learning models struggle to rectify due to the sparse nature of the data.

Proposed Methodology

We propose two deep learning-based enhancements implemented at the BS to improve performance without increasing the UE's computational load.

Deep Learning-Based Port Selection

A Convolutional Neural Network (CNN) is employed to identify dominant ports from noisy uplink CSI. To better represent the physical characteristics of the signals, we utilize circular padding instead of standard zero-padding, respecting the cyclic nature of Discrete Fourier Transform (DFT) bases. Furthermore, Focal Loss is applied to address the class imbalance problem, as only a small fraction of ports are dominant. This forces the model to focus on "hard-to-classify" ports rather than being biased toward the majority of low-power ports.

Proposed deep learning architecture for port selection.

Deep Learning-Based CSI Reconstruction

To mitigate quantization errors, we use a CNN-based reconstruction module at the BS. This module employs a weighted shortcut approach, focusing on learning the residual (the difference between the standard codebook reconstruction and the perfect channel). A small weighting coefficient ($w$) is used to facilitate convergence and capture fine-grained details.

Experimental Setup

The performance was evaluated using synthetic data generated from the 3GPP 38.901 clustered delay line (CDL) channel model for Urban Macro-cell (UMa) scenarios. The setup included a single BS with 32 antennas and 5 single-antenna UEs. The training and validation sets consisted of 102,400 and 2,560 samples, respectively.

Results and Discussion

The proposed deep learning methods demonstrate significant improvements over traditional benchmarks.

Sum Rate Performance

As shown in the simulation results, the proposed deep learning-based port selection significantly outperforms the traditional Type-II codebook, especially in low SNR regimes. At an uplink SNR of -5 dB, the proposed method achieves performance close to that of the Type-II codebook with perfect uplink CSI.

Sum rate performance comparison across different uplink SNR levels.

Comparison with Benchmarks

Our reconstruction method also outperformed EVCsiNet, a popular deep learning benchmark. By preserving the sparse structure inherent in the R17 codebook, our approach avoids the "smearing" effects often seen in standard autoencoder-based reconstructions.

Conclusion

Integrating deep learning into the R17 Type-II codebook framework provides a robust solution for CSI feedback in massive MIMO systems. By utilizing circular padding, focal loss, and weighted residual learning, we achieve higher sum rates and better robustness to noise. Future work will explore the trade-offs between model complexity and real-time latency, as well as generalization across diverse deployment environments.

Summary updated

Abstract

The Type-II codebook in Release 17 (R17) exploits the angular-delay-domain partial reciprocity between uplink and downlink channels to select part of angular-delay-domain ports for measuring and feeding back the downlink channel state information (CSI), where the performance of existing deep learning enhanced CSI feedback methods is limited due to the deficiency of sparse structures. To address this issue, we propose two new perspectives of adopting deep learning to improve the R17 Type-II codebook. Firstly, considering the low signal-to-noise ratio of uplink channels, deep learning is utilized to accurately select the dominant angular-delay-domain ports, where the focal loss is harnessed to solve the class imbalance problem. Secondly, we propose to adopt deep learning to reconstruct the downlink CSI based on the feedback of the R17 Type-II codebook at the base station, where the information of sparse structures can be effectively leveraged. Besides, a weighted shortcut module is designed to facilitate the accurate reconstruction. Simulation results demonstrate that our proposed methods could improve the sum rate performance compared with its traditional R17 Type-II codebook and deep learning benchmarks.
BibTeX
@misc{ma2023improving,
  author = {Ke Ma and Yiliang Sang and Ming Yang and Lian Jin and Chang Tian and Zhaocheng Wang},
  title = {{Improving the Performance of R17 Type-II Codebook with Deep Learning}},
  year = {2023},
  eprint = {2310.05962},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2310.05962},
}
Figure