Attention Mechanism Based Intelligent Channel Feedback for mmWave Massive MIMO Systems

Yibin Zhang, Jinlong Sun, Guan Gui, Yun Lin, Haris Gacanin, Fumiyuki Adachi, Adachi, FumiyukiarXiv 2022 · DOI · arXiv

Abstract: The potential advantages of intelligent wireless communications with millimeter wave (mmWave) and massive multiple-input multiple-output (MIMO) are based on the availability of instantaneous channel state information (CSI) at the base station (BS). However, no existence of channel reciprocity leads to the difficult acquisition of accurate CSI at the BS in frequency division duplex (FDD) systems. Many researchers explored effective architectures based on deep learning (DL) to solve this problem and proved the success of DL-based solutions. However, existing schemes focused on the acquisition of complete CSI while ignoring the beamforming and precoding operations. In this paper, we propose an intelligent channel feedback architecture using eigenmatrix and eigenvector feedback neural network (EMEVNet). With the help of the attention mechanism, the proposed EMEVNet can be considered as a dual channel auto-encoder, which is able to jointly encode the eigenmatrix and eigenvector into codewords. Simulation results show great performance improvement and robustness with extremely low overhead of the proposed EMEVNet method compared with the traditional DL-based CSI feedback methods.

Introduction

The evolution towards 6G wireless communication systems relies heavily on millimeter wave (mmWave) and massive multiple-input multiple-output (MIMO) technologies. These technologies enable high-throughput and low-latency communication by utilizing large antenna arrays at the base station (BS) to perform precise beamforming. A critical requirement for effective beamforming is the availability of accurate downlink Channel State Information (CSI) at the BS. In Frequency-Division Duplex (FDD) systems, the lack of channel reciprocity necessitates that the User Equipment (UE) estimates the downlink channel and feeds this information back to the BS. However, as the number of antennas increases, the dimensionality of the CSI matrix grows, leading to a "feedback storm" that consumes significant uplink bandwidth.

The EMEVNet Architecture

To mitigate this overhead, we propose the Eigenmatrix and Eigenvector feedback neural Network (EMEVNet). The fundamental insight behind EMEVNet is that the BS does not require the full CSI matrix for beamforming; rather, it needs the principal spatial components and gains.

Overall architecture of the EMEVNet system showing the dual-channel feedback flow.

The EMEVNet workflow begins with Singular Value Decomposition (SVD) at the UE. The channel matrix $H$ is decomposed into an eigenmatrix $V$ and an eigenvector $S$. $V$ contains the spatial information necessary for precoding, while $S$ represents the channel gains. By feeding back these components instead of the raw matrix, the system focuses on the most relevant information.

Furthermore, EMEVNet incorporates a Channel Identification Network (EMEV-IdNet) that classifies the propagation environment into specific Clustered Delay Line (CDL) models (e.g., CDL-A through CDL-E). This classification allows the selection of an auto-encoder specifically trained for either Line-of-Sight (LOS) or Non-Line-of-Sight (NLOS) conditions, enhancing compression efficiency.

Dual-Channel Attention Mechanism

The core of the EMEVNet is a dual-channel auto-encoder. The encoder processes the eigenmatrix $V$ through a 3D Convolutional (Conv3D) path to handle its complex-valued nature, while the eigenvector $S$ is processed via a 2D Convolutional (Conv2D) path. These features are then fused in a transcoding module utilizing a residual attention mechanism. This attention mechanism allows the network to dynamically prioritize the most critical features of the heterogeneous input data during compression into a low-dimensional codeword.

Training and Optimization

The training process for EMEVNet is designed to minimize the reconstruction error of both the eigenmatrix and the eigenvector. As detailed in Algorithm 2, the network parameters are optimized using the Adam optimizer based on a joint loss function that accounts for the differences between the original and reconstructed components ($̂{V}$ and $̂{S}$).

Algorithm 2: The proposed EMEVNet training procedure.

The training includes an adaptive learning rate strategy to ensure convergence to a global minimum, with the loss function calculated as the expectation of the norms of the reconstruction errors for both $V$ and $S$.

Results and Discussion

Simulation results using standard 3GPP CDL models demonstrate that EMEVNet significantly outperforms the traditional CsiNet baseline. EMEVNet achieves lower Normalized Mean Squared Error (NMSE) and higher Cosine Similarity across all tested environments. Most notably, EMEVNet maintains high reconstruction quality even at extremely high compression ratios, where CsiNet's performance typically collapses. Visualizations of the reconstructed eigenmatrix show that the attention mechanism effectively preserves the diagonal "backbone" of the spatial information, which is paramount for accurate beamforming.

Conclusion

EMEVNet offers a robust and efficient framework for CSI feedback in massive MIMO systems. By combining SVD-based feature extraction with a dual-channel attention auto-encoder, it achieves superior performance with minimal feedback overhead. Future research will explore the integration of temporal correlations and the impact of hardware impairments on the feedback process.

Summary updated

Abstract

The potential advantages of intelligent wireless communications with millimeter wave (mmWave) and massive multiple-input multiple-output (MIMO) are based on the availability of instantaneous channel state information (CSI) at the base station (BS). However, no existence of channel reciprocity leads to the difficult acquisition of accurate CSI at the BS in frequency division duplex (FDD) systems. Many researchers explored effective architectures based on deep learning (DL) to solve this problem and proved the success of DL-based solutions. However, existing schemes focused on the acquisition of complete CSI while ignoring the beamforming and precoding operations. In this paper, we propose an intelligent channel feedback architecture using eigenmatrix and eigenvector feedback neural network (EMEVNet). With the help of the attention mechanism, the proposed EMEVNet can be considered as a dual channel auto-encoder, which is able to jointly encode the eigenmatrix and eigenvector into codewords. Simulation results show great performance improvement and robustness with extremely low overhead of the proposed EMEVNet method compared with the traditional DL-based CSI feedback methods.
BibTeX
@misc{zhang2022attention,
  author = {Yibin Zhang and Jinlong Sun and Guan Gui and Yun Lin and Haris Gacanin and Fumiyuki Adachi and Adachi, Fumiyuki},
  title = {{Attention Mechanism Based Intelligent Channel Feedback for mmWave Massive MIMO Systems}},
  year = {2022},
  eprint = {2208.06570},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2208.06570},
}
Figure