| 摘要: |
| 针对大规模多输入多输出系统中信道状态信息在反馈时重构精度低、复杂度高的问题，提出了一种基于注意力机制的反馈方法。首先，考虑到信道状态信息矩阵数据分布特点，采用一种高效多尺度注意力模块提取信道状态信息矩阵局部和全局的特征，并关注重要数据点的分布，提升网络模型的特征学习能力。其次，使用增强的可重参数化的卷积替代普通的卷积核，提升卷积对于局部特征的提取能力，使整个神经网络自编码器在保持轻量化的基础上达到更高的压缩重构精度。仿真结果表明，与轻量化网络CRNet和ACRNet-1x相比，所提出的网络模型在复杂度方面分别平均降低了19%和5%，重构精度分别平均提高了3%和8%，同时展现出了更好的鲁棒性。 |
| 关键词: 大规模MIMO 信道状态信息反馈 神经网络自编码器 高效多尺度注意力 轻量化网络 |
| DOI：10.20079/j.issn.1001-893x.240516001 |
|
| 基金项目:国家自然科学基金资助项目（61931004） |
|
| A Channel State Information Feedback Method Based on Multi-scale Attention Lightweight Network |
| LIU Qingli,XIE Jiajun |
| (School of Information Engineering,Dalian University,Dalian 116622,China) |
| Abstract: |
| In order to address the issues of low reconstruction accuracy and high complexity in channel state information(CSI)feedback for large-scale multiple-input multiple-output systems,a feedback method based on attention mechanisms is proposed.The method initially considers the characteristics of the data distribution in the channel state information matrix,employs an efficient multi-scale attention module to extract both local and global features of the channel state information matrix,and focuses on the distribution of important data points to enhance the feature learning capacity of the network model.Secondly,the enhanced re-parameterizable convolution is used to replace the ordinary convolution kernel in order to improve the convolution’s ability to extract local features.This allows the entire neural network autoencoder to remain lightweight,while achieving higher compression and reconstruction accuracy.The simulation results demonstrate that the proposed network model exhibits a reduction in complexity by an average of 19 |
| Key words: massive multiple-input-multiple-output channel state information feedback neural network autoencoder efficient multi-scale attention lightweight network |
