Scholarlink

<p>In frequency-division duplex massive multiple-input multiple-output system, the base station relies on the downlink channel state information provided by the user equipment to optimize the system performance. However, due to the sheer size of the channel matrix, direct feedback of complete channel information would result in significant overhead. To minimize overhead and enhance feedback efficiency, this paper proposes a new deep neural network architecture MRWANet, which cleverly integrates an attention mechanism with a multi-scale convolution module. The multi-scale convolution module captures key features at varying granularities of channel information, while the attention mechanism precisely assesses their significance for the downstream task, jointly enabling efficient and precise channel information feedback. Experiments based on COST2100 channel data show that MRWANet has distinct advantages compared with traditional compressive sensing methods and conventional deep learning methods. Even under a low compression ratio, it can maintain a high level of feedback accuracy.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CSI Feedback for Massive MIMO System with Joint Attention Mechanism and Multi-scale Convolution

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月20日

查看全文
  • Zeyi Li,
  • Yinnian Zhang,
  • Rong Luo

摘要

In frequency-division duplex massive multiple-input multiple-output system, the base station relies on the downlink channel state information provided by the user equipment to optimize the system performance. However, due to the sheer size of the channel matrix, direct feedback of complete channel information would result in significant overhead. To minimize overhead and enhance feedback efficiency, this paper proposes a new deep neural network architecture MRWANet, which cleverly integrates an attention mechanism with a multi-scale convolution module. The multi-scale convolution module captures key features at varying granularities of channel information, while the attention mechanism precisely assesses their significance for the downstream task, jointly enabling efficient and precise channel information feedback. Experiments based on COST2100 channel data show that MRWANet has distinct advantages compared with traditional compressive sensing methods and conventional deep learning methods. Even under a low compression ratio, it can maintain a high level of feedback accuracy.