Massive MIMO basestations, operating with frequency-division duplexing (FDD), require the users to feedback their channel state information (CSI) in order to design the precoding matrices. Given the powerful capabilities of deep neural networks in learning quantization codebooks, utilizing these networks in compressing the channels and reducing the massive MIMO CSI feedback overhead has recently gained increased interest. Learning one model, however, for the full cell or sector may not be optimal as the channel distribution could change significantly from one \textit{zone} (an area or region) to another. In this letter, we introduce the concept of \textit{zone-specific} CSI feedback. By partitioning the site space into multiple channel zones, the underlying channel distribution can be efficiently leveraged to reduce the CSI feedback. This concept leverages the implicit or explicit user position information to select the right zone-specific model and its parameters. To facilitate the evaluation of associated overhead, we introduce two novel metrics named \textit{model parameters transmission rate} (MPTR) and \textit{model parameters update rate} (MPUR). They jointly provide important insights and guidance for the system design and deployment. Simulation results show that significant gains could be achieved by the proposed framework. For example, using the large-scale Boston downtown scenario of DeepMIMO, the proposed zone-specific CSI feedback approach can on average achieve around 6dB NMSE gain compared to the other solutions, while keeping the same model complexity.

翻译：在采用频分双工（FDD）模式运行的大规模MIMO基站中，用户需反馈信道状态信息（CSI）以设计预编码矩阵。鉴于深度神经网络在学习量化码本方面的强大能力，利用这些网络压缩信道并降低大规模MIMO CSI反馈开销的研究近期日益受到关注。然而，为整个小区或扇区训练单一模型可能并非最优方案，因为信道分布从某个“区段”（区域或地带）到另一个区段可能发生显著变化。本文提出“区段特定”CSI反馈的概念。通过将站点空间划分为多个信道区段，可有效利用底层信道分布特性来降低CSI反馈开销。该概念利用隐式或显式的用户位置信息，选择相应的区段特定模型及其参数。为便于评估相关开销，我们引入了两个新型指标：模型参数传输速率（MPTR）和模型参数更新速率（MPUR）。两者共同为系统设计与部署提供重要洞察和指导。仿真结果表明，所提框架可实现显著性能提升。例如，在DeepMIMO的大规模波士顿城区场景中，所提出的区段特定CSI反馈方法在保持相同模型复杂度的情况下，相比其他方案平均可获得约6dB的NMSE增益。