基于深度联合信源信道编码的CSI反馈技术

摘要：提出了一种基于深度联合信源信道编码的信道状态信息（CSI）反馈方法。该方法使用非线性编码对原始的CSI 信息进行降维，之后使用多层网络生成信道输入符号，并利用注意力机制实现了针对信道噪声的自适应功能。与现有CSI 压缩反馈方法相比，得益于信源和信道的深度联合编码，该方法可以在有限的带宽下获得更好的预编码任务的性能。此外，所提的方法可使用近似量化方法将复信道输入符号转换为有限的星座点符号，能够与现代移动通信系统有效兼容。

关键词：联合信源信道编码；深度学习；CSI 反馈；预编码

Abstract: A deep joint source-channel coding based channel state information (CSI) feedback method is proposed. The proposed method uses nonlinear coding to reduce the dimensionality of the original CSI information and a multi-layer network to generate channel input symbols, and employs an attention mechanism to realize the adaption for channel noise. Compared with existing compression-based CSI feedback methods, the proposed method can obtain better performance of the precoding task under limited bandwidth, which benefits from deep joint source-channel coding. Moreover, the proposed method can employ approximation quantization to convert infinite channel input symbols to finite quantization constellation symbols, which is effectively compatible with modern mobile communication systems.

Keywords: joint source-channel coding; deep learning; CSI feedback; precoding