摘要： 5G规模化商用可提供高速低延时的通信服务，但由于逐块设计的通信模型可解释性程度较低，增加了其物理层优化的复杂度。针对该问题，利用深度学习在结构化信息表示和数据提取上的优势，在其基础上提出一种自编码器端到端物理层优化方案。通过两阶段训练模式提高神经网络的泛化性，同时利用自编码器压缩特性量化信道状态信息（CSI）并进行重建，降低CSI反馈导致的系统开销。仿真结果表明，该方案通过分阶段训练能有效提升收敛速率，而压缩量化CSI则可缓解系统负载。

关键词: 深度学习, 神经网络, 自编码器, 物理层优化, 压缩感知

Abstract: Large-scale commercial use of 5G will enable high-speed low-latency communication services.However,the physical layer optimization gets complicated by the limited interpretability of the modular communication model.To address the problem,this paper proposes an end-to-end physical layer optimization scheme using auto-encoder based on deep learning,as deep learning has endogenous advantages in structured information representation and data extraction.The scheme adopts a two-phase training mode to improve the generalization of neural network,and uses the compression features of auto-encoders to quantify Channel State Information(CSI) for reconstruction,so as to reduce the system cost generated by CSI responses.Simulation results show that the phased training by the proposed scheme can effectively increase the convergence rate,and the compression of quantified CSI can reduce system loads.

Key words: deep learning, neural network, auto-encoder, physical layer optimization, compressed sensing

中图分类号: 

• TP391