SCMA-D2D混合网络卷积自编码器  

SCMA-D2D hybrid network convolutional autoencoder

作　　者：许耀华[1,2] 周鑫源 黄兴 蒋芳[1,2] 王翊[1,2] 王跃[1,2,3] XU Yaohua;ZHOU Xinyuan;HUANG Xing;JIANG Fang;WANG Yi;WANG Yue(Key Laboratory of Intelligent Computing and Signal Processing,Ministry of Education,Anhui University,Hefei 230601;The IoT Spectrum Sensing and Testing Engineering Center of Anhui University,Hefei 230601;Anhui Telecom Planning and Design Co.,Ltd.,Hefei 230031)

机构地区：[1]安徽大学智能计算与信号处理教育部重点实验室,合肥230601 [2]安徽大学物联网频谱感知与测试工程中心,合肥230601 [3]安徽电信规划设计有限责任公司,合肥230031

出　　处：《高技术通讯》2023年第12期1244-1252,共9页Chinese High Technology Letters

基　　金：安徽省高校自然学科研究项目(No.2022AH050109);国家自然科学基金(No.62071002,62001001)资助。

摘　　要：为满足物联网通信大连接、低功耗的需求,高效利用有限的频谱资源成为一项重要的挑战。在应用了稀疏码分多址接入(SCMA)技术的蜂窝网络中增加设备对设备(D2D)用户对,两者共享频谱资源,可以进一步提高频谱利用率,满足大规模连接和低功耗的通信需求。然而,当不同类型的用户共享相同的频谱资源时会导致严重的用户间干扰,导致多用户检测精度降低,译码复杂度增高。本文使用卷积神经网络(CNN)进行SCMA-D2D混合网络自编码器设计,通过端到端的联合训练,设计出合适的神经网络结构。用CNN单元实现混合网络的编码,学习SCMA蜂窝用户和D2D用户的有效码本;将混合网络的多用户检测问题建模为一个基于共享层机制的多任务分类解码问题,建立多用户分类解码器。实验结果表明,本文提出的自编码器能够生成对系统适应性更强的码本,结合接收端的多任务分类解码器能够有效地提高整个混合网络系统的误码率性能,同时减小译码计算复杂度。In order to meet the needs of massive connectivity and low power consumption in Internet of Things communication,efficient use of limited spectrum resources has become an important challenge.Adding D2D(device to device)user pairs to SCMA(sparse code multiple access)cellular network and sharing spectrum resources among them can further improve spectrum utilization and meet the communication needs of massive connectivity and low power consumption.However,serious inter-user interference will be caused when different types of users share same spectrum resources,resulting in lower accuracy of multi-user detection and higher complexity of decoding.In this paper,convolutional neural network(CNN)is used to design the SCMA-D2D hybrid network autoencoder,and a suitable neural network structure is designed through end-to-end joint training.Using CNN unit to implement hybrid network coding,learn the valid codebook of SCMA cellular users and D2D users;The multi-user detection problem of hybrid networks is modeled as a multi-task classification decoding problem based on shared layer mechanism,and a multi-user classification decoder is established.The results show that the proposed autoencoder can generate codebooks with better adaptability to the system,and combined with the multi-task classification decoder at the receiving end,it can effectively improve the bit error rate performance of the whole hybrid network system and reduce the computational complexity of decoding.

关 键 词：卷积神经网络(CNN) 稀疏码分多址接入(SCMA) 设备对设备(D2D)通信 无线通信 多用户检测 

分 类 号：TP391.44[自动化与计算机技术—计算机应用技术] TN929.5[自动化与计算机技术—计算机科学与技术]