相关噪声下基于深度学习的LDPC码联合降噪译码算法设计  

Joint Denoising and Decoding of LDPC Codes Under Correlated Noise Based on Deep Learning

作　　者：杨恩鑫 袁磊[1] 郭毅[2] 岳新东 YANG Enxin;YUAN Lei;GUO Yi;YUE Xindong(School of Information Science and Engineering,Lanzhou 730000,China;Xi'an Institute of Optics and Precision Mechanics of CAS,Xi'an 710119,China;Gansu Radio Monitoring Station,Lanzhou 730000,China)

机构地区：[1]兰州大学信息科学与工程学院,甘肃兰州730000 [2]中国科学院西安光学精密机械研究所,陕西西安710119 [3]甘肃省无线电监测站,甘肃兰州730000

出　　处：《移动通信》2024年第5期83-88,共6页Mobile Communications

基　　金：甘肃省科技计划资助项目“智能反射面辅助的超可靠低延时通信研究”(22JR5RA490)。

摘　　要：为了改善采用低密度奇偶校验(LDPC)码和高阶调制的无线通信系统在相关噪声下的译码性能,提出了一种联合优化降噪和译码的深度学习算法。降噪器中采用了残余收缩模块(RSBU),译码器采用了基于循环神经网络的神经网络最小和译码算法。在提出的联合降噪译码(JDD)算法中,利用复数神经网络在处理复信号方面比实数神经网络更有优势的特点,提出了一个复数RSCNN(CRSCNN),接收的复信号直接输入CRSCNN并利用新颖的联合优化降噪损失函数和译码损失函数的多任务学习策略来改善译码性能。仿真结果显示基于CRSCNN的JDD算法获得了比基于循环神经网络的神经网络最小和译码算法更好的译码性能。To enhance the decoding performance of wireless communication systems utilizing low-density parity-check codes and high-order modulation in correlated noise environments,we propose a deep learning-based joint denoising and decoding algorithm.The denoiser incorporates a residual shrinkage building unit,and the decoder employs a neural network-based min-sum decoding algorithm using recurrent neural networks(RNNs).In our joint denoising and decoding(JDD)approach,a complex residual shrinkage convolutional neural network(CRSCNN)capitalizes on the superiority of complex neural networks over real-valued networks for processing complex signals.The received complex signals are directly fed into the CRSCNN,utilizing a novel multi-task learning strategy that jointly optimizes denoising and decoding loss functions to improve decoding performance.Simulation results demonstrate that the CRSCNN-based JDD algorithm achieves superior decoding performance compared to the neural network-based min-sum decoding algorithm using RNNs.

关 键 词：低密度奇偶校验码 高阶调制 相关噪声 深度学习 残余收缩模块 

分 类 号：TN92[电子电信—通信与信息系统]