基于深度学习的Turbo码译码研究  

Research on Turbo Code Decoding Based on Deep Learning

作　　者：朱进蓉 苟明亮 秦明伟[1,2] ZHU Jinrong;GOU Mingliang;QIN Mingwei(School of Information Engineering,Southwest University of Science and Technology,Mianyang 621000,China;Key Robot Laboratory of Sichuan Province,Southwest University of Science and Technology,Mianyang 621000,China)

机构地区：[1]西南科技大学信息工程学院,四川绵阳621000 [2]西南科技大学四川省重点机器人实验室,四川绵阳621000

出　　处：《自动化仪表》2021年第6期34-38,共5页Process Automation Instrumentation

基　　金：四川省科技计划基金资助项目(2019YJ0309)。

摘　　要：译码器通常被设计用于高斯信道,对于非高斯信道而言却是次优的。深度学习方法为设计译码器提供了一种新的方法,可以对任意信道的数据进行训练和学习。为了克服传统模式存在的弊端,提出了一种新的基于卷积神经网络(CNN)的Turbo码译码方案。通过搭建神经译码网络模型,采用交叉熵函数(BCE)作为损失函数,对生成的编码数据进行网络训练,使得训练网络能够更好地提取特征关系,从而拟合出Turbo码的最佳译码函数。试验结果表明:基于卷积神经译码网络在加性高斯白噪声(AWGN)信道上和标准的Turbo解码器性能接近,但在非AWGN信道上却表现出相比于标准译码器更好的适应性和鲁棒性,适用于任意信道。此外,卷积神经译码网络还可以用于低密度奇偶校验(LDPC)码和Polar码的译码,有利于提出更多端到端训练有素的神经译码器。The decoder is usually designed for Gaussian channels,but it is sub-optimal for non-Gaussian channels.The deep learning method provides a new method for designing decoders,which can be used to train and learn for any channel data.In order to overcome the disadvantages of the traditional model,a new Turbo code decoding scheme based on convolutional neural network(CNN)is proposed.By building a neural decoding network model,it uses binary cross entropy(BCE)as the loss function to do network training for the generated decode data.The training network can better extract feature relationships and fit the best decoding function of Turbo code.The experimental results show that the performance of the convolutional neural decoding network is closing to that of the standard Turbo decoder on the additive white Gaussian noise(AWGN)channel,but it shows better adaptability and robustness than the standard decoder on the non-AWGN channel.It is suitable for any channel.In addition,the convolutional neural decoding network can also be used for the decoding of low density parity check(LDPC)code and Polar codes,which is helpful to propose more end-to-end well-trained neural decoders.

关 键 词：TURBO码 标准译码器 深度学习 卷积神经网络 加性高斯白噪声 鲁棒性 适应性 

分 类 号：TH-39[机械工程]