Research Of Channel Decoding Based On Deep Learning

In the past few years,the popularization of 4G has promoted the rapid development of the mobile Internet industry.In the next few years,the full rollout of 5G will surely bring new development to the Internet of Things.As network speeds become faster and faster,the requirements for coding and decoding delays in the channel transmission process are getting higher and higher.In recent years,deep learning has made major breakthroughs in many fields,and is increasingly being used in various industries.It's "one-time training,inferencing everywhere" feature is very suitable for processing channel decoding tasks.Using deep learning to learn a certain coding feature in the neural network,after saving the model with this feature,you can perform the decoding task at any time,and only use the forward propagation inference decoding.In terms of using time compared to the traditional,the method also has advantages.Therefore,it is of great theoretical and practical significance to study the application of deep learning in channel decoding.This paper uses deep learning for channel decoding,explores its feasibility,and focuses on the research of various network structures in convolutional neural networks and the influence of network parameters on decoding performance.The main works of the paper is as follows:(1)Studying of the concept and related algorithms of convolutional neural networks.For the structure of common convolutional neural networks,understanding its advantages and disadvantages to reproducing the network structure in one-dimensional convolutional neural networks for decoding.(2)Generating data sets of Polar codes,BCH codes and LDPC codes for training and testing.In the process of generating data sets,doing a simple analysis of the signalto-noise ratio distribution of the training set and the testing set.Choosing the training signal-to-noise ratio distribution with the best performance of subsequent training set signal-to-noise ratio distribution.(3)Focus on designing the convolutional neural network to decode the polarization code.First,design a simple convolutional neural network to decode polarization codes,and analyze the influence of several parameters of the convolutional neural network on the decoding performance,including the size of the convolution kernel,the number of convolutional layers,and the number of fully connected layers.Then introduce some complex connection structures into the convolutional neural network for polarization code decoding,such as Inception and Res Net structures,and properly adjust the parameters according to the appropriate parameters obtained from the experiments of the simple convolutional neural network to analyze its decoding performance influences.Finally,compare the performance of various network decoding and analyze the reasons for the pros and cons of various network decoding performances.After many experiments,a more suitable neural network structure is finally designed,which can achieve good results in the decoding task of polarization codes.(4)Investigating the generalization of neural networks to decoding tasks.For BCH codes and LDPC codes,use the empirical network derived from the decoding of polarized codes to decode them,and analyze the reason why the decoding results are worse than that of polarized codes.