Design And Implementation Of Polar Codes Decoder Based On Neural Network

Polar codes were first proposed by Professor Ar(?)kan of Bilkeng University in Turkey in 2009 as a theoretically proven coding scheme that can reach the channel capacity.For the decoding scheme of polar codes,there are two mainstream algorithms,Successive Cancellation(SC)decoding algorithm and Belief Propagation(BP)decoding algorithm,but both of them are insufficient to meet the requirements of next-generation communication system due to the limited decoding performance and latency.Therefore,we need a new decoding scheme.At the same time,with the rapid development of Deep Neural Network(DNN),neural networks have made breakthrough achievements in many fields,which have attracted widespread attention.In view of the channel decoding problem can be regarded as a classification problem,it is of great significance to apply neural network theory to the decoding of polar codes.This thesis combines the neural network model with decoding of polar codes,implements different neural network decoders for polar codes,further analyzes the influencing factors of neural network decoding,and finally designs two neural network decoders to improve decoding performance.The main contents are as follows:Based on the traditional neural network decoding architecture,The Multilayer Perceptron(MLP)decoder,Convolutional Neural Network(CNN)decoder,Long Short Time Memory(LSTM)decoder are implemented,The Gated Recurrent Unit(GRU)decoder and Residual Networks(Res Net)decoder are designed.In the two cases of noise and noiseless,five neural network decoders are used to simulate the decoding performance of polar codes with code lengths of 8,16,and 32 respectively.Though the analysis,it is found that each neural network decoder has a saturation code length due to its limited learning ability.For MLP and CNN decoders,the saturation code length is 16.For LSTM,GRU,and Rest Net,the saturation code length is 32.In the case of noise,polar codes can be effectively decoded when the code length is less than the saturated code length and the proportion of the training set is 100%,and the Bite Error Rate(BER)is close to MAP decoding.Compared with the delay of traditional decoding algorithms,the neural network decoder not only has well decoding performance,but also has advantages in decoding delay compared with traditional SC and BP decoding.Based on the generalization ability of neural network in the case of noiseless,an improved MLP-GRU decoder is proposed.MLP-GRU decoder successfully improves the decoding performance of GRU decoder with code length of 32,and it can approach the performance of MAP decoding when the training set accounts for a relatively high proportion.MLP-GRU decoding model improves the generalization ability of GRU decoder with noise,and the decoding rules can be learned only by learning small-scale data accounting for 9% of the total data set,which makes it possible for neural network to decode longer polar codes accurately and quickly.Next,in order to improve the performance of the SC decoder in the channel with coherent noise,a concatenated GRU-SC decoder is designed to combine the neural network with the traditional decoding algorithm.We compare the decoding performance of GRU-SC decoder with the SC decoder with code length of 64 under four different coherence conditions of noise.With the biggest performance gap of the coherence of noise between the two decoders,GRU-SC decoder has about 2d B performance gain compared with the SC decoder when BER is 0.02.GRU-SC decoder breaks through the dimension limitation and significantly improves the decoding performance with the coherent noise.