Research On Channel Decoding Based On Deep Learning

The diversification and standardization of communication services have put forward higher requirements for the reliability of information transmission.Channel encoding is an important guarantee of system transmission accuracy,among which Low Density Parity Check(LDPC)code and Turbo code are two key channel encoding technologies in the field of communication.In order to reduce bit error rate and decoding time,deep learning is introduced into channel decoding in this paper.According to the structural characteristics of neural network and different application scenarios,LDPC decoder based on threshold cycle Unit(Gated Recurrent Unit,GRU)network and turbo decoder based on Long and short-term Memory(Long Short-Term Memory,LSTM)network are designed and implemented respectively.Firstly,for LDPC code,this paper builds a neural network decoder based on GRU network,introduces in detail the design process of GRU decoder and the steps of updating the weight parameters of the time back propagation algorithm.Then,the influence of the selection of unit number,learning rate,Dropout retention ratio and loss function on the bit error rate performance of GRU decoder is discussed through simulation experiments.Simulation results show that,compared with the traditional sum-product Algorithm(SPA),the proposed decoder requires shorter decoding time and higher decoding efficiency when achieving the same bit error rate performance.With the increase of data volume,the performance advantages of low bit error rate and low delay of neural network decoding are more significant.Secondly,a neural network decoder based on LSTM network is proposed to solve the problem of high decoding delay caused by poor parallelism of existing decoding algorithms of Turbo codes.The powerful parallel computation and feature learning ability of neural networks can reduce the decoding delay of Turbo codes and make it obtain lower bit error rate.This Turbo decoder references code component encoding idea.First,each component decoder is designed based on LSTM network,then each layer of component decoder is pre trained,and the trained weights are loaded into the turbo code decoding network as initialization parameters,and then the turbo code decoding network is trained end-to-end,finally a complete turbo code decoder is obtained.The simulation results show that the performance of the decoder is 0.5?1.5d B better than that of traditional serial decoding algorithm in the environment of white Gaussian noise and t-distribution noise.At the same time,it shows certain generalization ability for communication systems of different types of Turbo codes,and effectively solves the problem of high delay in serial iterative decoding.