Research On Neural Network Decoding Model Of LDPC Codes

With the continuous improvement of computer computing power in recent years,deep learning and neural networks have once again become popular research areas.Deep learning and neural networks have achieved great success in many traditional fields.More and more researchers are also trying to apply deep learning and neural networks to the field of wireless communication to build smart wireless communication networks.As the standard coding of5G(fifth generation mobile communication)data channel,the decoding algorithm of lowdensity parity-check code(LDPC)based on neural network and deep learning has naturally become a research hotspot.This paper first introduces the Gallager and Mackay construction method of LDPC codes,and then illustrates the direct coding method of generator matrix,approximate lower triangle method,and analyzes the traditional bit flip algorithm of LDPC code(BF)and Log-Likelihood Ratio Belief Propagation(LLR BP).Secondly,it studies the fully connected neural network decoder for LDPC decoding,and introduces the basic structure of fully connected neural network,including activation function,loss function,forward propagation and back propagation and so on.The influence of network parameter selection,activation function and loss function selection on the decoding of LDPC codes is analyzed.The simulation results show that the fully connected neural network decoder can decode LDPC codes with shorter code lengths,and the decoding delay is less than the LLR BP algorithm.However,when the code length increases,the codeword space will increase exponentially,resulting in a fully connected neural network decoder unable to perform effective decoding.To this end,this paper analyzes the LLR BP and the Tanner graph,constructs a non-fully connected deep neural decoding network based on the LLR BP decoding algorithm,and analyzes the construction method of the network,activation function,loss function,etc.At the same time,according to the characteristics of the network itself,an improved weighted loss function is proposed.The simulation results show that the decoding network can effectively decode LDPC medium codes,and has a faster decoding convergence speed,lower decoding delay and bettter decoding performance than the LLR BP decoding algorithm.