The Research For Small Size LDPC Decoder Based On Deep Learning

The LDPC code is an error correction code whose performance can be approximated to the Shannon limit.It has been widely used for channel coding of mobile communication and error correction coding of memory.The LLR-BP algorithm is one of the LDPC soft-decision decoding algorithms.It has the characteristics of iterative decoding and decoding performance becomes better as the code length increases,but there is an upgraded space for improvement in decoding performance and decoding throughput rate of short code length.Deep learning has proven its powerful recognition,classification and fitting capabilities in applications such as speech,image,and natural language processing.Therefore,this paper combines deep learning network with LDPC decoding of medium and short code length as the main research content.In this paper,the decoding process of the received LDPC codeword received by the receiver is regarded as the process of data identification,classification and fitting.Three methods of decoding the LDPC code using the deep neural network for short and medium code length are proposed:This paper proposes a multi-classification decoding method based on fully connected neural network.This method treats codeword types as classification categories,and uses multi-join feedforward neural networks with several hidden layers and Softmax layers for multi-classification output.The method has the characteristics of simple network structure and low training difficulty.The simulation experiment proves that the decoding performance of the fully connected neural network multi-class decoding method leads the LLR-BP algorithm in LDPC short code,which the maximum can reach 1dB.This paper proposes a binary classification decoding method based on semi-supervised classification neural network.Firstly,the auto-encoder is used to realize the unsupervised extraction of features from LDPC codeword data,and then use a supervised classification networks to decode each symbol in the codeword.The method converts multi-class decoding into binary decoding,which solves the problem that the network output layer dimension of multi-class decoding is too large as the LDPC codeword grows.Simulation experiments show that the decoding performance based on semi-supervised neural network classification decoding method can exceed the LLR-BP algorithm.This paper proposes an LDPC decoding scheme based on SPNs network for fast probability inference.In this paper,according to the LDPC belief propagation iterative decoding process graph model,proposes a method of constructing deep SPNs network.The depth SPNs decoding network structure is relatively simple,and it can realize medium-length LDPC decoding with a code length of up to hundreds of bits.Simulation experiments show that the decoding performance can approach the LLR-BP algorithm.According to the LDPC decoding method using deep learning,the decoding output is directly calculated by input and network parameters.Compared with the LLR-BP decoding algorithm of LDPC,the decoding efficiency is improved because of whithout iteration process,and the decoding performance on the short code length is superior.