LDPC Decoding Of Low Error Floor Based On Machine Learning

Low density parity check code has the characteristics of low density parity check matrix and approaching Shannon limit.It has been widely used in the field of communication and storage.As a low complexity iterative algorithm,belief propagation(BP)decoding plays an important role in the large-scale commercial use of LDPC codes.However,due to the local information processing characteristics of BP algorithm,if there is a trapping set in Tanner graph of LDPC code,the error floor phenomenon will occur in the high SNR region during the iterative decoding process of belief propagation.The emergence of error floor has seriously restricted the further development of LDPC code error correction performance,so the research on decoding algorithm with better performance and moderate complexity has high theoretical application value.In recent years,the field of artificial intelligence and deep learning has been booming,and has made great success in image classification,natural language processing,optimization algorithm and so on.Similarly,deep learning is also introduced into the field of communication system as a new technology.In the aspect of decoding,some scholars have put forward neural network decoder,which transforms belief propagation between variable nodes and check nodes in Tanner graph into an equivalent depth network model.Through training,the optimal parameters of message propagation processing is obtained,and the great decoding performance is achieved.Due to the excessive network parameters,the existing neural network decoder is only suitable for the decoding of medium or short length error correcting codes.How to reduce the number of training parameters and the computational complexity of neural network decoder,so that it can be applied to LDPC codes with a code length of more than 1000,is an important issue in academic and industrial circles,and also the focus of this paper.This paper points out that the reduction of traditional neural network decoder can be studied from two aspects: reducing the number of hidden layers and reducing the number of parameters.According to the weight parameter value distribution of traditional neural network decoder,this paper proposes two low complexity neural network decoders:sparse weight neural network decoder and sharing weight neural network decoder.Both of them reduce the number of redundant parameters by different algorithmic ideas,and thereduction of parameters is close to or more than 50%.Two algorithm have their own advantages.The sparse weight decoder is more flexible in the field of short and medium code,while the sharing weight neural network decoder is more versatile and can be used in the research of long code.Aiming at the error floor problem of BP algorithm in LDPC code decoding process,the traditional two-stage decoding algorithm for optimizing trapping set problem has the problems of slow convergence speed,instability of stop criterion and empiricism of parameter selection,instability and suboptimal performance.In this paper,a low complexity two-stage neural network decoding algorithm is proposed.In the first stage of the algorithm,traditional belief decoding is used,and messages which fail to decode in the first stage will enter the second stage of neural network decoder to post-processing.The simulation results show that this algorithm has better decoding performance than the conventional BP algorithm and improved two-stage selective BP decoding algorithm in AWGN channel.Compared with the same type of deep learning decoding algorithm,the algorithm not only reduces the model complexity in terms of the number of parameters,but also reduces the scale of neural network in terms of the number of hidden layers,so it achieves lower model complexity in the same number of iterations.