Research On Deep Learning Based Decoding Of Medium-long LDPC Codes

In recent years,LDPC codes have been adopted by many industry standards,such as WLAN(IEEE 802.11n)and Wi MAX(IEEE 802.16e),etc.,and are used as a key component of the 5G standard to improve the reliability of information exchange.With the rapid development of deep learning technology and the rapid increase of modern communication network complexity which results in the appearance of traditional communication methods’ limitations,a large number of researchers have adopted deep learning methods in communication systems and achieved many related results.However,most of the existing researches combined with deep learning focus on short-medium LDPC codes,which will cause dimension explosion with long code length,and do not well solve the problem of falling into the trapping set in the decoding process of LDPC codes in high SNR region.In this paper,taking binary LDPC codes as the research object,combining the strong learning ability of deep learning and the advantages of traditional LDPC decoding methods,two kinds of decoding algorithms based on deep learning are proposed: one is to implement the traditional decoding algorithm through deep learning model,and the other is to solve the trapping set problem and further improve the decoding performance of LDPC codes through deep learning on the basis of the traditional decoding algorithm of LDPC codes.The main contents of this paper are as follows:(1)In this paper,a decoder suitable for LDPC long codes is proposed.Combined with the characteristics of the Tanner graph of LDPC codes,a depth neural network is established to learn the conventional normalized min-sum algorithm.In addition,in order to further reduce the complexity of the algorithm implementation,two improved decoders are proposed in this paper,that is neural normalized min-sum decoder with weight sharing and neural offset minsum decoder.In addition,LDPC code decoder with a structure of relaxation is proposed in this paper to further improve the decoding performance.Simulation results show that the proposed decoding algorithm achieves the decoding performance comparable to or even better than the traditional decoding algorithm,and the decoding convergence speed is faster and the delay is lower.(2)A deep learning-aided post-processing decoding algorithm based on the conventional normalized min-sum decoding algorithm is also proposed in this paper.The algorithm performs deep learning based bit flip two-stage decoding for error frame codewords,which aims to further reduce the frame error rate of LDPC codes on the premise of ensuring the excellent decoding performance of conventional iterative decoding algorithms,thus the trapping set problem in the high SNR region can be solved to some extent.Simulation results show that,compared with the conventional normalized decoding algorithm,the DLAPPS decoder can significantly improve the FER of the codeword and the decoding gain,while the increased complexity of the DLAPPS decoding algorithm is negligible compared with the decoding gain.When the model parameters are trained,the DLAPPS decoder has a lower delay than the existing methods to solve the trapping set problem.(3)The above two LDPC decoding methods based on deep learning proposed in this paper are suitable for long LDPC codes and solve the problem of dimension explosion in most of the existing methods.The simulation results show that the performance of the two methods is better than the conventional min-sum decoding of LDPC codes,so it can be considered to be used in practical applications.