Research On LDPC Code Decoding Algorithm Based On Deep Learning

In recent years,with the help of powerful computing power of the computer,deep learning has been increasingly improved and successfully applied in various fields of social life.In order to reduce end-to-end transmission latency and improve the accuracy of information transmission,researchers in the field of communications are dedicated to using deep learning techniques to improve the limitations of traditional communication algorithms.The decoding model integrating deep learning and low density parity check(LDPC)has received extensive attention and research,such methods are capable of achieving superior performance compared to traditional decoding algorithms.This paper does the following work around deep learning-assisted LDPC code decoding:(1)According to the min-sum decoding algorithm of LDPC code,a normalized min-sum decoding model based on deep learning and a offset minsum decoding model based on deep learning are proposed.To improve the performance of the decoding model,it is trained using both intra-layer parameter sharing and inter-layer parameter sharing.In addition,the deep learning decoding model is used in the study of related Gaussian channels to improve the problem of colored noise weakening the performance of LDPC codes.Since the LDPC code can be decoded after each iteration,this paper optimizes the loss function of the decoding model.The simulation results show that in the proposed decoding model,the offset min-sum decoding model performs better than the traditional min-sum algorithm,and better than the normalized min-sum decoding model;the offset min-sum decoding model with intra-layer parameter sharing exhibits superior decoding performance and effectively reduces the impact of noise correlation on LDPC code decoding.After optimizing the loss function,the decoding model shows improved performance compared to the original one.(2)The flooding message passing mechanism in the min-sum decoding algorithm results in lag in the update of node information,while the information transmission of the decoding algorithm in hierarchical scheduling mode is more real-time.Therefore,this paper proroses a normalized layer min-sum decoding model and an offset layer min-sum decoding model,and trains these models using intra-layer parameter sharing.The simulation results indicate that this type of decoding model exhibits superior decoding performance compared to decoding models using conventional scheduling methods.(3)To further enhance the performance of the neural network decoding model,a denoising network-assisted LDPC code decoder is proposed by using a convolutional neural network to filter the input layer data for noise reduction.In order to investigate the ability of the "denoising" network to extract relevant noise data features,the decoding performance of LDPC codes with different noise correlation coefficients is compared and analyzed,and it is concluded that the stronger the noise correlation,the stronger the ability of the "denoised" network to filter noise,and the greater the performance gain of the decoder.