Deep Learning-based Decoding For LDPC Codes

In order to improve the decoding performance of Belief Propagation(BP)algorithm for LDPC codes with short block length under impulsive channel and correlated Gaussian channel,deep learning-based decoding for LDPC codes is presented in this thesis.The specific research contents are as follows:(1)In the Symmetric alpha-stable(S?S)noise model,first,two deep neural network models are constructed by Tanner graphs,where weights of the edges are reassigned to improve the decoding performance,and channel adaptor is used to solve the problem of dimension explosion of training data when using high order modulation mode.Then,to obtain optimal parameters for approximate calculation of LLR of channel outputs,the BP decoder robust to parameter ? is constructed.Finally,a robust training set is constructed to obtain the BP decoder robust to parameters ? and ?.At high code rate,the decoding performance of the BP decoder is effectively improved over traditional BP algorithm.Furthermore,when using the approximate calculation method,the decoding performance is robust over S?S noise models with different parameters.(2)In the BG noise model,first,a BP decoder based on deep learning is constructed by using a recurrent neural network(RNN).Then the clipping processor and the deep BP decoder are optimized together,simplifying the calculation method of the channel LLR into linear calculation,and thresholds that fixed under different Signal Noise Ratio(SNR)values are obtained.Finally,based on the BP decoder robust to parameter H,the threshold that fixed under different H and SNR values is obtained.When using the simplified calculation method of channel LLR,the BP decoder can significantly improve the decoding performance.(3)In the correlated Gaussian channel,residual learning is used to construct residual denoiser,and the combination of denoising and decoding is used to construct denoising BP decoder.Firstly,MLP,LSTM and CNN were used to construct denoiser.Then,the LDPC codes are decoded under different correlation coefficients.The performance of the three denoising BP decoders is improved respectively over traditional BP algorithm.Finally,the SNR and distribution characteristics of the transmitted information are analyzed in detail.