SC Decoding In Association With Deep Learning Of Polar Codes

In the process of digital communication,the information will be affected by noise and interference,which will produce error code.Through the channel coding technology,the above influence and the bit error rate can be reduced.Polar Codes,a new channel coding method,have been theoretically proved to be able to reach the Shannon limit.It has the advantages of low computational complexity,fixed generation matrix,low bit error rate,and has been approved by 3GPP as the core coding technology for error correction of 5G control channel.However,the decoding performance of Polarizing Codes under short codes is not high,and the Successive Cancellation(SC)decoding algorithms developed for Polarizing Codes at present have poor performance.Therefore,in order to further improve the performance of polarized code decoding,this paper uses deep learning and other technologies to explore the polarized code decoding,and proposes a scheme combining neural network decoding with SC decoding,which provides a new idea for improving the performance of polarized code decoding.Firstly,this paper introduces the polarization code and analyzing the characteristics of the neural network algorithm,and create the polarization code contains 1.5 x 106 data of the training set and test article 2.5 x 105 data sets,on this basis,evaluate the decoding performance of feedforward neural networks,under low signal-to-noise ratio decoding performance of feedforward neural networks than SC,but under the high signal-to-noise ratio is better than that of SC.It shows that it is feasible to use neural network to decode polarization codes.On the basis of the evaluation,considering the time diversity of block codes,and combining the characteristics of LSTM(Long Short-Term Memory)chain ring structure and the ability to process sequences,a decoding scheme based on original sequences is proposed.The learning rate was set as 0.003,Dropout was added to prevent overfitting,and Adam optimizer was used to carry out the simulation test.The results show that the performance is almost the same as that of SC at low SNR,and the performance is obviously better than that of SC at medium and high SNR.In order to further improve SC decoding performance,LSTM decoding results were combined with SC decoding results,and a joint decoding algorithm was proposed.On the basis of SC decoding results,for the positions with different decoding results,the single-bit flip is carried out one by one,and the new candidate codewords are recoded.The similarity of the equivalent index of maximum likelihood is used to evaluate the reliability of all the candidate codewords,which is equivalent to the Euclidean distance between the candidate codewords and the receiver vector.The candidate codewords with the largest similarity are selected as decoding output.The simulation results show that,compared with SC decoding results,the error correction rate of the joint algorithm for the different positions of decoding results is more than98%,and the bit error rate is close to the theoretical lower bound.