Study On Unsupervised Learning For Polar Code Decoding Based On Syndrome Loss Function

With the advent of the 5G era,there is a growing demand for high-speed,large-capacity,low-latency communications.Polar code is the first error correction code that has been proven to reach the Shannon limit of the communication channel.With its excellent performance,it has been determined as a coding scheme for 5G enhanced mobile broadband scene control channels.In recent years,deep learning has shown a strong ability to handle complex tasks,and has shown good prospects in the field of digital communications(channel coding and decoding,signal detection and classification,millimeter wave communication,end-to-end communication).The current decoding algorithms of polarization codes mainly include Belief Propagation(BP)decoding algorithms and Serial Cancellation List(SCL)decoding algorithms.Among them,the decoding error rate of the SCL decoding algorithm is lower and the delay is higher,while the decoding error rate of the BP decoding algorithm is higher but has a good low delay characteristic.In order to further improve the decoding error rate of the BP decoding algorithm,this paper mainly studies the decoding problem of polarization codes based on deep learning,explores the related performance of polarization code decoding based on deep learning,and improves the confidence propagation decoding algorithm The decoding error rate performance.This paper mainly studies the problem of polarized code decoding algorithm based on deep learning.Explore the performance of polar code decoding based on deep learning,and improves the Belief Propagation(BP)decoding algorithm on the decoding bit error rate.The main work of this article includes:Firstly,a neural network decoder is designed based on the existing deep learning models(MLP model,CNN model and LSTM model).Experimental research is carried out from the aspects of bit error rate,frame error rate,normalized verification error,generalization ability,and code length tolerance.The results show that the existing deep learning models approach the maximum a posteriori probability(MAP)decoding performance in the case of extremely short code lengths.As the code length increases,the training sample set grows exponentially,and existing deep learning models cannot complete decoding tasks.Then,based on the coding structure of the polar code,a codeword expansion method is proposed,which improves the decoding performance of existing deep learning models.The proposed method can extract the hidden information of the coding structure.Experimental results show that the method can reach the decoding bit error rate of the belief propagation algorithm when the code length is 16.Finally,based on the existing syndrome function,the frozen code check matrix of polarized codes is studied.An unsupervised learning polar code decoding algorithm based on the syndrome loss function is proposed.This method is suitable for polar code decoding scenarios with arbitrary code lengths.Simulation results show that the proposed algorithm can effectively reduce the decoding error rate and save about 0.2d B compared with the traditional belief propagation decoding algorithm.In addition,the algorithm proposed in this paper extends the training method of polar code decoding model based on deep learning.