Study On Decoding Algorithms Of Polar Codes Based On Neural Network

5G has now officially entered the commercial stage,and its high-speed,low-latency perfor-mance indicators have also promoted the development of baseband communication technology.As a newly selected 5G coding scheme promoted by Huawei,Polar Codes was theoretically proved that can reach the Shannon limit and has excellent error performance.However,for traditional Polar Codes decoding algorithms,complicated node calculations and high-delay decoding structures are acceptable on the control channel of short codes,but they also limit the development of Polar Codes to data channels of long codes.In addition,the diversification of5G application scenarios also makes the channel noise model more complicated.The decoding design that only considers white noise in the decoding process will bring a large performance loss,and the traditional whitening filter processing method is not only complicated in struc-ture.And the performance is related to the parameter estimation accuracy,which is not stable enough.Therefore,in this paper,with the popular deep learning technology,the neural network is introduced into the channel decoding process,and some research results have been achieved for the above two problems.First,this paper studies the Polar Codes decoding algorithm based on neural network.For general neural network decoder designs,although it has the low-latency characteristic and a close to the maximum a probability(MAP)decoding performance,but its network design is limited by dimensions,that is,the network training complexity increases exponentially with the length of the information bits.To this end,we propose a partition-based BP-NN decoding algorithm.We divide the decoding process into several sub-blocks based on the recursive structure of the Polar Codes,and design the corresponding neural network decoder for each sub-block,and then use the BP algorithm to implement the information interaction between the sub-blocks.At the same time,we also designed a Weight function to measure the reliability of information during interaction.The simulation results show that compared with the general neural network decoder,the proposed scheme greatly reduces the amount of network parameters and training difficulty,and the bit error rate(BER)performance is basically consistent with the BP algorithm.In terms of decoding delay,the delay overhead varies with the size of the neural network part of the decoder.When the sub-block size is 32,the decoding delay is reduced by about 50%compared with the traditional BP.In addition,we also propose a BP-NN optimization algorithm that terminates iteratively in advance,which further reduces the number of iterations of the scheme under high signal-to-noise ratio.Then we studied the decoding structure optimization method in the correlated noise en-vironment.We borrowed the successful application of the feed-forward denoising network DnCNN in the field of image denoising,introduced it to the signal processing,and proposed a DnCNN noise processing Decoding structure.In the scheme,we divide the decoding process into two steps.The first step is to estimate the correlated noise.The decoding result of the first decoder is used to estimate the correlated noisen to obtain the estimated value?n.Then the DnCNN network is used to eliminate the estimation errors that do not meet the correla-tion constraints,and the accurate estimation value?n is obtained.The second step uses?n to denoise the original signal and then perform final decoding.In the design of the loss function of the DnCNN network,we introduce Jarque-Bera test based on the traditional mean square error function to make the residual noise more consistent with the Gaussian distribution.We performed simulations under the standard correlation noise model and the color noise model.The results show that the scheme has good performance gain under different models,At BER of 10^-3,a gain of 0.8-1.5d B can be obtained.At the same time,compared with the traditional whitening filter scheme,the scheme has better performance at low SNR.Although it is slightly insufficient at high SNR,compared with the whitening filter scheme,our proposed scheme does not require additional operations such as parameter estimation and signal pre-coloring,and the system overhead is smaller.Finally,we summarized the above work,analyzed the deficiencies in the current stage results,and looked forward to future research.