Research On Polar Codes Decoding Algorithm Based On Deep Learning

Polar Codes are theoretically proven channel coding schemes that can reach the Shannon limit,and are characterised by low compile code complexity and simple construction compared to other error correction codes,but traditional decoding algorithms are more complex to model and solve.The decoding process in communication systems can be regarded as the classification of signals,and deep learning can be used for the decoding process in communication systems as it can effectively process a large amount of training data and learn relevant features from it,and can solve many non-linear and complex modelling tasks.In view of this,this thesis conducts an in-depth study on the decoding scheme combining polarized codes and deep learning,and optimizes the Successive Cancellation(SC)and Belief Propagation(BP)decoding algorithms based on deep learning respectively,the main research achievements are as follows:(1)The neural network is used to learn the structural features of the polarization code word,and the deep learning decoding scheme is designed to replace the traditional polarization code decoder.Considering that Deep Neural Networks(DNN)are suitable for efficient parallel computation and Convolutional Neural Networks(CNN)are good at extracting data features,While Long Short-Term Memory(LSTM)is suitable for processing time series signals,so the polarization code decoder models based on DNN,CNN and LSTM are designed respectively,and the Bit Error Rate(BER),Frame Error Rate(FER),normalized validation error(NVE),model generalization ability,storage overhead and computational complexity were studied.The experimental results show that the DNN decoder model not only has excellent decoding performance,but also has the lowest memory cost and computational complexity under the same parameter scale,which is more in line with the requirements of practical application scenarios of communication systems.(2)By deeply analyzing the influence of weight storage overhead and time overhead of Neural Network Decoder(NND)on the decoding algorithm of neural network SC,this thesis proposes a low-precision deep neural network successive cancellation(LP-DNN-SC)decoding algorithm.In this algorithm,firstly,the weights of a finite number of bits of the deep neural network decoder are quantized with weights to realize low-precision decoding,and all arithmetic operations are performed by quantizing the input and output of 8 bits in the Q8.4 fixed-point digital format to realize the int8 calculation.Using the block idea,deep learning decoding is performed on each polar codes sub-block,and finally the trained DNN decoders of each subblock are coupled by the traditional SC decoding algorithm.Compared with Neural Successive Cancellation(NSC)decoding algorithm,this algorithm reduces the weight of neural network decoder through weight quantization,reduces the high memory requirement of floating-point decoder,and achieves lower decoding delay.Experimental results show that compared with NSC decoding algorithm and NND,this algorithm has achieved better decoding performance gain.For the(32,16)polar codes,the number of time steps required by this algorithm is reduced by 10%compared with NSC algorithm,and by 85.4% compared with SC algorithm,and a lower decoding latency is obtained.(3)Aiming at poor block error rate(BLER)performance and high computational complexity of BP decoding algorithm,this thesis combines Offset Min-Sum(OMS)approximation algorithm and bit-flipping(BF),A Recurrent Neural Network Offset Min-Sum Belief Propagation Bitflip(RNN-OMS-BPF)decoding algorithm is proposed.In this algorithm,a Critical Set(CS)is constructed based on the codeword structure of the polar codes.The information bits are sorted in descending order according to the error rate of the corresponding polarization channel in the CS set.Then,according to the similarity between BP decoding message update factor map and neural network structure,BP decoding message update factor map was expanded on the basis of RNN architecture to construct BP decoding algorithm of recurrent neural network,and OMS was used to improve BP decoding algorithm of recurrent neural network.If the decoding results of the BP decoding algorithm are not verified by Cyclic redundancy check(CRC),candidate information bits are selected from the CS set to perform bit-flipping.The algorithm uses RNN architecture to realize parameter sharing among BP decoding iterations,reducing the number of iterations required to achieve the convergence effect.OMS is used to replace the multiplication operation of the algorithm,improving the computing resource consumption and extra memory overhead of BP decoding algorithm,and significantly improving the error correction ability of the algorithm by BF.Experimental results show that compared with CRC Aided Successive Cancellation List(CRC-SCL)decoding algorithm with L=4,the proposed algorithm has a performance gain of about 0.50 d B.Compared with the Recurrent Neural Network Belief Propagation(RNN-BP)decoding algorithm,it reduces addition operation by 50% and memory overhead by 28.7%.