Research And Implementation Of The Polar Code Denoising-Decoding Algorithm Based On Deep Learning ResNetXt

The fifth generation of mobile communication(5G)is a mobile communication system nowadays,which integrates communications,semiconductors,intelligent hardware,artificial intelligence,and new applications and services.Channel encoding and decoding is a crucial technology in mobile communication systems.This technology adds redundant check bits to the transmitted information sequence,and corrects errors through decoding at the receiving end,thereby improving the reliability of information transmission and realizing the correct reception of the transmitted information sequence.Polar code is the final control channel coding scheme of the 5G.This code is the first structural code to reach the capacity of Shannon's channel.It has excellent processing capability of encoding and decoding with high reliability.However,the sequential decoding structure of the traditional Polar code's decoding method causes a large latency,which cannot meet the requirements of low-latency tasks.To reduce the latency of decoding,deep learning(DL)-based decoders have been proposed for their strong learning ability and high robustness.Meanwhile,the DL-based decoders can be trained and optimized easily.However,in the case of a low signal-to-noise ratio(SNR),the reliability of DL-based decoders is low.Therefore,in order to meet the requirements of ultra-reliability and low latency tasks at the same time,this thesis adds a denoiser based on the parallel residual network(ResNetXt)structure to the DL-based decoder,and proposes a ResNetXt denoiser-decoder,which consists of a ResNetXt-based denoiser and a DL-based decoder.The sequence to be decoded is first denoised by the ResNetXt-based denoiser,the SNR performance of the denoised sequence is improved,and then as the input of the decoder for decoding,the BER performance of the decoding is improved.ResNetXt denoiser-decoder adopts a multi-task learning(MTL)method for optimization.Denoising tasks and decoding tasks enhance each other,and the denoiser can significantly improve the performance of the decoder.The main contributions of this thesis are summarized as follows:1.In view of the large latency of traditional decoders,this thesis implements and compares the performance of four types of DL-based decoders,which are Multilayer Perceptron(MLP)-based decoders,Convolutional Neural Networks(CNN)-based decoders,Recurrent Neural Network(RNN)-based decoders,and Transformer-based decoders.The first three structures are widely used in decoding tasks,and to the best of the author's knowledge,this is the first thesis that uses the Transformer structure in decoding tasks.Compared with the traditional Successive Cancellation(SC)decoders,the above-mentioned four deep learning-based decoders can increase the decoding speed by more than seventy times with little difference in reliability.2.Aiming at the problem of low decoding reliability of DL-based decoders when the SNR is low,four types of ResNetXt-based denoiser are implemented,which are MLP-based ResNetXt denoiser,CNN-based ResNetXt denoiser,RNN-based ResNetXt denoiser and Transformer-based ResNetXt denoiser.The four ResNetXt denoisers above can significantly improve the SNR performance of information,and SNR will directly improve the BER performance of the decoder.Therefore,this thesis adds the above-mentioned ResNetXt denoiser to the DL-based decoder to form a ResNetXt denoiser-decoder.The two tasks can promote each other by using multi-task learning to optimize the denoising and decoding tasks together.Numerical simulation results show that the ResNetXt denoiser-decoder still has good decoding performance when the SNR is low,which is better than the non-parallel denoiser-decoder and the purely DL-based decoder.The performance is close to the SC decoder.3.Due to the fact that most existing DL-based decoders only consider the fixed code length,this thesis analyzes the decoding effect of ResNetXt denoiser-decoder for Polar codes with different code lengths.The simulation results show that by matching the model size for Polar codes with different code lengths,the ResNetXt denoiser-decoder can achieve good decoding results for Polar codes with different code lengths.It can replace SC decoder in delay-sensitive tasks.