Research On Deep Neural Network Based On Polar Code Decoding

In recent years,the fifth generation of mobile communication technology develops rapidly.Based on the fourth generation of communication technology,it puts forward higher requirements for mobile communication technology.Since the polarization code was determined as the short code coding scheme of 5G enhanced mobile broadband scene control channel,the polarization code has re entered the attention of researchers.As the first coding scheme proved to be able to reach the Shannon limit in theory,the research value of the decoding algorithm of polarization code is undoubted.In the case of long code length,the decoding complexity and delay of the current decoding algorithm are relatively high,which makes the use of polar code in the communication system with low delay(such as 5G urllc scenario)need to be further improved.With the improvement of hardware facilities in recent decades,the application of deep neural network is becoming more and more extensive.Because of the similarity between neural network and decoding network,compared with the existing traditional decoding methods,we try to apply the deep neural network to the polar code decoding.By combining with the traditional decoding algorithm,we can reduce the decoding complexity and delay or improve the decoding performance.In theory,neural network can effectively reduce the decoding delay and make the polar code competitive in 5g other scenarios.In this thesis,we study the decoding algorithm of polarization code based on deep neural network,and use deep neural network as decoder instead of traditional decoding algorithm.The main work of this paper is as follows:(1)The common deep neural networks are introduced: MLP,CNN,RESNET,long short term memory network Memory(LSTM)is directly decoded as a decoder.The purpose of the experiment is to explore the applicability and feasibility of various networks and the specific parameters of various networks to lay the foundation for subsequent experiments.In the experiment of neural network decoder,we adopt(16,8)polar code.The simulation results show that several kinds of neural networks have certain decoding ability,among them,LSTM neural network has the best comprehensive performance under the full signal-to-noise ratio,and the bit error rate is below 5dB,which has about 0.2db performance gain compared with the traditional serial cancellation algorithm(SC)and belief propagation algorithm(BP).(2)Three algorithms are proposed,which are the combination of LSTM neural network and traditional polar code decoding.The first is LSTM-SC decoding.The combination of LSTM network and SC decoding can decode the long code in blocks.The neural network decoder is successfully applied to the decoding of the longer code.The experimental results show that the performance of LSTM decoder will be lower than that of SC decoder when the codeword is changed to(32,16)(this is also consistent with the theoretical network learning ability is affected by the information bit dimension and the decoding performance is reduced),while the performance of LSTM-SC decoder will be 0.5dB higher than that of SC decoder when the BER is 0.5dB,and the performance is close to that of CA-SCL(4)when the BER is 5dB.For(1024,512)polar code,BER has a performance gain of about 0.2db at 0.09 compared with SC decoding algorithm.In the second LSTM-BP decoding model,the neural network block is used instead of a part of BP network.At the same time,a method of initializing the right information is set.Experiments are carried out with(64,32)codes.The experimental results show that the LSTM-BP decoder can reduce the decoding delay at the same iteration times slightly better than BP decoding,and has a performance gain of about 0.7 dB at a bit error rate of 0.05.The third is the BP-LSTM-BP decoding model under the coherent noise.In this model,the neural network block learns the coherent noise characteristic instead of the code character characteristic.The experimental results show that the performance of the traditional decoding model is worse when the noise coherent coefficient is larger.On the contrary,the new decoding model has a gain of about 1.9db when the noise coherent coefficient is 0.9 and the bit error rate is 0.05.By increasing the number of neural network blocks,all three models can be applied to long code decoding in theory.