Channel Coding Based On Deep Learning

The channel coding theory can be understood as a mapping from the original information to the encoded information and then restoring the original information from the encoded information destroyed by noise.As the neural network can learn the internal laws from input information to output information,and its performance continues to improve as the scale of data increases.Therefore,the introduction of neural networks into channel coding technology has become one of the current research hotspots.In this thesis,we use a variety of common network models to explore the implementation of the combination of deep learning technology and channel coding technology.The specific research work is as follows:Based on the existing Viterbi decoding algorithm and combined with Recurrent Neural Network(RNN),a convolutional code decoder Network is proposed(CONV-DEC).The decoder structure uses bi-directional RNN.When each bit is decoded,the decoding output depends on the information of the before and after bits,which is beneficial to improve the decoding accuracy.Simulation experiments show that for different convolutional codes,the decoding performance of CONV-DEC is close to that of Viterbi decoding,which shows a certain generalization ability under longer block length,and has higher reliability under burst channel.On this basis,combined with the idea of the BCJR decoding algorithm,a Turbo decoder network is proposed by using CONV-DEC instead of component code decoder(Turbo-DEC).The BCJR iterative decoding structure is constructed by using a neural network.The number of network layers is increased according to the number of iterations,and the number of posterior information is increased to improve the system performance.Simulation experiments show that at high SNR,Turbo-DEC shows an error performance close to the BCJR decoding algorithm.In order to improve the reliability of the channel coding system,an end-to-end communication system based on Autoencoder is proposed.Through the joint optimization of the encoder and decoder,the reliability of the communication system is improved.In this thesis,an end-to-end communication system combining Convolutional Neural Networks(CNN)and RNN is proposed for the encoding algorithm of Convolutional code(CNN-RNN-AE).The CONV-DEC is used as the decoder of CNN learning Convolutional code encoding algorithm.Simulation experiments show that CNN-RNN-AE has an error performance close to(2,1,3)convolutional codes.In order to further improve the reliability of the channel coding system,an end-to-end communication system scheme based on RNN is proposed(RNN-AE),The encoder in CNN-RNN-AE is replaced by RNN.Simulation experiments show that the RNN-AE has higher reliability under different block lengths,and has certain adaptability under burst channel.In order to improve the use of feedback information in the channel feedback coding system,a design scheme for a highly reliable autoencoder system scheme with a feedback channel is proposed(FBC-AE).The scheme introduces a feedback decoder module in the transmitter,it divides coding and feedback information extraction into two parts.By using the end-to-end learning method,the transmitter and receiver can find a coding scheme and decoding algorithm suitable for the channel.In this thesis,two kinds of feedback decoders based on CNN and Bi-GRU are presented.In addition,a partial feedback scheme is proposed to save the power of feedback transmission.The power of feedback transmission is controlled flexibly by sampling interval,and the feedback information is recovered and extracted by the feedback decoder.Simulation experiments show that under the conditions of different code rates and different feedback channel signal-to-noise ratios,the system error performance of the two proposed channel feedback coding schemes is better than that of the existing literature.And when the feedback channel environment is poor,the partial feedback schemes can save feedback transmission power while still improving the error performance of 0.1dB.