Research On Convolutional Code Decoders Based On Deep Learning Under Correlated Noise

The traditional convolutional code decoding algorithm,such as soft-decision Viterbi decoding,can achieve ML performance under the AWGN channel,and the noise in the actual transmission has a certain correlation,so that its performance is significantly reduced.In the conventional traditional way,the method of solving the correlated noise is to whiten the noise by matrix multiplication,and it is difficult to obtain a priori knowledge about the correlated noise more completely in the matrix multiplication,and when the code length is long,the complexity is higher.Based on the characteristics of different neural networks,this paper designs convolutional code decoders based on convolutional neural networks and recurrent neural networks.Firstly,a DHDV decoder based on DnCNN network and traditional hard-decision Viterbi decoder is proposed.The architecture uses DnCNN to denoise using noise correlation to improve the SNR(signal-to-noise ratio)of the received signal,and then uses the hard-decision Viterbi algorithm to decode the denoised data.The stronger the noise correlation,the performance gains achieved by the DHDV decoder are more significant.It also exhibits strong robustness under different convolutions and different correlated noise models.Compared with the complexity of DnCNN network and matrix multiplication whitening method,DnCNN has lower complexity when the code length is longer,which is mainly affected by code length.When the code length is shorter,DnCNN has higher complexity,mainly from its network parameter.Then,based on the recurrent neural network GRU,the RNN neural network decoder is designed to completely replace the traditional convolutional code decoding algorithm.Using the bidirectional GRU network and the fully connected neural network,the sequence feature information of the convolutional code is extracted through the bidirectional GRU network,and the decoding result is calculated by the fully connected neural network.The RNN neural network decoder is better than the DHDV decoder and the traditional Viterbi decoding algorithm for convolutional codes with shorter memory lengths,such as(2,1,3)convolutional codes,under correlated noise channel.The greater the noise correlation,the greater the performance improvement of the decoder relative to the traditional Viterbi decoding algorithm.Besides,the RNN neural network decoder is close to the MAP performance under the AWGN channel.Due to the structure and complexity of the RNN decoder,as the memory length of the convolutional code increases,its decoding performance is gradually reduced,which is not suitable for convolutional codes with long memory lengths.The two neural network-based decoders proposed in this paper have their own advantages.The DHDV decoder has better versatility and exhibits excellent performance on different convolutional codes.The RNN neural network completely replaces the traditional convolutional code decoder,which is more convenient and has better performance on convolutional codes with smaller memory length,but the versatility is insufficient.