Research And Design Of Low Density Parity Check Code Decoder Based On Deep Neural Network

In recent years,with the increase of the parallel processing ability of computers and the development of artificial intelligence technologies,the deep neural network(DNN)has been widely used in many fields.The combination of the next generation mobile networks and DNN has become a research trend.The DNN can further improve the performance of channel decoder in complex environments.This paper focuses on the design of low-density parity-check code decoder in various channel environments,and achieves the following research results.(1)Considering the situation of additive white Gaussian noise channel,short code,and the characteristics of different DNNs,three kinds of channel decoders based on neural networks(NNs)are designed,which are respectively based on multi-layer perceptron(MLP),convolutional neural network(CNN)and long short-term memory network(LSTM).The DNN structure of the decoder realizes the function of denoising the received signal.The simulation results show that the CNN decoder has the advantages of decoding performance and network training efficiency compared with the MLP and the LSTM decoders.(2)Considering the performance degradation of traditional belief propagation(BP)decoder caused by correlated Gaussian noise channel and the dimensional curse of neural network decoder caused by medium and long code.A new channel decoder based on cascaded iteration of gated CNN and BP is proposed.The gated CNN estimates the correlated noise.The simulation results show that the proposed decoder structure can effectively overcome the performance degradation of BP decoder caused by noise correlation.Specifically,compared with the standard BP decoder,the proposed decoder effectively reduces the bit error rate at 576 bits.When the noise correlation is 0 to 0.8,the maximum performance gain is 4d B.(3)Considering the correlation of Rayleigh fast fading and the additive Gaussian noise channel environment,a non-iterative channel decoder based on the cascade of double CNN and BP is proposed.The decoder uses two one-dimensional CNNs to estimate correlation noise and channel gain respectively.The simulation results show that the proposed decoder structure can achieve a lower bit error rate than the traditional BP decoder.When the channel correlation is0.9 and the noise correlation is 0 to 0.9,the maximum performance gain is 7 d B.Finally,the work and the shortcomings are summarized and the future work is discussed.