Research On Turbo Coding Technology Based On Deep Learning

In wireless communication,channel coding is an effective means to improve the reliability of transmission systems.Turbo code is widely used in mobile communication,deep space communication,data link and other fields due to its excellent performance.In order to further improve the decoding efficiency and anti-interference performance of Turbo code,this thesis carries out the research of Turbo compilation code technology based on deep learning.This thesis elaborates on the principle of each module of Turbo compilation code,and simulates the typical algorithm of iterative decoding.Aiming at the serious performance loss of traditional decoding under burst noise,the convolutional neural network is used to realize the construction of deep learning decoding method.The encoding results of fixed-length code sequences under each register are stored as data labels,and then the network is trained with the noised coding dataset to obtain the best fit between the noisy information and the coding dataset.The simulation results show that the CNN-based decoding structure shows better anti-interference performance than iterative decoding for both Gaussian noise and burst noise,but CNN decoding adopts the form of block decoding,which increases the decoding delay by about half compared with iterative decoding.Furthermore,in order to improve the decoding delay and improve the efficiency of compilation code,a decoding method based on bidirectional recurrent neural network is proposed.This method uses the memory features of RNN to realize sequence coding,forms a training dataset by noising and extracts features from the data in time steps to best fit the encoded data.Furthermore,the weight parameters of the trained network are loaded into the decoding stage as initialization parameters,and the noise is filtered out in the decoding stage at the receiving end to improve the decoding performance.The simulation results show that in terms of reliability,the RNN-based decoding method is close to the CNN-based decoding method under high probability burst noise.In terms of effectiveness,the decoding delay of RNN-based decoding methods is much lower than that of CNN decoding at the same error level and not limited by code length,thus it has better universality.