Research On Neural Network Model Decoding Of Convolutional Codes

With the development of modern communication technology,modern society has put forward higher requirements for data communication.Convolutional codes have been widely used in various digital communication systems,and tail-biting convolutional codes are also used as one of the channel coding schemes in LTE systems.However,the traditional decoding algorithm has the characteristics of high delay and low throughput,and cannot achieve parallel decoding.Based on the characteristics of convolutional neural networks and recurrent neural networks,this paper studies the decoding performance of convolutional codes and the feasibility of parallel decoding respectively.The main work of this paper is as follows:1.Designed and implemented a convolutional code decoding model based on convolutional neural network.Since each bit of the codeword of the convolutional code is correlated with several bits before and after,two methods are proposed to construct the neural network input to the received codeword information,and the convolutional neural network adopts a sliding window mechanism to translate code.Simulation experiments show that regardless of the size of the convolutional code constraint length,a convolutional neural network with a certain depth can learn the mapping relationship of convolutional code codewords.For the return-to-zero convolutional code,its performance is similar to Viterbi decoding.A certain gap,but better than the performance of Viterbi decoding of tail-biting convolutional codes with shorter information sequence length,and the neural network decoding model based on sliding window decoding mechanism can realize parallel decoding,reduce decoding delay and improve the throughput.2.Because recurrent neural network is a kind of neural network with time sequence characteristics,considering the correlation of convolutional codewords,two neural network decoding models based on LSTM are designed and implemented.One of the models is to input all the received codeword sequences into the neural network,and the performance is equivalent to that of the Viterbi decoding algorithm in(7,5)₈ convolutional codes,but with the increase of the constraint length,its decoding performance is not as good as that of the Viterbi decoding algorithm.The other model is to segment the received codeword sequence and input it into the neural network for decoding based on the sliding window mechanism.this model has stronger adaptability than the former model and does not degrade with the increase of constraint length.3.According to the related research of neural network noise reduction,two Viterbi decoding models based on neural network noise reduction are designed and implemented.The performance difference between convolutional neural network noise reduction and recurrent neural network noise reduction is studied and analyzed.The noise network simulates and analyzes the Viterbi hard decision decoding and the Viterbi soft decision decoding.The simulation experiment shows that the noise reduction network has minimal improvement on the Viterbi soft decision decoding algorithm,but for the Viterbi hard decision decoding The improvement of the code algorithm is greater.