End To End Communications System Model Design Based On Deep Neural Network

At present,the traditional communication system is based on modular design,and its design is mostly abstracted into a mathematical model which is easy to handle.However,in real life,due to the complex and changeable channel environment,the traditional communication system based on the stable mathematical model often cannot achieve the optimal performance,and the overall performance may not be optimal if each module is individually optimized.Using deep learning can break the traditional modular design idea of communication system and build an overall end-to-end communication system model.However,there are still many improvements in the network structure of the model,and the research in this aspect is still in the initial stage.Based on the above problems,this paper uses the principle of autoencoder in deep learning,improves the structure of end-to-end communication system based on deep neural network,and realizes the end-to-end performance optimization of the communication system.In the first part of this paper,an end-to-end communication system based on autoencoder is built by using deep neural network.The model loss under different optimizers and the influence of different training SNR on the performance of autoencoder are discussed.The simulation results show that the end-to-end communication system based on deep neural network can achieve better system performance than the traditional communication system under the condition of additive Gaussian white noise channel and taking the bit error rate as the performance measure.On this basis,the paper further optimizes the network structure,changes the input neural network structure of the encoder part and the network structure of the normalization layer,and compares the error performance of the autoencoder under three different network structures.The simulation results show that the network combination structure of embedding layer and batch normalization layer can solve the problems of a large number of redundant sparse matrices and gradient explosion in the original network structure,and improve the bit error performance of the system.This paper also extends the application scenario from single user system to dual user system.A dual user communication system model based on deep neural network in interference channel is built.The influence of different weight values of loss function on the performance of a single autoencoder and the whole system is discussed.The dynamic weight method is used to adjust the weight automatically to achieve the end-to-end optimization of the whole communication system.In order to solve the problem of large amount of data transmission in dual user system,the embedded layer is used for dense coding at its input.In addition,the average bit error rate of the system is further reduced by changing the normalization function of the neural network,and the two autoencoders are trained more equitably.