Autoencoder-Based End-to-End Wireless Communication Systems

Autoencoder as a popular model in Deep Learning(DL),which has the similarity to communication systems,and has been feasible scheme in designing the physical layer of wireless communication systems.Moreover,the autoencoder system,which based on Convolutional Neural Network(CNN)has also been extensively studied.Based on the common models in DL and combining various neural network models with wireless communication system,this paper focuses on improving the performance of communication system based on CNN.The specific research work is as follows:Aimed at the problem of improving the generalization capability of the existing CNN-based end-to-end system,a novel CNN-based autoencoder communication system(CNN-AE)is proposed.Compared with the existing results,CNN-AE system not only inherits the characteristics that end-to-end autoencoder communication system can jointly performs the tasks of encoding/decoding and modulation/demodulation,but also has inherited CNN's characteristics such as local connection which can support more bits information.Simulation results show that the system can convergence fast under specific Signal to Noise Ratio(SNR).After training the system,it can work in the whole SNR.This scheme not only solves the problem of low generalization ability in the original system,but also improves the reliability of the system.Aimed at the performance of the imperfect channel state information(CSI)under Rayleigh fading channels,by around the idea of differential modulation,an improved scheme of CNN-based differential coding system(DCNN-AE)is proposed.This method introduces a differential encoding module to deal with the transmitted symbols by differential encoding,thereby it doesn't need channel estimation at the receiver and it can reduce network complexity.Simulation experiments show the performance of the DCNN-AE system that is similar to the existing differential modulation scheme under Rayleigh fading channels,which can demonstrate the effectiveness of the system.In order to improve the reliability of the system at the high SNR,around the idea of convolutional code,Recurrent Neural Network(RNN)is proposed.The scheme combine convolutional code with CNN/RNN layers.This scheme uses CNN to perform feature extraction on the input sequence to learn the best modulation scheme for sending symbols.Drawing support from RNN's characteristics,it can introduce correlation into the input sequence and learns the coding scheme of the transmitted symbols through training.Simulation experiments show that the CNN-RNN-AE system has good generalization and reliability in different channel environments.From the bottom of the network,the influence of factors such as the number of CNN layers,the number of kernels,the number of RNN layers and different types of RNN on the system performance has been studied.The simulation results are explained from the perspectives of modulation,coding,and deep learning.Simulation experiments show that the performance of CNN-AE system is related to the number of CNN layers and the number of kernels.The performance of CNN-RNN-AE system is related to the number of network layers,the number of units and the type of RNN.