End-to-end Coding And Modulation Systems Based On Deep Learning

The fifth generation(5G)mobile communication system for Mobile Internet and Internet of Things has become a major research hotspot in academia and industry.However,traditional communication systems are generally designed and optimized based on modules,such as channel estimation,channel decoding,etc.This approach does not guarantee the overall optimal performance of the communication system.The revival of artificial intelligence technology provides a new way of thinking for the design and optimization of communication systems.The end-to-end communication system based on the autoencoder models the physical layer communication as an end-to-end signal reconstruction problem.The input data and output data are jointly optimized,and the error between the transmitter and the receiver can be minimized.The rapid development of channel coding technology has important theoretical research significance and practical application value,and the gain brought by channel coding in the actual communication system cannot be ignored.This thesis designs and implements an end-to-end coding and modulation system based on a convolutional autoencoder.The system can learn the optimal mapping space for different channel coding schemes,improve the system BER performance,and has good reliability and generalization capabilities.The main contents are as follows:First of all,this thesis introduces the design ideas of the proposed structure in detail from the selection of coding schemes,the determination of neural network types,and the setting of power normalization layers.The network structure of the system under the Additive White Gaussian Noise(AWGN)channel is proposed.The simulation experiment in this thesis uses5 G NR LDPC codes.The system can block codeword sequences into one-hot vectors,and then perform constellation mapping in a multi-layer one-dimensional convolutional network.The receiver sends the soft information modulated by the neural network to the LDPC decoder for iteration to recover the original codeword sequence at the transmitter.The simulation results show that the end-to-end coding and modulation system based on the convolutional autoencoder has the characteristics of fast convergence,provide a reliable transmission environment for the channel coding scheme,and achieve similar block error rate performance under the AWGN channel.Next,under the Rayleigh flat fading channel of Single-Input Single-Output(SISO),the system structure is further optimized by adjusting the scale and parameters of the neural network.The simulation results show that the system can work effectively in different channel environments,achieving a balance between network complexity and system block error rate performance.On this basis,this thesis extends the end-to-end coding and modulation system to the Single-Input Multiple-Output(SIMO)system.Based on the idea of deep learning and integrated learning respectively,two end-to-end diversity reception structures are proposed.The simulation results show that the proposed structure can effectively combat fading in the channel,has a strong generalization ability,and its block error rate performance is better than that of traditional communication receiving diversity performance.