| Signal transmission is the most important part of the communication system,which determines the success of communication.Although the traditional communication technology has become mature,the technical realization process of signal transmission in the case of global optimization and unknown condition information is still in progress.There are many challenging problems.The rapid development of deep learning technology has made people realize that neural networks can be applied to the field of communication to solve existing problems.Therefore,in order to improve the accuracy of wireless communication system modeling and analysis,this thesis starts from the design concept of using deep learning to realize the physical layer of the communication system,and focuses on the problem of using the GAN(Generative Advertise Networks,GAN)model to realize the wireless communication channel modeling problem.Carry out research,the specific research work is as follows:(1)In order to solve the problems that traditional communication systems can only achieve individual module structure optimization but cannot achieve global optimization,and it is difficult to obtain channel state information in actual scenarios,the DCGAN(Deep Convolution Generator Advertise Networks,DCGAN)model is generated.Combined with autoencoder to build end-to-end communication system,a channel modeling method based on DCGAN model is proposed.This scheme introduces(Convolutional Neural Networks,CNN)into the GAN network to train with deeper neural network layers and fewer network parameters.Simulation experiments show that the overall training of the end-to-end communication system can achieve the global optimal state,and the bit error rate of the scheme is consistent with the traditional method when transmitting long sequences,which proves the feasibility of the experimental scheme.(2)Aiming at the problem that the channel modeling method based on the DCGAN network belongs to supervised learning and requires conditional information for labeling,and the acquisition of label information in the actual scene will consume a lot of manpower,material resources and financial resources,based on the improvement of the DCGAN network,it is proposed based on a channel modeling approach for Info GAN(Information Generative Adversarial Networks,Info GAN).This scheme introduces the concept of mutual information,uses the maximization of mutual information to make latent variables have the characteristics of real samples,and realizes unsupervised learning without increasing the amount of calculation.By introducing the unsupervised deep learning method,the neural network model can complete the learning and training when the label information is unknown.Simulation experiments show that the model can accurately model the channel without channel state information under different channel environments and modulation modes.(3)In view of the unstable training of the GAN model,which is prone to gradient disappearance or gradient explosion,resulting in the lack of diversity of generated samples,a relatively smooth Wasserstein distance is introduced into the model,and a new method based on WGAN(Wasserstein Generative Advertise Networks,WGAN)is proposed.model of the channel modeling method.This scheme can provide meaningful gradients with no or negligible overlap between generated samples and real samples,enabling the model to train stably.Simulation experiments show that the WGAN model has higher stability,and because the channel modeling method of the WGAN model is an improvement on the basis of the DCGAN method,it can not only be trained under a fixed signal-to-noise ratio,but also can be extended to 0~30,and has good adaptability and generalization.When applied to wireless transmission,the bit error rate performance is significantly improved. |
