Research On Key Technologies Of End-to-end Wireless Communication Channel Imitation And Signal Reception Under Few-shot Environments

End-to-End communication has been widly studied ever since it was proposed for the first time.Conventional deep neural networks achieved remarkable success with large volumes of training data as well as a large number of training epochs.However,for a more practical matter,wireless propagations vary depending on applications and are usually unknown,and readily available information resources are limited,which makes classical end-to-end communication unable to achieve performance comparable to traditional block-structure communication links as that under the additive Gaussian white noise channel,which has increasingly become one of the technical bottlenecks of end-to-end communication.To address the above problems,this thesis investigates the key technologies of end-to-end wireless communication channel imitation and signal reception for few-shot communication scenarios where only a few labeled data are available.The main work and contributions of this thesis are given in the following:(1)Aiming at the generation problem of unknown channel under few-shot environment,researches on channel imitation mechanisms based on meta-learning and generative adversarial networks are studied,and a few-shot channel imitation network of DAWSON framework which combines Reptile algorithm and generative adversarial networks is proposed.On the basis of this network,a variational sampling layer is added to the structure of the generator to enhance its ability of capturing channel characteristics.The proposed method is validated under Pytorch environment,and the results show that:Under symmetric alpha stable distribution noise channel,the proposed method can imitate the target channel with few labeled samples.(2)Aiming at the problem of signal reception and decoding at the receiver of end-to-end wireless communication under few-shot environment,decoding algorithms for the receiver that suits few-shot environment are studied,and an end-to-end wireless communication reception algorithm based on prototypical network is proposed.The prototype of each category is computed by processing query sample sets,computing prototype components after noise filtering and averaging over all prototype components,so as to improve the decoding performances of the receiver with clustering.Simulation results show that:the proposed algorithm has lower symbol error rate than existing end-to-end baseline algorithms in a few-shot environments.(3)Within the Pytorch open-source machine learning framework,the physical layer link simulations are performed for the end-to-end wireless communication system which concatenats the transmitter,the few-shot channel generative adversarial network based on Reptile algorithm,and the receiver network based on prototypical network.Results show that the proposed algorithm can successfully realize the transmission and reception of end-to-end wireless communication signals in the few-shot environment.Further,considering the problem that the training and testing channel environments may be dramatically different in practical applications,a performance enhancement method with few-shot channel imitation network is proposed.Computer simulations show that the proposed method can effectively improve the generalization performance of end-to-end wireless communications when the training and testing environments are different.