Design And Implementation Of Channel Adaptive Communication Autoencoder

With the rapid growth of Internet business and the continuous development of communication technology,it is more and more difficult for people to leave instant online communication in their lives.In the communication process,the channel generally changes with time,which increases the demand for adaptive channel technology in the communication system.However,it is difficult for traditional communication technology to adapt to the channel in time when the channel changes rapidly.On the one hand,the research of artificial intelligence in the field of communication has become more and more in-depth.The end-to-end(E2E)autoencoder system solves the local optimization problem caused by the sub-module design of traditional communication systems,and can achieve the best overall system.But it can only adapt to fixed channels,which has limitations.On the other hand,Meta-learning enables neural networks to learn the task distribution of different channels and adapt to changing channels.Therefore,Metalearning can be integrated into the autoencoder to meet the needs of channel adaptation and effectively learn and adapt to complex channels,breaking the limitation of poor performance of traditional communication algorithms in the face of real complex channels,which has great practical significance.This paper firstly introduces the basic theory of deep learning and communication autoencoder.The transmitter and receiver in the traditional communication structure are replaced by the end-to-end encoder and decoder respectively.In order to verify the performance of the autoencoder,a classical network model is designed using the fully connected layer,and the AWGN channel is carried out for the network.After comparing with traditional communication systems,it is concluded that the end-to-end autoencoder can realize the modulation and demodulation function of traditional communication systems;and it is pointed out that the autoencoder cannot adapt to the limitations of complex channels.There are two reasons,one is that the network model is relatively simple and cannot fully represent the channel features;the other is that the training method is limited and the features of different channels cannot be effectively extracted.Then,on the basis of introducing the principle of meta-learning and three classical algorithms,MAML,Meta-SGD and Reptile,this paper designs a communication autoencoder system model based on meta-learning.Task distribution and offline training and online fine-tuning algorithms are systematically designed.After the simulation of the bit error performance of the system under fixed channel and different channel conditions,it is concluded that the system can adapt to the channel.After that,the system is implemented based on FPGA hardware.After deriving the forward propagation and back propagation formulas of the model,the FPGA framework is designed,and the hardware implementation and optimization of the decoder are carried out based on the Vivado development kit.After the performance evaluation of the decoder,it is proved that its resource utilization and energy efficiency achieve good performance indicators,and its model test accuracy and fine-tuning performance are also relatively reliable.Finally,the full text is summarized at the end,and the areas for improvement and future research directions in the research on the combination of meta-learning and communication autoencoder and its hardware implementation proposed in this paper are pointed out.