Signal Detection Method At The OFDM Receiver Based On GAN

In the field of wireless communication,the performance at the receiver of single input single output(SISO)orthogonal frequency division multiplexing(OFDM)and massive multiple input multiple output(MIMO)OFDM systems depends greatly on signal detection.At present,traditional signal detection methods are difficult to strike a balance between complexity and bit error ratio(BER)performance.Generative adversarial network(GAN)is one of the most promising self-supervised learning methods in deep learning(DL).The advantage of GAN is that it can continuously improve the performance through adversarial training of two DL networks.In this thesis,signal detection technology at the receiver based on the generative adversarial network is studied.The main work of this thesis is summarized as follows:Firstly,to solve the signal detection problem of the SISO-OFDM receiver,this thesis introduces the conditional generative adversarial network(c GAN).Firstly,the c GAN model uses the traditional least square(LS)and zero-forcing(ZF)algorithms to preprocess the received signals,and then the c GAN model uses these signals to be the initial input of the DL network,which is also called model-driven.In contrast to the data-driven deep neural network(DNN)model,which needs to be trained which a large number of training parameters,the model-driven c GAN model improves the training performance.In addition,pilot sequences are introduced into the c GAN model as the conditional input of generator and discriminator,and then the two DL networks are trained against each other.The adaptive loss function suitable for the c GAN model is designed,which has a certain correction effect on the optimization of neural networks and can improve the training and learning efficiency.Compared with the existing DNN detection model,the c GAN model is robust in the Rayleigh channel and achieves the improved performance of BER,especially under the conditions of low signal-noise ratio(SNR)and short pilot sequences.Secondly,to solve the signal detection problem of the massive MIMO-OFDM system,this thesis designs a GAN model,in which the approximate message passing(AMP)iterative detection algorithm is expanded by the fully connected hidden layers inside the GAN’s generator.On the one hand,the GAN model improves the performance of the traditional AMP iterative detection method.On the other hand,it solves the problem that the data-driven method needs to rely on a large number of parameters to be trained.This method is based on model-driven,which has the advantages of both the data-driven model and the traditional approach.The GAN detection method has significant BER performance gain with fewer iterations.The deep learning neural network can learn and analyze the characteristics of the iterative structure.The GAN model is robust in the Gaussian channel,which reduces the system’s complexity,BER performance is improved obviously under the conditions of fewer iterations and higher SNR.To sum up,this thesis uses GAN as the research method,takes the signal detection problem of the receiver of SISO-OFDM system and massive MIMO-OFDM system in wireless communication as the research contents,and takes model’s BER improvement as the research requirement.It is concluded that the performance improvement and complexity reduction of receivers’ signal detection in OFDM systems are the research purpose.The corresponding model designs are respectively proposed for SISO-OFDM and massive MIMO-OFDM systems in the study of the above problems,and the effectiveness of the above designs are verified by simulation experiments.