Research On Wireless Precoding Algorithm Based On Complex-Valued Deep Learning In User-Centric Network

User centric network(UCN)enables users to always be in the center of the cell by building dynamic base station(BS)clusters to ensure the continuity and consistency of services,but it also brings higher scheduling complexity,which makes the traditional precoding algorithm in this scenario difficult to be competent for real-time tasks.In recent years,some researches use deep learning(DL)to extract the characteristics of traditional algorithms,and predict the results through simple operation during deployment to balance complexity and performance,so as to solve the problem of real-time.However,the attenuation and delay of wireless channel and the amplitude and phase of analog signal are mathematically more suitable to be expressed in complex number.In order to use the more mature real-valued DL technology,the existing research usually projects the data from complex-valued domain to real-valued domain,resulting in information loss.Therefore,this thesis uses the complex-valued DL model to fit the traditional precoding algorithm in UCN,directly process the complex-valued data and improve the performance of the model.The main work of this paper is divided into the following two parts:Aiming at the problem that the channel state and analog signal data need to be projected into the real-valued domain in UCN,this thesis uses complex-valued convolution neural network(CV-CNN)to fit the precoding algorithm based on weighted minimum mean square error(WMMSE).Combined with the data characteristics,this thesis designs the appropriate CV-CNN structure and loss function,so as to obtain the mapping relationship between channel state information and precoding.At the same time,the simulation compares the performance differences between the proposed model and the existing research in computational efficiency,sum rate and robustness,and demonstrates the value of complex-valued DL.In the real scene,users continue to join and leave the coverage of UCN,that is,the number of users changes dynamically.Aiming at the problem that the existing research is limited by the model structure,this thesis uses the fully convolutional network(FCN)to learn the precoding algorithm of UCN in the scene with dynamic number of users.Considering the performance gain of complex-valued DL,this thesis extends the realvalued FCN model to the complex-valued domain,and designs the appropriate data preprocessing process and model structure combined with the characteristics of dynamic data.Finally,the performance indexes of real-valued and complex-valued FCN,such as calculation efficiency,sum rate and generalization ability,are simulated and analyzed,and compared with the existing research to demonstrate the value of FCN in the face of the problem of dynamic precoding of the number of users.