Research And Design On Channel Estimation Scheme Based On Generative Adversarial Network In THz-MIMO System

Tera Hertz Multiple Input Multiple Output technology features huge bandwidth resources,high multiplexing gain,and can provide higher data rates and larger network capacity.It is the core technology of the sixth generation mobile communication.However,the severe frequency-selective fading effect in Tera Hertz MIMO systems and the high dimensional channels brought by a large number of antennas increase the number and difficulty of parameters to be estimated,which brings great challenges to channel estimation.The traditional signal estimation method is difficult to accurately estimate the channel in the increasingly complex wireless communication environment.Generative adversarial network is one of the most innovative network structures in recent years.It generates nearly real data distribution through game adversarial training mode.It is a hot research direction in the field of artificial intelligence.This thesis introduces this method to the channel estimation problem in wireless communication,hoping to improve the accuracy of channel estimation by means of the powerful feature extraction and data generation capability of the generated adversarial network,and explore the scheme to reduce pilot frequency cost and improve the efficiency of channel estimation.In the field of computer vision,image super resolution technology uses generative adversarial network to fill pixels and obtain outstanding image restoration effect.Inspired by this,this thesis firstly uses super resolution generative adversarial network to estimate Tera Hertz channels.The initial space-time channel response matrix in this scheme is regarded as a two-dimensional low-resolution picture.Then,the space-time characteristics of the Tera Hertz channel are extracted by the super resolution generative adversarial network,and the high-resolution picture is obtained by the adversarial learning of the generator and discriminator,namely the complete channel matrix.According to this scheme,the method of data set preprocessing is optimized to realize the effective fusion of multiple channel state information.Simulation results show that the performance of the proposed SRCGAN algorithm is superior to the traditional interpolation algorithm and super-resolution convolutional networks.At the same time,the information provided by the frequency correlation between adjacent subcarriers can make more effective use of the three-dimensional correlation of "time,frequency and space" of the channel.Compared with the space-time channel estimation scheme based on SRGAN,its performance is more outstanding.There is a certain correlation between the channel information of adjacent subcarriers in the coherent bandwidth.In order to make full use of the multi-dimensional channel characteristics of Tera Hertz channel,this thesis proposes a channel estimation algorithm based on super resolution conditional generative adversarial network.The data is optimized with pre-arranged methods and the structure of deep neural networks.In this scheme,we cache the channel matrix of the previous subcarrier based on the rough estimate of pilot frequency and input it into the current super resolution conditional generative adversarial network as a conditional label.By introducing additional prior information to constrain the generator and discriminator,the generated adjoint network is extended into a conditional model and the generator is guided to generate.According to this scheme,the data is optimized with pre-arranged methods and the structure of deep neural networks.Simulation results show that the proposed super resolution conditional generative adversarial network based space-time-frequency channel estimation is superior to traditional interpolation algorithms and super-resolution convolutional networks.The information provided by the frequency correlation between adjacent subcarriers can make more efficient use of the channel3 D correlation,thus improving the estimation performance.The above work uses super resolution conditional generative adversarial network to achieve high accuracy channel estimation problem,and the simulation results prove that the generative adversarial network is efficient in processing massive data.On this basis,in this chapter,we work from the perspective of reducing pilot overhead and operational complexity and improving the robustness of the system.By further utilizing the correlation between Tera Hertz channel frequencies,we propose a channel estimation algorithm with low pilot overhead,and carry out lightweight improvement on super resolution conditional generative adversarial network,and improve the online prediction module.In the low-cost channel estimation algorithm,the four subcarriers are combined into one channel estimation unit.The channel estimation units do not coincide with each other,and the correlation between subcarriers is used to reduce the pilot frequency overhead.More pilot frequency is inserted at the first subcarrier,and the pilot frequency is halved at the following three subcarriers.The preliminary results of channel estimation are recorded as cache at each subcarrier except the last one.Cached results and initial estimates of their own channels are fed into the lightweight super resolution conditional generative adversarial network.In order to realize real-time channel prediction and low-delay data communication,and solve the problems of high complexity and high training cost of the current super-resolution generating adversarial network,this thesis,based on the super-resolution model framework of generating adversarial network,replaces the original convolution layer of the generator with a lightweight deep separable convolution,which reduces network parameters and computation with almost no loss of performance.In this thesis,momentum-based gradient descent algorithm is used to optimize the online prediction module to enhance the robustness of super resolution generation adversarial network to time-varying channel prediction.Simulation results show that the proposed algorithm can achieve high estimation performance,low computational complexity,and strong robustness to the moving speed.