Research Of Channel Estimation Method Based On Deep Learning For OFDM Channel

Wireless communication channels are characterized by time-varying,multipath,time delay,Doppler shift,etc.Orthogonal Frequency Division Multiplexing,as a multi-carrier digital modulation technique,can convert frequency selective fading channels into a series of parallel narrowband flat fading channels.It has the advantages of strong multipath resistance,high spectrum utilization,and simple implementation.Facing the complex and changing communication environment,channel estimation is a necessary means to ensure reliable communication.The main problem faced in channel estimation methods is to ensure the effectiveness of channel estimation while considering the complexity of estimation methods in time-varying environments.Deep learning,as an emerging data analysis and processing tool,has a unique advantage in learning the intrinsic laws of data and input-output mapping relationships.In this thesis,we propose to investigate the channel estimation problem in OFDM systems by combining deep learning.(1)The interpolation method in frequency-guided channel estimation is based on channel correlation,and when the channel characteristics are not correlated due to the time-varying and frequency-varying characteristics of the wireless channel,the channel interpolation results become unsatisfactory.To address this problem,the SRGAN model is improved in this thesis for solving the interpolation problem in channel estimation.In the improved SRWGAN model,the estimates obtained at the guide frequency are analogous to the pixel points on a low-resolution image,and the channel features are first extracted by convolutional layers.Next,the mapping relationship is learned by multiple residual networks.Then it is amplified by Sub-Pixel Convolution Layer.Finally,the Wasserstein distance is substituted for JS and KL scattering,and the WGAN network is used as the adjudicator instead of the original GAN network to optimize the estimation results.The experimental results show that SRWGAN improves the SNR by about 3 d B when the BER is the same compared with the conventional interpolation algorithm and SRCNN model,and the SNR improves by about 5 d B when the MSE values are the same.(2)In order to make the wireless communication system learn the nonlinear mapping relationship between input and output directly and fit the channel characteristics,this thesis combines the residual structure with DNN model and proposes the improved model Res-DNN.Res-DNN replaces the 3 hidden layers in DNN with 3 residual blocks,takes the received at the receiving end The data signal and the guide frequency signal received at the receiving end are used as the input data of the model,and the known signal at the transmitting end is used as the label to learn the nonlinear mapping relationship between the transmitting signal and the receiving signal.The experimental results show that the channel estimation method based on Res-DNN model can still maintain high estimation accuracy with good robustness when the number of guide frequencies is small or there is no protection interval between OFDM symbols.