Research On OFDM Channel Estimation And Prediction Technology Based On Deep Learning

With the development of internet of vehicles(Io V),a growing number of wireless high-speed scenarios have emerged,which place higher demands on wireless communication systems.High mobility and multi-path transmission will introduce time-selectivity and frequency-selectivity to the channel.Estimating the channel state information(CSI)from the fading-affected received signal is one of the most important steps for transmitted symbols recovery.In high-speed mobile scenarios,traditional pilot-based channel estimation methods require lots of pilot symbols,resulting in bandwidth wastage.To address this issue,channel prediction techniques can be used to predict the CSI of fast-changing channels.However,the performance of common autoregressive(AR)prediction method is insufficient.Recently,deep learning(DL)with powerful data processing capabilities brings new design approaches to communication systems.Therefore,by analyzing existing research on DL-based channel estimation and prediction methods,we follows the model-driven idea to resolve the problems of traditional channel estimation and prediction methods,and conducts an in-depth study on channel estimation and prediction methods for orthogonal frequency division multiplexing(OFDM)systems under Rayleigh double-selective fading channels.To address the issue of poor performance of traditional channel estimation and prediction methods,we designs a DL-based channel estimation and prediction network.It consists of a fully-connected deep neural network(FC-DNN)with powerful curve fitting capability,two twodimensional convolutional neural networks(CNN)with powerful image reconstruction capability,and a long short-term memory(LSTM)network with powerful time series processing capability.In particular,these three nerual networks are deployed to perform channel estimation,channel interpolation and channel prediction,respectively.In addition,we propose a data aided decision feedback scheme in prediction to reduce the impact of error propagation.Simulation results show that the three neural networks outperform the corresponding traditional methods,and the proposed channel estimation and prediction network improves the transmission quality and transmission efficiency of the system.Based on the above results,we then proposes Branching Dueling Q-Network(BDQN)to adaptively adjust the pilot location and spacing,in order to further exploring the possibility of reducing the pilot overhead with the DL-based channel estimation and prediction network.The system state consists of the current noise intensity obtained with the aid of FC-DNN and the future CSI predicted by the multi-step LSTM network.The agent action consists of the transmitted symbol block length and the pilot position.The reward is determined by the pilot overhead and the channel prediction accuracy.Simulation results show that the BDQN can dynamically adjust the pilot overhead according to the current system state.BDQN can pay attention to transmission quality or transmission efficiency,or the trade-off.Compared to the fixed pilot pattern,BDQN can gradually improve the transmission efficiency as the signal-to-noise ratio(SNR)increases while guaranteeing the transmission quality.