Research On Key Technologies Of Physical Layer Based On Deep Learning In High-speed Scenario

In recent years,with high-speed railway becoming the main transportation for regional communication,high-speed railway scenario has become one of the important research scenarios of LTE,5G system and etc.Due to the high-speed mobility,the wireless channel performs fast time-varying and non-stationary characteriatics,which brings challenges to wireless transmission technologies such as channel estimation and prediction,signal detection and demodulation,channel equalization,channel decoding,and etc.On the other hand,compared with the public network scenario,the joint performance of different communication modules in the high-speed railway scenario is suboptimal,and they are expected to support different communication modes,which increases the adaptive pressure for the receiver.In order to reduce the negative effects of high-speed railway scenario on wireless transmission technology,it is meaningful to optimize the key technologies of physical layer based on deep learning.This paper analyzes the characteristics of wireless channel in high-speed railway scenario to realize high-speed transmission and high reliable communication,summarizes the particularity of Doppler frequency shift and nonstationary time-varying channel impulse response and improve the wireless transmission technology with the help of feature extraction and nonlinear mapping ability of neural networks.Firstly,this paper studies the dynamic characteristics of Doppler shift in high-speed railway scenario and proposes a novel Doppler shift prediction algorithm based on Long Short-Term Memory(LSTM)networks with the reference signal configuration in wireless communication system.Using the initial training sets based on theoretical values and the dynamic training sets based on estimated values,the tracking of channel is realized.Without changing the frame structure,the algorithm achieves better prediction performance than conventional prediction algorithms,and reduces the impact of estimation error on Doppler shift prediction.Then,this paper proposes a novel model-driven adaptive receiver for the fast timevarying and non-stationary channel.The Channel Net,which is realized by small-scale Convolutional Neural Network(CNN),is used to track the channel and provide accurate channel state information.The Signal Net,which is realized by CNN+LSTM,is used to detect and recover the received signal.The model-driven receiver are trained in real time by using the historical estimation of the reference signal to adapt to the channel changes in the high-speed railway scenario.Without changing the frame structure,the scheme achieves the better performance than traditional algorithms,realizing the high reliability and high transmission efficiency communication in the high-speed railway scenario.Finally,the paper simulates the traditional decoding algorithms of different channel coding schemes,discusses the impact of two types of training sets on the decoding performance and proposes the general decoding scheme based on neural network.Through the constrant features extracted by CNN,the general decoding network structure and decoding scheme are designed.The corresponding neural network weights are saved and the general decoding of different coding schemes is realized.