Research On Wireless Channel Simulation And Modeling Based On Neural Network

With the development of the fifth generation wireless communication system(5G),the demand for high data rate applications has exploded.Channel data expands in many dimensions such as time domain,frequency domain,airspace and scene,showing the 4V attribute of big data,such as volume,variety,low data value and fast velocity flow.The explosion of data and the demand for fast and efficient data exchange put forward higher requirements for the spectrum utilization and capacity of wireless communication systems.In order to improve the spectrum utilization rate and realize ultra-large capacity transmission,the premise is to accurately grasp the wireless channel characteristic information and establish an accurate channel model.However,although the measured data of wireless channels have obvious big data attributes,the research status of 5G channel modeling is still a continuation of the traditional channel modeling methods and theories,and the channels are still modeled and classified according to the classification methods of scenes and frequency points.However,with the explosion of measured data,the differences between different scenes are gradually reduced,and the continuity between measured frequency points is gradually enhanced.Therefore,finding an effective method to find the commonness of data from a large number of data,thus achieving the goal of modeling becomes the key.With the development of artificial intelligence,the combination of machine learning and wireless communication has attracted the attention of many scholars.Many research results show that neural networks have strong learning ability and can approach nonlinear systems in complex situations.Using neural network to estimate and predict the channel parameters and model the channel can process the wireless channel measurement data with high reliability and efficiency according to the characteristics of data source such as complexity,high dimension and sparsity,providing effective theoretical and technical support for wireless communication environment reconstruction and model construction.This paper studies the application of feedback neural network(BPNN)in Doppler power spectrum simulation of wireless time-varying channel and MIMO channel parameter extraction.In wireless time-varying channels,a large number of Doppler power spectrum training sample data are generated by using traditional sine wave superposition method and shaping filter method.The time-varying channels are predicted by training BPNN,and the simulation results are compared with traditional methods.The results show that the simulation effect of BPNN is closer to the theoretical value and the error is smaller than that of the traditional method.At the same time,the frequency domain of the output results of the model is U-shaped,and the time domain autocorrelation function satisfies the features of the first kind of Bessel function,which is in good accordance with the time-frequency domain conditions of Jakes model.Comparing the time complexity of the three methods,the results show that BPNN has the highest time complexity,sacrificing the time complexity for high precision and low error.For MIMO channels,the Qua Dri Ga simulation platform is used to generate the channel impulse response(CIR)in urban scenes,and SAGE algorithm is used to extract channel parameters such as azimuth angle of horizontal dimension,pitch angle of vertical dimension,time delay expansion,etc.Then BPNN is trained with sample data to extract channel parameters,and the results are compared with SAGE algorithm.The results show that the prediction effect of BPNN model is little different from SAGE algorithm,so BPNN model can replace SAGE algorithm to extract channel parameters for MIMO channel simulation.In addition,the time complexity of BPNN and SAGE are also compared.The results show that the time complexity of BPNN is greater than SAGE algorithm.Besides,the prediction results of three common neural network training algorithms are compared and analyzed.The results show that the Levenberg-Marquardt algorithm has the lowest mean square error and the best effect in training BPNN.