Research On Millimeter Wave Channel Clustering And Time-Varying Channel Playback Technologies By Machine Learning

With the development and popularization of the 5th generation mobile communication(5G),the lifestyle of people in modern society has undergone profound changes.The traditional demand for human-to-human communication have comprehensively shifted to the demand for interconnection of everything,causing massive communication devices and enormous data traffic,which puts higher quality of service requirements on mobile communication networks.As one of the key technologies for 5G,millimeter wave technology enables high transmission rates in hotspot scenarios with high user density and very large data traffic.Wireless channel modeling is the abstract description of wireless propagation environment and propagation characteristics,which is the primary problem to be solved for building wireless communication systems.Therefore,the study of millimeter wave channel modeling is crucial for 5G system design,simulation and advanced technology evaluation.Based on the measured channel data in a high speed railway station in power distribution scenario,this dissertation will focus on millimeter-wave channel modeling in 5G,explore millimeter wave channel clustering and time-varying channel playback technologies by machine learning in the face of the big data attributes in future wireless channels.Aiming at the problems that the traditional statistical-based channel simulation method cannot match the environment and is only applicable to the shortcomings of link-level simulation,we study the time-varying channel playback technology by machine-learning for millimeter-wave channel modeling to improve the accuracy of millimeter wave channel modeling,and serve for the design and optimization of 5G and next-generation mobile communication networks.The main work and innovations of this dissertation are as follows:(1)Research on support vector machine(SVM)-assisted adaptive kernel power density clustering algorithm for millimeter wave channelsTo address the problem that traditional algorithms depend on prior knowledge of clusters,adaptive kernel power density(AKPD)and SVM-assisted AKPD(SVM-AKPD)are proposed in this dissertation.Firstly,a new distance-based method is proposed to calculate an adaptive-K of multipath components(MPCs),thereby more accurately charistizing the local variation rules of MPC density.Then,the cluster centroids calculated by the AKPD are used as labeled data to train the SVM model.Finally,the SVM is applied to find the optimal hyperplane in feature space(channel parameter space)to realize the multi-dimensional and high-density nonlinear clustering of MPCs.The performance of the proposed AKPD and SVM-AKPD algorithms is verified based on the measured channel data and simulated data in the waiting hall of Qingdao high-speed railway station in power distribution scenario at 28 GHz.The results show that the proposed clustering algorithm outperforms the traditional channel clustering algorithms in different channel environments with good robustness,and does not require prior knowledge of channels.(2)Research on millimeter channel clustering by self-organizing maps with time-varying topological structureThe traditional channel clustering algorithm relies on the initialization of parameters,and its complexity is too high to be used for future channel modeling with large data attributes.In this dissertation,a clustering algorithm based on self-organizing map with time-varying topological structure(SOM-TVS)is proposed.Firstly,the SOM is initialized according to the distribution characteristics of MPCs by combining the locations and weights of the neurons in competition layer.Then,a new MPCs power weighted Gaussian-Sinc function are proposed to optimize the competitiveness performance of SOM.The clustering performance of the proposed SOM-TVS algorithm is verified based on the measured channel data and simulation data in the waiting hall of Qingdao high-speed railway station in power distribution scenario at 28 GHz.The results show that SOM-TVS outperforms traditional channel clustering algorithms,and achieves accurate clustering of millimeter-wave channels with lower complexity.(3)Research on MPC trajectory-based clustering algorithm for millimeter wave time-varying channelsFor non-stationary time-varying channels,a millimeter wave time-varying channel clustering algorithm based on MPC trajectory is proposed.Firstly,considering the distance and velocity similarities of the MPCs in different snapshots,fuzzy theory and global matching methods are used to achieve accurate tracking of MPC trajectories.Then,a fuzzy MPC trajectory clustering algorithm is proposed to cluster the MPCs.Based on fuzzy theory,different MPC trajectories are assigned to different MPC clusters according to their membership to achieve accurate clustering results.Finally,time-varying channels at 28 GHz are simulated to validate the performance of our proposed algorithm.The results show that the proposed algorithm is able to accurately identify the MPC trajectory clusters,and achieves accurate modeling of MPC time-evolution characteristics in millimeter wave time-varying channels.(4)Research on millimeter wave time-varying channel playback technology based on machine learningAiming at the problem that the simulated channel of the traditional stochastic channel modeling cannot match the environment well,this dissertation proposes a millimeter-wave time-varying channel playback technology based on machine learning.Utilizing the property that machine learning can adequately approximate complex nonlinear relationships,the path-loss plus shadow fading model is established to playback path-loss.And joint small-scale channel parameter model is established to generate small-scale channel parameters,which can be used to replace the random numbers generated by the statistical laws of channel parameters in the traditional stochastic channel simulations.Accordingly,the traditional channel modeling based on statistical laws is modified into more realistic time varying modeling.The performance of the proposed machine learing based millimeter wave channel playback technique is verified based on 28 GHz measured and simulated channel data in the waiting hall of Qingdao high-speed railway station in power distribution scenario.The results show that the proposed simulation technique overcomes the shortcomings of traditional simulation platforms,and can reproduce the time-varying channel propagation characteristics realistically.