Study On Propagation Characteristics Of Indoor 28 GHz Millimeter Wave Massive MIMO Wireless Channel Based On Machine Learning

With the application and development of multiple antennas C,the increasing size of the antenna array will hinder the deployment of some 5G application scenarios.Therefore,the millimeter-wave massive MIMO technology in indoor hotspot coverage scenarios relies on its advantages in antiinterference ability and energy efficiency,which has attracted a high degree of attention from researchers at home and abroad.Based on machine learning and SAGE estimation,the measurement activity is carried out at the 28 GHz massive MIMO channel in the waiting halls of high-speed railway stations.The channel characteristics are studied through machine learning and SBR/IM(Shooting and Bouncing Ray/Image)simulation methods.The actual measurement of the massive MIMO channel and the extraction of characteristic parameters in the high-hot spot scene,like the waiting hall,provide a theoretical basis for the coverage of the indoor 5G communication network.The main work of this thesis is as follows:(1)Measurement activities of indoor channels,high-resolution channel parameter estimation algorithms,and domestic and foreign research status of the massive MIMO channel modelings and parameter estimation based on machine learning are summarized.At the same time,the existing massive MIMO channel modeling methods and characteristics are summarized,and the technical route of channel estimation based on the SAGE(Space-alternating Generalized Expectation maximization)algorithm is explained in detail.(2)A data preprocessing method based on dimensionality reduction is proposed.The measured original CIR(Channel Impulse Response)is denoised by the PCA(Principal Component Analysis)method.By comparing the degree of multipath dispersion extracted by traditional denoising and PCA denoising,it is obtained that the standard deviation of the multipath received power obtained using the PCA method to denoise is smaller than the peak reduction method.Moreover,compared with the power delay spectrum,the PDP(Power Delay Profile)obtained by the PCA method has almost no noise,which is obviously better than the peak drop method.(3)The multipath cluster structure at the massive MIMO measurement channel in the waiting hall of the high-speed railway station is studied.The KPM(K Power Means)multipath clustering algorithm based on time delay,angle of arrival and received power is used for clustering.In addition,the CH criterion and the DB criterion is introduced to realize the self-adaptation of the number of clusters.By comparing the results of multipath clustering before and after adaptation,it is obtained that the separation between clusters of multipath channels is more obvious after adaptive interference.Moreover,the phenomenon of clusters birth and death changes and deviations exhibited by multipath clusters show the difference in the spatial non-stationary characteristics of massive MIMO channels at Lo S(Line of Sight)and NLo S(Non-Line of Sight)scenario.(4)Two channel models are established to study the impact of environmental information on the indoor 28 GHz massive MIMO channel: one is the RBFNN(Radial Basis Function Neural Network)reconstruction channel model based on measured data and environmental information,and the other is the SBR/IM simulation model that only depends on environmental information.In the SBR/IM simulation modeling,the simulation depth(7 reflections,1 transmission,1 diffraction)is determined.Comparing the received power obtained by these two models with the measured received power,it verifies the correctness of the results obtained by these models in Lo S and NLo S scenarios.By studying the APDP(Angular Power Delay Profile)and statistical parameter characteristics,the difference amomg these two models and the measured SAGE estimation is revealed.That is,the RBFNN reconstruction model is more consistent with the measured results,and can more completely reflect the spatial non-stationary characteristics of the actual channel.In contrast,the SBR/IM simulation model can only predict paths with higher power through environment and measurement information.