Machine Learning-Based Data Processing And Channel Modeling For Time-Varying Channels

This thesis focuses on the machine learning-based channel data processing and channel modeling.Beyond 5G(B5G)communications are developed to meet the requirements of next-generation mobile communications.According to the development history of mobile communications,B5 G is expected to offer a significant improvement in network capacity,network node density support,connection density support,and peak rate experience.Meanwhile,lower end-to-end delay and high-speed mobility support are also required for the next-generation communications system.In this case,B5 G becomes a key technique to make the breakthrough of the information age,which connects everything and offers a great communication society to this word.On the other hand,the analysis and research of wireless channels are the very basic foundation to develop any new wireless communication system,where the understanding of radio propagation is the key method to study the physical channels.Such that,designing and optimizing the communication network highly depends on the study of the radio propagation and channel modeling.Hence,channel modeling is one of the hot topics of research on wireless communications.Facing the dramatic development of 5G communications,the channel modeling solution is also evolved from conventional deterministic modeling or statistic modeling to a new type of modeling method which relies on machine learning to process the channel data.Supported by massive MIMO techniques and better propagation sounders,the channel measurement campaign is able to have a comprehensive observation of the channels and offers more accurate measurement data.However,the quantity of the measurement data is also increased significantly,which brings difficulties to model the channel.On the other hand,machine learning methods have a great advantage in measurement data processing,including data mining,data compressing,and data characterizing,which ultimately leads to better accuracy of modeling.This thesis studies the data processing methods for channel data processing: by combining the evolution pattern of the multipath components(MPCs),a tracking-based MPC's clustering algorithm is proposed,whereas an online clustering recognition solution is developed based on the image processing methods.Considering the requirements of scenario identification of the time-varying channels,an machine learning-based line-of-sight(LOS)/ non-line-of-sight(NLOS)identification algorithm is proposed by analyzing the channel characteristics in multi-dimensions.Finally,considering it is still a lack of the study of a cluster-based channel model for vehicle-to-vehicle(V2V)communications,a hybrid irregular-shape geometry stochastic channel model is proposed based on the research above.The contribution of the thesis mainly includes the following aspects:1)A trajectory-based MPC clustering algorithm is proposed for the time-varying channel modeling,which tracks the MPC by minimizing the globe matching distance between two consecutive snapshots.According to the relative position between each trajectory,two different scenarios are defined to measure the similarity between every two trajectories.The distance between two trajectories is measured based on the similarity of the shape and the actual distance,where all MPCs are clustered based on their evolution pattern.Finally,the proposed algorithm is evaluated by using the synthetic channel data and the measurement channel data,and the results show that the proposed algorithm is able to accurately identify the dynamic clusters in time-varying channels.2)For time-varying non-stationary channels,a power angle spectrum(PAS)-based online clustering algorithm is proposed.Instead of using high-resolution parameter estimation methods,the proposed clustering algorithm directly identifies the clusters from the PAS.Based on the physical feature,the MPC's clusters are characterized in three properties: shape,size,and position.Based on the obtained three properties,the clusters are tracked by using the tracking algorithm proposed in Section 1.In addition,a re-track method is proposed to identify the clusters that occasionally disappear in only one snapshot.Finally,the proposed algorithm is evaluated by using the synthetic channel data and the measurement channel data,and the results show that the proposed algorithm is able to efficiently identify the dynamic clusters from the PAS without using any high-resolution parameter estimation method.3)For V2 V time-varying channels,a novel channel scenario identification solution is proposed,where the channel angular characteristics are utilized to improve the identification accuracy.Meanwhile,three popular machine learning methods,i.e.,support vector machine(SVM),random forest(RF),and artificial neural network(ANN),are trained by using different channel characteristics to identify the LOS/NLOS channel scenario.The impact of using different channel characteristics and different training methods are evaluated,where three different training strategies are also evaluated to see the practical performance of the LOS/NLOS identification algorithm.4)Based on the studies introduced above,a hybrid irregular-shape geometry stochastic channel model is proposed,where the MPC parameters are extracted by using a high-resolution estimation algorithm.The clusters are identified by using the proposed trajectories-based clustering method and further classified into four types: LOS component,reflection from static scatterers,reflection from dynamic scatterers,and multi-bounced reflection,where the inter-and intra-cluster parameters are derived.Finally,the channel model is evaluated by comparing the synthetic data and measurement data from different streets,and the detailed implementation steps are given.