Study On LOS/NLOS Idenitification Techniques Based On Machine Learning

With the increasing demand for location based services,how to improve the location accuracy in indoor and outdoor environments has been a hot research issue.Achieving highprecision positioning of moving targets in non-line-of-sight(NLOS)environments is still a technical challenge.Compared with the line-of-sight(LOS)environment,the NLOS environment results in a dramatic increase in the estimation error of the characteristic parameters such as time of arrival(TOA)and angle of arrival(AOA),which results in a significant increase in the positioning error when applied to localization compared with the LOS environment.In wireless positioning technology,the negative impact of NLOS propagation on localization accuracy can be greatly reduced by correctly identifying the LOS/NLOS propagation path and handling it appropriately in the positioning algorithm.To this end,this thesis develops NLOS environment modeling and LOS/NLOS propagation path identification techniques,and explores the LOS/NLOS identification methods in outdoor cellular environment and indoor ultra-wideband environment.The main work is as follows.First,by reading the current references on NLOS environment modeling and path identification,the characteristics of existing NLOS scattering models and path identification methods are summarized,and it is found in the literature that none of the existing NLOS scattering models consider the case where the direct path is blocked.And NLOS identification methods can be mainly classified into methods based on statistical characteristics and machine learning.Then,two classical single scattering models in NLOS environment and IEEE standard ultra-wideband channel model are introduced,and simulation analysis is performed for both channel environments.Secondly,based on the analysis of existing scattering models,this thesis proposes a scattering model with restricted AOA in the NLOS case-cut ring/disk of scatterer(CDOS/CROS)model,which considers the case where the direct path is blocked,and by setting the blocking angle,the signal arrival angle is restricted to a certain range.In this thesis,this new scattering model is described in detail and closed-form solutions of its AOA and TOA probability density functions of received signal are given,and analytical verification was carried out by simulation.Finally,the model derived in this thesis is fitted to the measured signals in macrocell and microcell environments using genetic algorithms,and the fitting results show that the proposed CROS/CDOS model has more realistic TOA parameters compared with the conventional model.Then,the NLOS identification algorithm using TOA parameters in outdoor cellular environment is investigated.The LOS/NLOS binary judgments are performed using the generalized likelihood ratio test and the uniformly most powerful test(UMPT)for the signals modeled in two different ways,respectively,and the likelihood ratio test for the CDOS model derived in this thesis is given.The simulation shows that this NLOS identification using the statistical properties of the signal has good results.In addition,a combined training approach is applied to the cellular environment.This approach uses TOA as the training parameter and combines the two kinds of data modeled in the references and this thesis to obtain a machine learning-based LOS/NLOS identification method by combining different base stations.The simulation results show that the support vector machine classification method is the best in both environments,and the Adaboost and random forest methods have their own Pros and Cons in different environments,and using the identified LOS signal for localization can significantly reduce the positioning error compared to the unidentified signal.Next,the multi-parameter NLOS identification method is studied in the indoor ultrawideband channel environment.The k-s test is performed by log-normal fitting of various parameters of the IEEE802.15.4a standard channel model,including the skewness and kurtosis of the impulse response,the number of significant paths of multipath signals,etc.,to obtain the LOS/NLOS identification method based on the likelihood ratio test.And for the first time,these likelihood ratio testers are combined into the Adaboost machine learning method after being optimized by the minimum mean probability of error criterion,and the simulation verifies that the classification method proposed in this thesis has higher identification accuracy with different prior probabilities.