增强高斯混合模型与集成学习的室内定位方法  

Indoor positioning method for enhancing Gaussian mixture model and integrated learning

作　　者：胡锐 吴飞[1] 朱海 鄢松 韩学法 金霄 HU Rui;WU Fei;ZHU Hai;YAN Song;HAN Xuefa;JIN Xiao(School of Electrical and Electronic Engineering,Shanghai University of Engineering Science,Shanghai 201620,China)

机构地区：[1]上海工程技术大学电子电气工程学院,上海201620

出　　处：《导航定位学报》2021年第3期48-54,共7页Journal of Navigation and Positioning

基　　金：国家自然科学基金资助项目(61272097);上海市科技学术委员会重点项目(18511101600);上海市自然科学基金项目(17ZR1411900);上海市信息安全综合管理技术研究重点实验室项目(AGK2015006);上海高校青年教师培养资助计划项目(ZZGCD15090)。

摘　　要：传统构建指纹数据库是在参考点收集来自接入点的接收信号强度,但由于在同一参考点上来自同一接入点的接收信号强度变化的无规律性,以及实时定位时,单一分类器对接收信号强度的分类性能差,针对此问题,提出增强高斯混合模型重建指纹数据库,并提出确定分模型个数的方法,利用多分类器投票的集成学习方法进行实时定位。在离线阶段,通过贝叶斯信息准则确定分模型个数,并利用期望最大值算法,对高斯混合模型进行参数估计,将参数估计的结果融合进指纹数据库中,即重建指纹数据库;在在线阶段,利用多种分类器进行投票决策的方式得出实时位置。实验结果表明,本文提出的方法平均定位误差为0.96 m,定位误差小于1 m的概率为92.34%,相比与增强高斯混合模型与随机森林模型,本文集成学习模型的定位精度提高了2.79%和0.92%。Traditionally,the fingerprint database is constructed by collecting Received Signal Strength(RSS)from the Access Point(AP)at the reference point.However,due to the irregularity of the RSS change from the same AP at the same reference point and the poor performance of the single classifier for RSS classification in real-time positioning,this paper proposes to enhance the Gaussian mixture model to reconstruct the fingerprint database,and proposes a method to determine the number of sub-models,using the integrated learning method of multi-classifier voting for real-time positioning.In the offline stage,the Bayesian information criterion is used to determine the number of sub-models,and the Expectation Maximization(EM)algorithm is used to estimate the parameters of the Gaussian mixture model,and the parameter estimation results are fused into the fingerprint database,that is,the fingerprint database is reconstructed;the classifier makes the voting decision to get the real-time location.The experimental results show that the proposed method has an average positioning error of 0.96 m and a probability of positioning error less than 1 m is 92.34%,which is 2.79%and 0.92%higher than that of the enhanced Gaussian mixture model and random forest and the integrated learning model of this paper.

关 键 词：增强高斯混合模型 数据库重构 集成学习 期望最大值算法 

分 类 号：P228[天文地球—大地测量学与测量工程]