井下WLAN位置指纹定位中改进区域划分方法研究

井下WLAN位置指纹人员定位系统主要是通过聚类算法来实现位置指纹样本的整体性划分,但现有的聚类算法只是针对接收信号强度的统计分布特性进行聚类划分,并没有充分考虑奇点问题。针对该问题,提出了一种基于类关系的K-Means(CRK-Means)算法,该算法以类内离散度和类间离散度的比值为目标函数,通过使该比值最小的聚类的聚合、分离过程即可得到避免了奇点问题的最优聚类,完成定位区域的合理划分。针对采用随机森林(RF)算法对聚类划分后的定位区域进行粗定位存在误判的问题,提出了遗传算法与随机森林相结合的(GA-RF)算法,该算法以GA中的选择、交叉和变异优化过程确保了RF算法的选择树总数和位置指纹参考点特征数的最优取值。实验结果表明:CRK-Meams算法有效解决了奇点问题,且在一定程度上提升了系统定位精度;采用CRK-Meams算法和GA-RF算法后,子区域粗定位的准确率相比RF算法提升了4%,达到98%;置信概率大于90%的最小定位误差达到了3m,优于传统的聚类算法。

Research on improved region division method in underground WLAN location fingerprints positioning

Underground WLAN location fingerprinting personnel positioning system mainly realizes overall division of location fingerprinting samples through clustering algorithm,but existing clustering algorithm only carries out the clustering division according to the statistical distribution characteristics of received signal strength(RSS),and does not fully consider singularity problem.For the above problem,a class relationship K-Means(CRK-Means)algorithm was proposed.CRK-Means algorithm takes the ratio of intra class dispersion and inter class dispersion as the objective function,and the optimal clustering without singularity problem can be achieved by aggregation and separation process of clustering with the minimum ratio,so as to complete reasonable division of positioning area.Genetic Algorithm-Random Forets(GA-RF)algorithm was proposed to solve the problem of misjudgment in rough localization of clustering area by using Random Forest(RF)algorithm.The optimization process of selection,crossover and mutation in GA ensures the optimal value of the total number of selection trees and the feature number of location fingerprints reference points in RF algorithm.The experimental results show that the CRKMeans algorithm solves the singularity problem effectively,and improves the positioning accuracy of the positioning system.The accuracy of sub-region rough positioning by CRK-Means algorithm and GA-RF algorithm is 4% and higher than RF algorithm,it is 98%.The minimum positioning error with a confidence probability greater than 90%is 3 m,which is better than the traditional clustering algorithms.