基于改进LOF算法的窃电检测方法研究  

Research on detection method of electricity stealing based on improved LOF algorithm

作　　者：殷锋[1] 周绍军[2] 漆翔宇 曹旭 YIN Feng;ZHOU Shaojun;QI Xiangyu;CAO Xu(School of Computer Science and Engineering,Southwest Minzu University,Chengdu 610041,China;Department of Information Engineering,Sichuan Water Conservancy Vocational College,Chengdu 200051,China)

机构地区：[1]西南民族大学计算机科学与工程学院,成都610041 [2]四川水利职业技术学院信息工程系,成都611231

出　　处：《中南民族大学学报（自然科学版）》2022年第5期579-585,共7页Journal of South-Central University for Nationalities：Natural Science Edition

基　　金：国家社会科学基金重大招标资助项目(19ZDA284);四川省科技资助项目(2020JDR0141,2020JDRC0040);成都市哲学社会科学规划项目(2022BS027);中央高校基本科研业务费专项资金资助项目(2022SZL20)。

摘　　要：异常值检测作为数据挖掘领域研究的热点问题之一,广泛应用于窃电识别、反信息欺诈等领域.而LOF算法作为一种依赖数据密度进行异常值识别的算法,因其具有检测精度高、应用场景多元等优势常被应用于窃电识别与检测过程中,但该算法往往存在较高的时间复杂度.针对该问题,提出了一种基于混合剪枝树模型改进的RBT-LOF算法,并在此基础上提出了相应的窃电用户识别模型.RBT-LOF算法首先对混合剪枝树的超平面划分方式进行调整,采用数据第一特征向量找出平衡分割位并重构数据对象;其次使用混合剪枝查询加速数据对象的搜索.实验结果表明:基于RBT-LOF的窃电识别模型较LOF算法、SVM、CNN和WDNet模型具有更高的执行效率和检测精确率.Outlier detection,as one of the hot topics in data mining research,is widely used in the fields of electricity stealing identification,anti-information fraud and so on.The LOF algorithm,which is often used in the identification and detection of electricity theft due to its advantages of high detection accuracy and diverse application scenarios as an algorithm for outlier identification relying on data density.Aiming at this problem,an improved RBT-LOF algorithm based on the hybrid pruning tree model is proposed,and a corresponding electricity stealing user identification model is proposed on this basis.The RBT-LOF algorithm firstly adjusts the hyperplane division of the hybrid pruning tree,uses the first eigenvector of the data to find the balanced segmentation bit and reconstructs the data object;secondly,it uses the hybrid pruning query to speed up the search of the data object.The experimental results show that the power theft identification model based on RBT-LOF has higher execution efficiency and detection accuracy than the LOF algorithm,SVM,CNN and WDNet model.

关 键 词：窃电检测 RBT-LOF算法 球树模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]