基于多粒度联合孤立森林的机器人实时异常监控方法  

Real-time anomaly monitoring method for robots based on multi-granularity cascade isolation forests

作　　者：于振中[1] 洪辉武 徐斌 江瀚澄 Yu Zhenzhong;Hong Huiwu;Xu Bin;Jiang Hancheng(Internet of Things Engineering,Jiangnan University,Wuxi Jiangsu 214122,China;HRG International Institute for Research&Innovation,Hefei 230601,China)

机构地区：[1]江南大学物联网工程学院,江苏无锡214122 [2]哈工大机器人(合肥)国际创新研究院,合肥230601

出　　处：《计算机应用研究》2021年第6期1785-1789,共5页Application Research of Computers

基　　金：国家重点研发计划资助项目(2018YFB1306100)。

摘　　要：针对机器人实时异常监控问题,提出了基于多粒度联合孤立森林的实时异常监控方法。该方法对机器人的历史数据进行多粒度参数重组得到一系列具有不同参数组合的数据集,每一个数据集都可以训练一个孤立森林,最终产生了一个孤立森林集合。多个孤立森林联合投票策略是使用一个异常数据集对各个森林进行测试,求出异常数据集在每个孤立森林上的平均异常分数,并由此确定每个孤立森林在联合孤立森林模型总体决策中的话语权。以机器人的碰撞异常作为监控对象对该方法进行评价,监控准确率达99.8%,且报警平均延迟仅为26.72 ms,说明该方法能够有效地实现机器人实时异常监控。Aiming at the problem of robot real-time anomaly monitoring,this paper proposed a real-time anomaly monitoring method based on multi-granularity cascade isolation forests.This method recombined historical data of the robot to obtain a series of data sets with different parameter combinations.Each data set could train an isolation forest,and finally generated an isolation forest set.The strategy of the joint voting decision of multiple isolation forests was to test each forest with an abnormal data set,calculate the average abnormal score of the abnormal data set on each isolation forest,and then determine the discourse power of each isolation forest in the overall decision of the cascade isolation forest model.This paper used collision anomaly of robot as monitoring object to test the method,the experimental results show that the monitoring accuracy is up to 99.8%,and the average alarm delay is only 26.72 ms,indicating that this method can effectively realize the real-time monitoring of robot anomalies.

关 键 词：孤立森林 机器人 异常监测 多粒度联合 

分 类 号：TP393.04[自动化与计算机技术—计算机应用技术]