基于耦合自适应距离的高维异常检测算法

High-dimensional anomaly detection algorithm based on coupling-adaptive distance

距离聚类方法是航天器等复杂系统实现遥测参数异常检测的常用方法之一,但在面对高维遥测数据进行异常检测任务时,往往会暴露出效率低下、精度劣化等严重问题。针对基于高维遥测数据的航天器异常检测难题,提出了一种基于耦合自适应的改进距离定义,并针对归纳监视系统(IMS)算法这一经典距离聚类算法进行了改进。该方法利用历史数据的分布特征,在进行聚类的同时,对于参数耦合性进行动态挖掘,并将挖掘到的知识高效地投入到异常检测任务。最后,采用运载火箭电源系统的真实高维遥测数据对所提方法进行了应用验证。在与多种传统基于IMS的异常检测方法的对比实验中,该改进算法检测效率与准确率较另两类IMS算法中的最优方法分别提升了41.83%和69.03%,验证了运用该距离定义的检测方法在效率与精确率上的优越性。

The distance-based clustering is one of the common methods to realize the anomaly detection of telemetry parameters in complex systems,such as spacecraft.However,when it is applied to high-dimensional remote measurement data,it often exposes serious problems,such as low efficiency and degraded accuracy.To overcome the difficulty in anomaly detection on the high-dimensional telemetry data,this article proposes an improved distance definition based on coupling adaptation.The inductive monitoring system(IMS)algorithm which is a classical distance clustering algorithm is improved.Based on the intrinsic distribution characteristics of historical telemetry data,this method mines dynamically the couplings among telemetry parameters while clustering.Then,it takes efficiently advantage of this mined knowledge of telemetry parameters’couplings into the following task of anomaly detection.Finally,this article evaluates the application of the proposed method on a high-dimensional telemetry data of a real rocket power supply system.Compared with a variety of classic high-dimensional anomaly detection methods based on IMS algorithms,this article demonstrates its advantages for high-dimensional anomaly detection as well,which is 69.03%and 41.83%better than the best method in other two categories of IMS algorithms respectively on efficacy and accuracy of anomaly detection.It shows the superiority of the detection methods using the proposed distance definition in efficiency and accuracy.