基于多元时间序列分割聚类的异常值检测方法  

Outlier detection method based on multivariate time series segmentation clustering

作　　者：邓春宇 吴克河[1] 谈元鹏 胡杰 DENG Chun-yu;WU Ke-he;TAN Yuan-peng;HU Jie(School of Control and Computer Engineering,North China Electric Power University,Beijing 102206;Artificial Intelligence Application Department,China Electric Power Research Institute,Beijing 100192;School of Electrical and Electronic Engineering,North China Electric Power University,Beijing 102206)

机构地区：[1]华北电力大学控制与计算机工程学院,北京102206 [2]中国电力科学研究院有限公司人工智能应用研究所,北京100192 [3]华北电力大学电气与电子工程学院,北京102206

出　　处：《计算机工程与设计》2020年第11期3123-3128,共6页Computer Engineering and Design

摘　　要：为解决多元时间序列中的异常数据问题,在分析已有研究方法的基础上,提出一种基于分割聚类算法和长短期记忆网络结合的大数据异常检测方法。建立多元时间序列聚类模型,采用流水线模型和交替方向乘子法求解,得到子数据分段;使用长短期记忆网络重构各子序列,比较与原始序列的残差检测出异常数据点。以变压器监测数据为例进行异常检测,检测结果表明,该方法具有较高的检测精度。To solve the problem of abnormal data in multivariate time series,after a detailed analysis of existing methods,a big data anomaly detection method based on segmentation clustering algorithm and long-term and short-term memory network was proposed.A multivariate time series classification model was established,which was solved using the pipeline model and the alternating direction multiplier method to obtain sub-data segments.Each sub-sequence was reconstructed using a long short-term memory network,and the abnormal data points were detected by comparing the residuals with the original sequence.An abnormality detection was carried out with transformer monitoring data as examples.The results show that the proposed method has high detection accuracy.

关 键 词：异常检测 多元时间序列 分割聚类 长短期记忆网络 变压器 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]