摘要
针对传统的单分类器不适用于周期时间序列的异常检测,提出了一种基于移相加权球面单簇聚类的单分类器PS-WS1M-OCC.通过在聚类过程中增加高效的循环移位操作,解决了时间序列记录之间相似度计算的问题.另一方面,基于时间序列记录的权重分布,提出了新的阈值自适应确定方法,从而使单分类器对训练集包含的异常数据和参数设置不敏感.实验表明,本文提出的单分类器可以用于周期时间序列的异常检测;与传统的单分类器相比,可以成功地从包含异常数据的训练集中进行无监督学习,对训练集包含的异常数据鲁棒,并且对参数不敏感.
The traditional one-class classifiers are not suitable for detecting discords in periodic time series. A novel one-class classifier PS-WSIM-OCC is proposed in this paper. In our method, the phase problem in time series is solved by introducing phase shift into the clustering procedure. Meanwhile, a novel criterion for adaptively choosing threshold is proposed. In this way, the proposed classifier is insensitive to noise inthe training set. Experimental results show that our PS-WSKM-OCC is more robust than the existing one-class classifiers when it is applied to the problem of discord detection in the periodic time series.
出处
《自动化学报》
EI
CSCD
北大核心
2011年第8期984-992,共9页
Acta Automatica Sinica
基金
国家自然科学基金(90820002,60903100)
江苏省自然科学基金(BK2009067)
中央高校基本科研业务费专项基金(JUSRP21128)资助~~
