基于子空间重构的无监督时间序列异常检测

Unsupervised time series anomaly detection based on subspace reconstruction

时间序列异常检测旨在寻找时间序列中不符合预期的数据,为相关人员提供有价值的信息,一直以来都受到学术界和工业界的广泛关注。然而,现有时间序列异常检测方法大多忽略了复杂数据中的多种模式,不能充分利用已有模式信息进行有效的特征学习,造成检测效果不理想。为此,本文提出了一种基于子空间重构的无监督时间序列异常检测模型。首先,将原始时间序列转换至低维潜在空间,利用高斯混合模型在潜在空间聚类,将原始时间序列分割为多个独立子空间。之后,各个子空间训练子模型,实现多模式捕获。最后,通过各个子模型重构,实现异常检测。该模型在UCR和MIT-BIH的6个数据集上的检测效果显著地优于已有方法,证明了方法的有效性。

Time series anomaly detection has long been a subject that has attracted wide attention in academia and industry,which aims to find the data that deviate significantly from the excepted behavior of time series,and then provide valuable information to those interested.However,most existing anomaly detection methods ignore the multiple patterns in complex data,and thus fail to make full use of the existing patterns information for effective feature learning,resulting in unsatisfactory detection results.To address the above problems,this paper proposes an unsupervised time series anomaly detection method based on subspace reconstruction.Firstly,the original time series is converted to latent space with lower dimensions.Then,based on the result of gaussian mixture model clustering in latent space,the original time series is divided into multiple independent subspaces.To achieve the goal of extracting multiple patterns,one sub-model is trained for each subspace separately.Finally,samples are reconstructed by all sub-models at the same time,and anomaly detection is performed based on the reconstruction errors.Experimental results on six public datasets of UCR and MIT-BIH show that the proposed method is significantly superior to the existing methods,and thus demonstrates the effectiveness of the method.