关键词: 异常检测, 多元时间序列, 权重更新, 多粒度表示, 主动学习

Abstract:

Unsupervised anomaly detection methods for Multivariable Time Series （MTS） have attracted wide attention due to their low labeling costs. However， traditional unsupervised anomaly detection methods are often based on two assumptions： 1） Independent and Identical Distribution （IID） assumption， i.e.， there is no dependency between samples and attributes of MTS. 2） High-purity starting assumption， i.e.， it is assumed that a completely normal time series should be used for training. The above assumptions are often difficult to satisfy in practical scenarios. To address this problem， an iterative active MTS anomaly detection algorithm based on margin anomaly candidate set （called EraseMTS） was proposed. Firstly， a multi-granularity representation learning method was utilized to capture the dependencies within subsequences and between subsequences， and then represent the original MTS. Secondly， a selection strategy was proposed to interact with experts based on margin anomaly candidate set， which was determined by the subsequence anomaly score and the uncertainty of its anomaly degree. Finally， an iterative subsequence weight update mechanism was designed to integrate the abnormal feedback information into the training process of the unsupervised anomaly detection model. The performance of the initial training model was continuously optimized through iteration. The proposed algorithm was verified in detection performance， scalability， and stability respectively on four datasets in UCR time series archive and one synthetic dataset. Experimental results show that the proposed algorithm can run effectively and stably.

Key words: anomaly detection, Multivariable Time Series (MTS), weight update, multi-granularity representation, active learning