An extreme learning machine for unsupervised online anomaly detection in multivariate time series

Abstract

Unsupervised anomaly detection in time series remains challenging, due to the rare and complex patterns of anomalous data. Previous change point detection methods based on extreme learning machine and mutual information (ELM-MI) are potential solutions for this problem. However, the kernels in these methods are randomly initialized on test data, which imposes a constraint that these methods can only be used for offline inference. Moreover, these methods are limited in utilizing temporal contexts, and require the ensemble of multiple models to improve the robustness. To tackle these problems, we introduce a multivariate ELM-MI framework, and combine it with a dynamic kernel selection method, which performs a hierarchical clustering procedure on unlabeled training data and utilizes the clusters to determine the kernels in ELM-MI. In this way, our method can tackle the unsupervised online detection of various anomalous (e.g., point anomalies and group anomalies) and reduce the computational cost. Extensive experiments on three public datasets and our collection of real-life 4G Long-Term Evolution data demonstrate that the proposed method outperforms state-of-the-art methods in terms of effectiveness and efficiency. For demo, see this link: https://personal.ntu.edu.sg/ezplin/NC-demo.htm.