A Novel Sliding Window-Based Model for Outlier Detection in Multivariate Time Series

Articles in Press

Document Type : Original Article

Authors

1 Department of Mathematics and Computer Sciences, Damghan University, Damghan, Iran

2 Department of Engineering, Damghan University, Damghan, Iran

10.22052/scj.2024.253155.1160

Abstract

Anomaly detection in multivariate time series has been an active research area due to its widespread application in various fields. Window-based methods are popular in the anomaly detection domain. These methods identify anomalous windows rather than specific anomalous points, even if not all points within the window are anomalies. It is a critical limitation of window-based methods. We propose an unsupervised sliding window-based model for detecting anomalies in multivariate time series to address this limitation. Our model employs a sliding mechanism to iterate through the input time series multiple times and utilizes a consensus function to aggregate different window anomaly scores. This mechanism facilitates the discovery of more anomalous subsequences, even if they are not precisely confined within a specific window. To evaluate the performance of the proposed method, several experiments on synthetic and real-world datasets, including SKAB and MSL, with multiple indices. The results confirm the superiority of the proposed method. The method achieves an Fscore of 0.902 for SKAB and 0.620 for MSL, which are twice as good as the results achieved by other methods.

Keywords

Main Subjects

Soft Computing Journal

Articles in Press, Accepted Manuscript
Available Online from 02 October 2023

Files

History

  • Receive Date: 26 June 2023
  • Revise Date: 29 August 2023
  • Accept Date: 30 September 2023

Share

How to cite

Statistics