Skip to main navigation Skip to search Skip to main content

TS-Bert: Time Series Anomaly Detection via Pre-training Model Bert

  • Weixia Dang
  • , Biyu Zhou*
  • , Lingwei Wei
  • , Weigang Zhang
  • , Ziang Yang
  • , Songlin Hu
  • *Corresponding author for this work
  • University of Chinese Academy of Sciences
  • CAS - Institute of Information Engineering

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Anomaly detection of time series is of great importance in data mining research. Current state of the art suffer from scalability, over reliance on labels and high false positives. To this end, a novel framework, named TS-Bert, is proposed in this paper. TS-Bert is based on pre-training model Bert and consists of two phases, accordingly. In the pre-training phase, the model learns the behavior features of the time series from massive unlabeled data. In the fine-tuning phase, the model is fine-tuned based on the target dataset. Since the Bert model is not designed for the time series anomaly detection task, we have made some modifications thus to improve the detection accuracy. Furthermore, we have removed the dependency of the model on labeled data so that TS-Bert is unsupervised. Experiments on the public data set KPI and yahoo demonstrate that TS-Bert has significantly improved the f1 value compared to the current state-of-the-art unsupervised learning models.

Original languageEnglish
Title of host publicationComputational Science – ICCS 2021 - 21st International Conference, Proceedings
EditorsMaciej Paszynski, Dieter Kranzlmüller, Dieter Kranzlmüller, Valeria V. Krzhizhanovskaya, Jack J. Dongarra, Peter M.A. Sloot, Peter M.A. Sloot, Peter M.A. Sloot
PublisherSpringer Science and Business Media Deutschland GmbH
Pages209-223
Number of pages15
ISBN (Print)9783030779634
DOIs
StatePublished - 2021
Externally publishedYes
Event21st International Conference on Computational Science, ICCS 2021 - Virtual, Online
Duration: 16 Jun 202118 Jun 2021

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume12743 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference21st International Conference on Computational Science, ICCS 2021
CityVirtual, Online
Period16/06/2118/06/21

Keywords

  • Anomaly detection
  • Pre-training model
  • Time series analysis

Fingerprint

Dive into the research topics of 'TS-Bert: Time Series Anomaly Detection via Pre-training Model Bert'. Together they form a unique fingerprint.

Cite this