Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Open-Set Multivariate Time-Series Anomaly Detection

Thomas Lai, Thi Kieu Khanh Ho, Narges Armanfard

TL;DR

This work tackles open-set multivariate time-series anomaly detection (OS-TSAD) by introducing MOSAD, a three-headed framework with a shared feature extractor tailored for end-to-end detection of seen and unseen anomalies. MOSAD combines a Generative head (masked reconstruction), a Discriminative head (deviation-based scoring), and an Anomaly-Aware Contrastive head (normal-focused contrastive Learning) to create a rich representation space and robust anomaly scoring via s = s_rec + s_dev + s_con. The authors demonstrate state-of-the-art performance on three real-world datasets (SMD, PTB-XL, TUSZ) under general and hard open-set settings, with extensive ablations validating the contribution of each head and the anomaly augmentation techniques COE and WMix. The method’s open-set capability and threshold-independent evaluation offer practical advantages for deploying TSAD systems in domains where labeled anomalies are scarce and unseen anomalies are expected, such as industrial monitoring and biomedical signal analysis.

Abstract

Numerous methods for time-series anomaly detection (TSAD) have emerged in recent years, most of which are unsupervised and assume that only normal samples are available during the training phase, due to the challenge of obtaining abnormal data in real-world scenarios. Still, limited samples of abnormal data are often available, albeit they are far from representative of all possible anomalies. Supervised methods can be utilized to classify normal and seen anomalies, but they tend to overfit to the seen anomalies present during training, hence, they fail to generalize to unseen anomalies. We propose the first algorithm to address the open-set TSAD problem, called Multivariate Open-Set Time-Series Anomaly Detector (MOSAD), that leverages only a few shots of labeled anomalies during the training phase in order to achieve superior anomaly detection performance compared to both supervised and unsupervised TSAD algorithms. MOSAD is a novel multi-head TSAD framework with a shared representation space and specialized heads, including the Generative head, the Discriminative head, and the Anomaly-Aware Contrastive head. The latter produces a superior representation space for anomaly detection compared to conventional supervised contrastive learning. Extensive experiments on three real-world datasets establish MOSAD as a new state-of-the-art in the TSAD field.

Open-Set Multivariate Time-Series Anomaly Detection

TL;DR

This work tackles open-set multivariate time-series anomaly detection (OS-TSAD) by introducing MOSAD, a three-headed framework with a shared feature extractor tailored for end-to-end detection of seen and unseen anomalies. MOSAD combines a Generative head (masked reconstruction), a Discriminative head (deviation-based scoring), and an Anomaly-Aware Contrastive head (normal-focused contrastive Learning) to create a rich representation space and robust anomaly scoring via s = s_rec + s_dev + s_con. The authors demonstrate state-of-the-art performance on three real-world datasets (SMD, PTB-XL, TUSZ) under general and hard open-set settings, with extensive ablations validating the contribution of each head and the anomaly augmentation techniques COE and WMix. The method’s open-set capability and threshold-independent evaluation offer practical advantages for deploying TSAD systems in domains where labeled anomalies are scarce and unseen anomalies are expected, such as industrial monitoring and biomedical signal analysis.

Abstract

Numerous methods for time-series anomaly detection (TSAD) have emerged in recent years, most of which are unsupervised and assume that only normal samples are available during the training phase, due to the challenge of obtaining abnormal data in real-world scenarios. Still, limited samples of abnormal data are often available, albeit they are far from representative of all possible anomalies. Supervised methods can be utilized to classify normal and seen anomalies, but they tend to overfit to the seen anomalies present during training, hence, they fail to generalize to unseen anomalies. We propose the first algorithm to address the open-set TSAD problem, called Multivariate Open-Set Time-Series Anomaly Detector (MOSAD), that leverages only a few shots of labeled anomalies during the training phase in order to achieve superior anomaly detection performance compared to both supervised and unsupervised TSAD algorithms. MOSAD is a novel multi-head TSAD framework with a shared representation space and specialized heads, including the Generative head, the Discriminative head, and the Anomaly-Aware Contrastive head. The latter produces a superior representation space for anomaly detection compared to conventional supervised contrastive learning. Extensive experiments on three real-world datasets establish MOSAD as a new state-of-the-art in the TSAD field.
Paper Structure (21 sections, 7 equations, 2 figures, 7 tables)

This paper contains 21 sections, 7 equations, 2 figures, 7 tables.

Table of Contents

  1. Proposed Method
  2. Feature Extractor
  3. Multi-head Network
  4. Generative Head
  5. Discriminative Head
  6. Anomaly-Aware Contrastive Head
  7. Anomaly Scoring
  8. Detecting Unseen Anomalies
  9. Experiments
  10. Experimental Settings
  11. Datasets
  12. Baselines
  13. Implementation Details
  14. Experimental Results
  15. General Setting
  16. ...and 6 more sections

Figures (2)

  • Figure 1: The overall framework of MOSAD.
  • Figure 2: (Left) $|\mathcal{P}|$ versus APR under O$^{\text{Class}}_H$ on PTB-XL. (Right) Anomaly augmentation techniques versus APR under O$^{\text{MI}}_H$.