View in EDS

SAformer: A time series anomaly detection model based on Similarity-Aware attention.

Saved in:
Bibliographic Details
Title: SAformer: A time series anomaly detection model based on Similarity-Aware attention.
Authors: Tang, Yunkai1 (AUTHOR) tangyk@cug.edu.cn, Li, Guiling1,2 (AUTHOR) guiling@cug.edu.cn, Wu, Zongda3 (AUTHOR) zongda1983@163.com, Yu, Philip S.4 (AUTHOR) psyu@uic.edu
Source: Knowledge-Based Systems. Jun2026, Vol. 344, pN.PAG-N.PAG. 1p.
Subjects: Outlier detection, Deep learning, Pattern perception
Abstract: Time series anomaly detection has become a prominent research topic in recent years. Reconstruction-based methods have demonstrated strong performance on this task. However, existing approaches still struggle to detect shapelet anomalies and multi-peaked anomalies, as traditional attention mechanisms fail to comprehensively capture potential abnormal patterns. To address this limitation, we propose SAformer , a novel time series anomaly detection model based on S imilarity- A ware attention. SAformer incorporates shape similarity between subsequences to enhance traditional attention mechanisms and leverages a global dictionary to improve the detection of multi-peaked anomalies. In addition, it introduces the concepts of strong and weak similarity domains to further enrich similarity representation, while integrating sequence-level information to refine anomaly scores. Experimental results on five public benchmark datasets demonstrate that SAformer achieves state-of-the-art performance in anomaly detection, validating the effectiveness and robustness of the proposed SAformer attention mechanism. • SAformer fuses inter-subsequence similarity with Transformer. • Similarity-guided loss regulates attention patterns for effective learning. • Sequence-level information ensures sensitive anomaly scoring. • Extensive experiments on real datasets validate the model's effectiveness. [ABSTRACT FROM AUTHOR]
Copyright of Knowledge-Based Systems is the property of Elsevier B.V. and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database: Engineering Source
Description
Abstract:Time series anomaly detection has become a prominent research topic in recent years. Reconstruction-based methods have demonstrated strong performance on this task. However, existing approaches still struggle to detect shapelet anomalies and multi-peaked anomalies, as traditional attention mechanisms fail to comprehensively capture potential abnormal patterns. To address this limitation, we propose SAformer , a novel time series anomaly detection model based on S imilarity- A ware attention. SAformer incorporates shape similarity between subsequences to enhance traditional attention mechanisms and leverages a global dictionary to improve the detection of multi-peaked anomalies. In addition, it introduces the concepts of strong and weak similarity domains to further enrich similarity representation, while integrating sequence-level information to refine anomaly scores. Experimental results on five public benchmark datasets demonstrate that SAformer achieves state-of-the-art performance in anomaly detection, validating the effectiveness and robustness of the proposed SAformer attention mechanism. • SAformer fuses inter-subsequence similarity with Transformer. • Similarity-guided loss regulates attention patterns for effective learning. • Sequence-level information ensures sensitive anomaly scoring. • Extensive experiments on real datasets validate the model's effectiveness. [ABSTRACT FROM AUTHOR]
ISSN:09507051
DOI:10.1016/j.knosys.2026.116087