Scholarlink

<p>The critical task of detecting anomalies in multivariate time-series data faces challenges due to the lack of anomaly labels and the unpredictable nature of the data. Despite deep learning’s advancements in improving anomaly detection, few models adeptly handle these complexities. This study introduces a novel deep learning model called DASTAD, which is designed to effectively detect anomalies in multivariate time-series data. Unlike current models that focus on time-based differences but overlook the connections between different univariate time-series, DASTAD utilizes transformer-based attention mechanism to process the time-series data both across time and between features. This dual approach ensures a deeper understanding of the data’s dynamic patterns. The model further overcomes the limitations of conventional encoder–decoder approaches by incorporating self-conditioning and adversarial training, ensuring reliable feature extraction and training stability. The model has been extensively tested on six publicly accessible datasets. The results show that it outperforms the current leading approaches in terms of detection and diagnostic accuracy. Additionally, the model achieves these results while using less data for training. Our approach specifically enhances F1 scores by a maximum of 7.97%.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DASTAD: Dual Aspect Self-supervised Transformer-based Anomaly Detection in Multivariate Time-Series

  • 期刊论文

  • 发表日期：

    2025年5月23日

查看全文
  • Kartik Aggarwal,
  • Kella Sowmya,
  • K. Ramesh

摘要

The critical task of detecting anomalies in multivariate time-series data faces challenges due to the lack of anomaly labels and the unpredictable nature of the data. Despite deep learning’s advancements in improving anomaly detection, few models adeptly handle these complexities. This study introduces a novel deep learning model called DASTAD, which is designed to effectively detect anomalies in multivariate time-series data. Unlike current models that focus on time-based differences but overlook the connections between different univariate time-series, DASTAD utilizes transformer-based attention mechanism to process the time-series data both across time and between features. This dual approach ensures a deeper understanding of the data’s dynamic patterns. The model further overcomes the limitations of conventional encoder–decoder approaches by incorporating self-conditioning and adversarial training, ensuring reliable feature extraction and training stability. The model has been extensively tested on six publicly accessible datasets. The results show that it outperforms the current leading approaches in terms of detection and diagnostic accuracy. Additionally, the model achieves these results while using less data for training. Our approach specifically enhances F1 scores by a maximum of 7.97%.