Scholarlink

Anomaly detection is vital in industrial applications but remains challenging due to the rarity and complexity of anomalies. Existing methods either reconstruct input sequences or learn discriminative representations, each with inherent limitations. In this paper, we propose a hybrid framework that integrates reconstruction-based and representation-based approaches for time series anomaly detection. Specifically, we leverage a pretrained large language model, GPT2, as a reconstruction module within a dual-branch architecture inspired by DCdetector. GPT2-generated reconstructions are used to enhance representation discrepancy and improve detection performance. Extensive experiments on real-world datasets demonstrate the superiority of our approach over state-of-the-art methods.

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

Time Series Anomaly Detection via LLM-Based Reconstruction and Contrastive Representation Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yusong Zhou,
  • Weiping Zheng,
  • Huile Wang

摘要

Anomaly detection is vital in industrial applications but remains challenging due to the rarity and complexity of anomalies. Existing methods either reconstruct input sequences or learn discriminative representations, each with inherent limitations. In this paper, we propose a hybrid framework that integrates reconstruction-based and representation-based approaches for time series anomaly detection. Specifically, we leverage a pretrained large language model, GPT2, as a reconstruction module within a dual-branch architecture inspired by DCdetector. GPT2-generated reconstructions are used to enhance representation discrepancy and improve detection performance. Extensive experiments on real-world datasets demonstrate the superiority of our approach over state-of-the-art methods.