Scholarlink

<p>This research presents a time series representation learning method based on unsupervised and self-supervised contrastive learning, which does not require complete data labeling. Important model differences are compared for time series to evaluate the impact on anomaly detection, namely, the form of the encoder and whether supervised fine-tuning is applied. Different representations are obtained, frozen ones from unsupervised training that are independent of any downstream task and liquid representations produced with self-supervised learning, enhanceable for specific end-tasks. Also, a simple, purely empirical procedure is implemented to detect anomalies. The methodology is applied over sensor-based data gathering devices from human activities recognition, ECG readings and industrial processes.</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

Time series anomaly detection models with contrastive representation learning

  • 期刊论文

  • 发表日期：

    2026年7月29日

查看全文
  • Francisca Quijada,
  • Klim Drobnyh,
  • George Runger

摘要

This research presents a time series representation learning method based on unsupervised and self-supervised contrastive learning, which does not require complete data labeling. Important model differences are compared for time series to evaluate the impact on anomaly detection, namely, the form of the encoder and whether supervised fine-tuning is applied. Different representations are obtained, frozen ones from unsupervised training that are independent of any downstream task and liquid representations produced with self-supervised learning, enhanceable for specific end-tasks. Also, a simple, purely empirical procedure is implemented to detect anomalies. The methodology is applied over sensor-based data gathering devices from human activities recognition, ECG readings and industrial processes.