Scholarlink

<p>Anomaly detection aims to identify anomalies and uncover their potential implications, a challenging problem that has drawn significant research attention. This study presents a multi-scale framework based on the principle of justifiable granularity, incorporating two distinct multi-scale approaches: type-1 and type-2. The type-1 approach employs bottom-up granular representation to construct higher-order information granules, while the type-2 approach generates various higher-type granules iteratively from the original numeric data. Two similarity measurement algorithms are developed to support anomaly detection for each approach, respectively. Extensive experimental studies on various datasets, including periodic and non-periodic patterns, demonstrate the proposed framework’s robustness and versatility. The results show that the type-1 and type-2 multi-scale approaches outperform state-of-the-art methods across several performance metrics, effectively handling data with trends, instability, and diverse anomaly characteristics.</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

A multi-scale granular framework for detecting anomalies in time series data

  • 期刊论文

  • 发表日期：

    2025年8月8日

查看全文
  • Junjun Yang,
  • Bin Yang,
  • Yanjun Zhou

摘要

Anomaly detection aims to identify anomalies and uncover their potential implications, a challenging problem that has drawn significant research attention. This study presents a multi-scale framework based on the principle of justifiable granularity, incorporating two distinct multi-scale approaches: type-1 and type-2. The type-1 approach employs bottom-up granular representation to construct higher-order information granules, while the type-2 approach generates various higher-type granules iteratively from the original numeric data. Two similarity measurement algorithms are developed to support anomaly detection for each approach, respectively. Extensive experimental studies on various datasets, including periodic and non-periodic patterns, demonstrate the proposed framework’s robustness and versatility. The results show that the type-1 and type-2 multi-scale approaches outperform state-of-the-art methods across several performance metrics, effectively handling data with trends, instability, and diverse anomaly characteristics.