Scholarlink

Accurate detection and classification of machine anomalies are essential tasks for performing machine maintenance. Recently, deep learning methods have been widely utilized for anomaly detection to improve performance. For this study, we utilized TimesNet as a feature extractor. TimesNet had achieved state-of-the-art performance in time series analysis tasks, including anomaly detection, in 2023. TimesNet enables efficient feature extraction by segmenting and stacking the vibrations at critical periods during the preprocessing. However, we found that TimesNet lacks exactness in the period length used for the segmentation. Thus, we additionally proposed two methods to address this issue and compared the accuracy of our proposed models to that of TimesNet. As a result, the proposed methods outperformed the original TimesNet.

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

Improvement of Anomaly Detection Through Enhanced Feature Extraction in TimesNet

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Hiroyoshi Nagahama,
  • Michifumi Yoshioka,
  • Katsufumi Inoue,
  • Masayoshi Todorokihara,
  • Keishi Omori

摘要

Accurate detection and classification of machine anomalies are essential tasks for performing machine maintenance. Recently, deep learning methods have been widely utilized for anomaly detection to improve performance. For this study, we utilized TimesNet as a feature extractor. TimesNet had achieved state-of-the-art performance in time series analysis tasks, including anomaly detection, in 2023. TimesNet enables efficient feature extraction by segmenting and stacking the vibrations at critical periods during the preprocessing. However, we found that TimesNet lacks exactness in the period length used for the segmentation. Thus, we additionally proposed two methods to address this issue and compared the accuracy of our proposed models to that of TimesNet. As a result, the proposed methods outperformed the original TimesNet.