Scholarlink

The detection of anomalies in time series is a critical task across various domains including time series generated by a diversity of distributed systems. Despite extensive research, no single algorithm is the undisputed leader in all respects. This paper introduces KDEAE, a novel unsupervised approach for univariate time series anomaly detection. This method combines autoencoders (AE) and kernel density estimation (KDE). The proposed approach segments time series into non-overlapping segments, encodes them using convolutional AE and constructs anomaly scores based on their density obtained from KDE. The method is evaluated on GutenTAG dataset collection using TimeEval framework and compared with 25 algorithms of the same category. The results show that KDEAE outperform 40% algorithms by ROC AUC and 60% by PR AUC scores and 60% by runtime. The paper concludes with suggestions for further research to enhance the method’s performance and extend it to multivariate time series.

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

Autoencoder and Kernel Density Estimation Based Approach for Time Series Anomaly Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Anton Arzha,
  • Vladimir Korkhov

摘要

The detection of anomalies in time series is a critical task across various domains including time series generated by a diversity of distributed systems. Despite extensive research, no single algorithm is the undisputed leader in all respects. This paper introduces KDEAE, a novel unsupervised approach for univariate time series anomaly detection. This method combines autoencoders (AE) and kernel density estimation (KDE). The proposed approach segments time series into non-overlapping segments, encodes them using convolutional AE and constructs anomaly scores based on their density obtained from KDE. The method is evaluated on GutenTAG dataset collection using TimeEval framework and compared with 25 algorithms of the same category. The results show that KDEAE outperform 40% algorithms by ROC AUC and 60% by PR AUC scores and 60% by runtime. The paper concludes with suggestions for further research to enhance the method’s performance and extend it to multivariate time series.