|
| 基于TranAD-FD的多变量时间序列异常检测方法 |
| A Multivariate Time Series Anomaly Detection Method Based on TranAD-FD |
| 投稿时间：2025-09-11 修订日期：2025-11-18 |
| DOI： |
| 中文关键词: 特征融合 异常检测 频域 多变量时间序列 |
| 英文关键词: Feature fusion Anomaly detection Frequency domain Multivariate time series |
| 基金项目:全龄友好完整社区建设及居家适老化环境提升关键技术研究与应用 |
|
| 摘要点击次数: 44 |
| 全文下载次数: 0 |
| 中文摘要: |
| 针对面对标签缺失时无监督学习难以辨识多变量时间序列数据中正常与异常数据的问题，提出一种基于TranAD-FD架构的多变量时间序列异常检测方法。利用FD模块与TranAD相结合，能够有效地避免捕获短期时间趋势的问题。采取滑动窗口划分时间序列，特征融合技术融合多个时间窗的特征，并通过多头自注意机制对其进行加权融合，可以增强模型的性能和泛化能力，进而解决上述问题。实验结果表明，该方法不仅具备较高的训练效率与良好的稳定性，而且在精确率、召回率、F1 值等核心异常检测性能指标上，表现也优于当前已有的生成式异常检测模型。 |
| 英文摘要: |
| To address the issue that unsupervised learning struggles to distinguish normal from abnormal data in multivariate time series when labels are missing, a multivariate time series anomaly detection method based on the TranAD-FD architecture is proposed. By combining the FD module with TranAD, it can effectively avoid the problem of capturing short-term time trends. Time series are divided into windows using a sliding window approach, and feature fusion technology is employed to integrate features from multiple time windows, which not only has high training efficiency and good stability, but also outperforms the existing generative anomaly detection models in core anomaly detection performance indicators such as precision rate, recall rate and F1 value. |
|
View Fulltext
查看/发表评论 下载PDF阅读器 |
| 关闭 |
