基于时频帧上下文建模的重构式多维时序数据异常检测
Time-Frequency Frame Context Modeling for Reconstruction-Based Multivariate Time-Series Anomaly Detection
王欣玮 1李佳颐 2温自强 1徐鹏1
作者信息
- 1. 网络与交换技术全国重点实验室，北京邮电大学，100876
- 2. 中国工业互联网研究院，100102
- 折叠
摘要
重构式异常检测由于无需密集异常标注、能够学习正常模式而被广泛应用于多变量时间序列。短时傅里叶变换（STFT）等时频表示有助于揭示具有频率结构的异常；但 STFT 会将连续序列转换为短时帧序列，若对帧进行独立处理，往往会造成帧级时间上下文不足，从而削弱对跨帧异常或频谱渐变型异常的检测能力。本文提出一个简洁的重构式框架，在时频域中对 STFT 帧序列显式建模时间上下文：以对数幅度谱为重构目标，先进行轻量跨变量融合，再在每个频率通道上沿帧维建模帧序列依赖（采用紧凑的序列编码器实现）。随后将重构幅度与原始相位组合得到复谱，并通过逆 STFT 回到时域，以时域重构误差作为异常分数。在 Server Machine Dataset（SMD）上的实验表明，与 PCA、Autoencoder 和 CATCH 相比，引入时频域帧级时间上下文能够提升重构式多变量时间序列异常检测性能。
Abstract
Reconstruction-based anomaly detection is widely used for multivariate time series because it can learn normal patterns without requiring dense anomaly labels. Time-frequency representations such as the short-time Fourier transform (STFT) further help reveal frequency-structured irregularities; however, STFT converts a continuous sequence into a sequence of short-time frames, and processing frames independently often leads to insufficient frame-level temporal context, degrading the detection of anomalies that span multiple frames or exhibit gradual spectral changes. In this paper, we propose a simple reconstruction-based framework that explicitly models temporal context over STFT frames in the time-frequency domain. We reconstruct the log-magnitude spectrogram, apply lightweight cross-variable mixing, and model frame-sequence dependencies along the frame dimension for each frequency channel using a compact sequence encoder. The reconstructed magnitude is combined with the original phase to form a complex spectrum, and an inverse STFT is used to obtain time-domain reconstructions for anomaly scoring via reconstruction errors. Experiments on the Server Machine Dataset (SMD) show that incorporating frame-level temporal context in the time-frequency domain improves reconstruction-based multivariate time-series anomaly detection compared with PCA, Autoencoder, and CATCH.关键词
多维时序数据/异常检测/重构式方法/时频表示Key words
multivariate time series/anomaly detection/reconstruction-based methods/time-frequency representation引用本文复制引用
王欣玮,李佳颐,温自强,徐鹏.基于时频帧上下文建模的重构式多维时序数据异常检测[EB/OL].(2026-01-22)[2026-01-25].http://www.paper.edu.cn/releasepaper/content/202601-51.学科分类
计算技术、计算机技术
评论