.png){kind=icon}
.png){kind=icon}
Return
EnsDiffAD: Ensemble Diffusion Models for Multivariate Time Series Anomaly Detection
.png){kind=icon}
PRE
AI
Q
Y
M
X
S
Y
DOI：10.1109/tkde.2026.3708334.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Multivariate time series anomaly detection (MTSAD) is critically important across a wide range of research fields and applications. However, achieving accurate anomaly detection in multivariate time series (MTS) presents several challenges. First, existing MTSAD methods mainly emphasize modeling long short-term temporal dependencies, while neglecting the intricate and strongly correlated inter-variable dependencies, which limits detection accuracy. Second, distributional shifts caused by external factors further hinder a single model’s ability to generalize across evolving patterns, risking missed subtle anomalies and compromising detection robustness. To address these issues, we propose EnsDiffAD, a novel MTSAD framework leveraging an ensemble of improved diffusion models. A key component of our framework is TiT-Diffusion, a denoising architecture designed to model the entangled yet distinct nature of temporal and inter-variable dependencies. Instead of directly combining existing models, TiT-Diffusion decouples these aspects into complementary modules: a temporal Transformer captures long-term and short-term temporal patterns, while an inter-variable iTransformer models global cross-variable correlations. This decomposition mitigates representation conflicts and facilitates more expressive and disentangled feature learning. Additionally, we enhance dynamics and uncertainty modeling via a novel loss function and ensemble multiple TiT-Diffusion models during training. Experiments on five real-world datasets demonstrate EnsDiffAD’s superiority.
Keywords:
Anomaly detection
Multivariate time series
Deep learning
Ensemble learning
Generative Networks
Journal
IF:
10.4
Papers:
6.7K
Citations:
3.2W
Organization
Cited Papers