EUVIMEDEUVIMED entdecken
Crossref

EnsDiffAD: Ensemble Diffusion Models for Multivariate Time Series Anomaly Detection

Qian Ma, Yanyang Li, Mei Bai, Xite Wang, Shikai Guo, Yu Gu, Ge Yu

IEEE Transactions on Knowledge and Data Engineering · 2026-09 · 38 (9) · 5841-5853

DOI: 10.1109/tkde.2026.3708334 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Knowledge and Data Engineering, 2026-09. Verfasst von Qian Ma und 6 weiteren Autor:innen. Die Arbeit nennt 56 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Applied Basic Research Project of Liaoning Province, Liaoning Provincial Natural Science Foundation Project.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

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.

Förderung

  • National Natural Science Foundation of China
  • Applied Basic Research Project of Liaoning Province
  • Liaoning Provincial Natural Science Foundation Project
  • Dalian Science and Technology Innovation Fund project
  • Open Research Project of the State Key Laboratory of Maritime Technology and Safety
  • Fundamental Research Funds for the Central Universities

Zitieren

APA

Ma, Q., Li, Y., Bai, M., Wang, X., Guo, S., Gu, Y., & Yu, G. (2026). EnsDiffAD: Ensemble Diffusion Models for Multivariate Time Series Anomaly Detection. IEEE Transactions on Knowledge and Data Engineering, 38(9), 5841-5853. https://doi.org/10.1109/tkde.2026.3708334

Vancouver

Ma Q, Li Y, Bai M, Wang X, Guo S, Gu Y, et al. EnsDiffAD: Ensemble Diffusion Models for Multivariate Time Series Anomaly Detection. IEEE Transactions on Knowledge and Data Engineering. 2026;38(9):5841-5853. doi:10.1109/tkde.2026.3708334

BibTeX
@article{Ma2026,
  author = {Ma, Qian and Li, Yanyang and Bai, Mei and Wang, Xite and Guo, Shikai and Gu, Yu and Yu, Ge},
  title = {EnsDiffAD: Ensemble Diffusion Models for Multivariate Time Series Anomaly Detection},
  journal = {IEEE Transactions on Knowledge and Data Engineering},
  year = {2026},
  volume = {38},
  number = {9},
  pages = {5841--5853},
  doi = {10.1109/tkde.2026.3708334}
}
Kostenlos starten →