EUVIMEDEUVIMED entdecken
Crossref

Unsupervised Industrial Time-Series Anomaly Detection by Fusing Raw–Differential Dual Views with Adaptive Thresholding

Shengkai Qu

Advances in Engineering Technology Research · 30.06.2026 · 17 (1) · 896

DOI: 10.56028/aetr.17.1.896.2026 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Advances in Engineering Technology Research, online erschienen am 30.06.2026. Verfasst von Shengkai Qu.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

As smart factories proliferate, ensuring uninterrupted, reliable operation of industrial equipment is increasingly critical. This paper addresses unsupervised anomaly detection for multivariate multi-sensor industrial time series and proposes a lightweight, dual‑view 1D convolutional autoencoder tailored for predictive maintenance. Using a sliding‑window sampling scheme, the model is trained on normal operation data to learn baseline behavior. To better capture dynamic anomalies, the input adopts two complementary views: the raw signal (describing operating state) and its first‑order difference (capturing temporal change and trends). Reconstruction errors from the two views are combined via a weighted dual‑score fusion scheme to produce a unified anomaly score. An adaptive thresholding policy and a consecutive‑alarm mechanism are introduced to suppress false positives in practical industrial deployments. We validate the approach on a multi‑sensor industrial compressor dataset and on synthetic industrial time series, and perform ablation studies on threshold strategy, window length, network depth, and loss design. Results show the proposed method substantially improves Precision, Recall, and F1‑score while keeping model size and inference cost low. The method’s simple architecture and high computational efficiency make it well suited for real‑world engineering deployment.

Zitieren

APA

Qu, S. (2026). Unsupervised Industrial Time-Series Anomaly Detection by Fusing Raw–Differential Dual Views with Adaptive Thresholding. Advances in Engineering Technology Research, 17(1), 896. https://doi.org/10.56028/aetr.17.1.896.2026

Vancouver

Qu S. Unsupervised Industrial Time-Series Anomaly Detection by Fusing Raw–Differential Dual Views with Adaptive Thresholding. Advances in Engineering Technology Research. 2026;17(1):896. doi:10.56028/aetr.17.1.896.2026

BibTeX
@article{Qu2026,
  author = {Qu, Shengkai},
  title = {Unsupervised Industrial Time-Series Anomaly Detection by Fusing Raw–Differential Dual Views with Adaptive Thresholding},
  journal = {Advances in Engineering Technology Research},
  year = {2026},
  volume = {17},
  number = {1},
  pages = {896},
  doi = {10.56028/aetr.17.1.896.2026}
}
Kostenlos starten →