EUVIMEDEUVIMED entdecken
Crossref

A modified evidential deep learning fault diagnosis method for detecting out-of-distribution samples

Heng Lu, Weiwei Li, Bo Xiong, Peng Han, Jinju Zhou, Jiayi Ke

Measurement Science and Technology · 05.06.2026 · 37 (23) · 236202

DOI: 10.1088/1361-6501/ae6fda ↗

Kurz gefasst

Fachartikel in der Zeitschrift Measurement Science and Technology, online erschienen am 08.06.2026. Verfasst von Heng Lu und 5 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 44 Literaturangaben. Gefördert von: the National Natural Science Foundation of China.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Deep learning-based fault diagnosis has consistently demonstrated state-of-the-art performance, making it a focal point of recent research. However, conventional deep learning methods often operate under the closed-set assumption that all test data belong to predefined fault classes. This limitation leads them to neglect epistemic (model) and aleatoric (data) uncertainties, particularly the uncertainty of out-of-distribution data caused by invisible fault or noise, compromising their reliability in real-world scenarios. Therefore, a modified evidential deep learning method capable of quantifying both epistemic and aleatoric uncertainties is proposed. Firstly, a reliable uncertainty modeling framework is established through evidence theory. Subsequently, epistemic uncertainty is quantified by the concentration parameters of the Dirichlet distribution. This allows the model to express a lack of confidence when encountering unknown data. Aleatoric uncertainty is calculated as the Dirichlet distribution’s expected entropy. This value quantifies inherent data noise and helps assess the reliability of individual predictions. Finally, separate thresholding methods are applied to epistemic and aleatoric uncertainty respectively to distinguish between unknown samples and noisy samples. The effectiveness of the proposed method is validated through two industrial case studies, demonstrating its ability to not only accurately diagnose known faults but also reliably identify unknown samples and noisy samples.

Förderung

  • the National Natural Science Foundation of China

Zitieren

APA

Lu, H., Li, W., Xiong, B., Han, P., Zhou, J., & Ke, J. (2026). A modified evidential deep learning fault diagnosis method for detecting out-of-distribution samples. Measurement Science and Technology, 37(23), 236202. https://doi.org/10.1088/1361-6501/ae6fda

Vancouver

Lu H, Li W, Xiong B, Han P, Zhou J, Ke J. A modified evidential deep learning fault diagnosis method for detecting out-of-distribution samples. Measurement Science and Technology. 2026;37(23):236202. doi:10.1088/1361-6501/ae6fda

BibTeX
@article{Lu2026,
  author = {Lu, Heng and Li, Weiwei and Xiong, Bo and Han, Peng and Zhou, Jinju and Ke, Jiayi},
  title = {A modified evidential deep learning fault diagnosis method for detecting out-of-distribution samples},
  journal = {Measurement Science and Technology},
  year = {2026},
  volume = {37},
  number = {23},
  pages = {236202},
  doi = {10.1088/1361-6501/ae6fda}
}
Kostenlos starten →