EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Unsupervised detection of overshooting defects using hybrid generative and boundary classification models

Paula Arcano-Bea, Anabel Díaz-Labrador, Alejandro Vidal-Bralo, Francisco Zayas-Gato, Héctor Quintián

Journal of Logic and Computation · 02.11.2026 · 36 (8)

DOI: 10.1093/logcom/exag054 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Logic and Computation, online erschienen am 06.11.2026. Verfasst von Paula Arcano-Bea und 4 weiteren Autor:innen. Die Arbeit nennt 15 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Quality control in modern manufacturing increasingly relies on unsupervised anomaly detection to identify structural defects when abnormal samples are critically scarce. This study presents an evaluation of hybrid generative architectures designed to detect complex overshooting anomalies in industrial simple sub-assemblies base plates. We analysed eight distinct model combinations by pairing four advanced generative backbones, including a latent space resampling autoencoder and a denoising diffusion model, with two boundary classification strategies: the one class support vector machine and the deep support vector data description. The experimental results demonstrate that the configuration pairing the latent space resampling autoencoder with the one class support vector machine achieves superior detection capabilities, securing the highest mean area under the curve of 0.7870 alongside absolute statistical dominance. However, our findings reveal a critical architectural vulnerability, as pairing heavily regularized generative latent spaces with the rigid hyperspherical boundaries of the deep data description model invariably degrades detection, driving in the most strongly regularized configuration a systematic inversion of the anomaly score rather than a loss of discriminative power. Ultimately, this research establishes that flexible margin optimization is required to preserve anomaly detection accuracy in complex industrial topologies, providing a framework for deploying reliable automated quality assurance systems.

Zitieren

APA

Arcano-Bea, P., Díaz-Labrador, A., Vidal-Bralo, A., Zayas-Gato, F., & Quintián, H. (2026). Unsupervised detection of overshooting defects using hybrid generative and boundary classification models. Journal of Logic and Computation, 36(8). https://doi.org/10.1093/logcom/exag054

Vancouver

Arcano-Bea P, Díaz-Labrador A, Vidal-Bralo A, Zayas-Gato F, Quintián H. Unsupervised detection of overshooting defects using hybrid generative and boundary classification models. Journal of Logic and Computation. 2026;36(8). doi:10.1093/logcom/exag054

BibTeX
@article{ArcanoBea2026,
  author = {Arcano-Bea, Paula and Díaz-Labrador, Anabel and Vidal-Bralo, Alejandro and Zayas-Gato, Francisco and Quintián, Héctor},
  title = {Unsupervised detection of overshooting defects using hybrid generative and boundary classification models},
  journal = {Journal of Logic and Computation},
  year = {2026},
  volume = {36},
  number = {8},
  doi = {10.1093/logcom/exag054}
}
Infokarte | EUVIMED

Artikeldaten derzeit nicht verfügbar.

Recherche