EUVIMEDEUVIMED entdecken
Crossref

Self-Supervised Anomaly Detection in Extremely Imbalanced Ultrasonic Railway Data: An Adaptive Triplet-Center Approach

Huiying Li, Timo Hartmann

e-Journal of Nondestructive Testing · 2026-07 · 31 (7)

DOI: 10.58286/33076 ↗

Kurz gefasst

Fachartikel in der Zeitschrift e-Journal of Nondestructive Testing, 2026-07. Verfasst von Huiying Li und 1 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Ultrasonic B-scan inspection of railway tracks generates vast grayscale data where structural anomalies are extremely rare. This extreme class imbalance challenges automated detection, as supervised methods lack labels. While contrastive learning enables training without annotations, traditional sampling often fails to distinguish domain noise from true anomalies in low-contrast environments. We propose a robust framework driven by an Adaptive Triplet-Center Loss. By introducing a structure-focused sampling strategy and dynamic margin adjustment, our model effectively suppresses acoustic artifacts while maintaining sensitivity to structural deviations. Validated on a 10km real-world dataset, our approach significantly improves anomaly prioritization, demonstrating robustness in handling extremely imbalanced industrial data.

Zitieren

APA

Li, H., & Hartmann, T. (2026). Self-Supervised Anomaly Detection in Extremely Imbalanced Ultrasonic Railway Data: An Adaptive Triplet-Center Approach. e-Journal of Nondestructive Testing, 31(7). https://doi.org/10.58286/33076

Vancouver

Li H, Hartmann T. Self-Supervised Anomaly Detection in Extremely Imbalanced Ultrasonic Railway Data: An Adaptive Triplet-Center Approach. e-Journal of Nondestructive Testing. 2026;31(7). doi:10.58286/33076

BibTeX
@article{Li2026,
  author = {Li, Huiying and Hartmann, Timo},
  title = {Self-Supervised Anomaly Detection in Extremely Imbalanced Ultrasonic Railway Data: An Adaptive Triplet-Center Approach},
  journal = {e-Journal of Nondestructive Testing},
  year = {2026},
  volume = {31},
  number = {7},
  doi = {10.58286/33076}
}
Kostenlos starten →