EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

VibroAgent: Multiresolution Soft-Token Fusion with Gemma 4 for Reference-Conditioned Structural Vibration Monitoring

Niks Kordjukovs, Alfredo Cigada, Danilo Pietro Pau

Computers · 22.10.2026 · 15 (11) · 742

DOI: 10.3390/computers15110742 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Computers, online erschienen am 22.10.2026. Verfasst von Niks Kordjukovs und 2 weiteren Autor:innen. Die Arbeit nennt 57 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

VibroAgent represented ten-second acceleration windows from one reference and five monitored locations as 84 continuous soft tokens for Gemma 4 E2B, letting one model decide and potentially explain; only its decisions were evaluated. Tasks were global healthy/changed classification and affected-zone flags for released or loosened connections on a lattice mast (LUMO), grandstand (QUGS), and scaled offshore jacket (UPC). Against tuned comparators on the paper’s 984-window split, published Gemma had the highest overall global accuracy (0.9787) but pooled affected-target F1 0.1802. Its target predictions tracked slot position. An Abdeljaber-style CNN adaptation had higher pooled target F1 (0.3107) but LUMO global accuracy below 0.47. In an even three-seed comparison (equal data, pretraining, tuning, selection) on 1483 held-out windows, mean LUMO/QUGS target F1 was 0.290/0.407 (binary Gemma) and 0.623/0.813 (CNN); finer supervision gave no benefit. With more data and healthy/changed-balanced sampling for all methods, binary Gemma matched the CNN in healthy/changed detection on 792 never-used damaged and 72 healthy UPC windows (balanced accuracy 0.949 versus 0.935 under equal loss weighting; paired difference 0.014 ± 0.030); LUMO/QUGS target F1 did not improve. Results established neither localization advantage, unseen-structure generalization, nor field validity. Six STWIN.box nodes on a Qualcomm IQ-9075 EVK demonstrated local integration only.

Zitieren

APA

Kordjukovs, N., Cigada, A., & Pau, D. P. (2026). VibroAgent: Multiresolution Soft-Token Fusion with Gemma 4 for Reference-Conditioned Structural Vibration Monitoring. Computers, 15(11), 742. https://doi.org/10.3390/computers15110742

Vancouver

Kordjukovs N, Cigada A, Pau DP. VibroAgent: Multiresolution Soft-Token Fusion with Gemma 4 for Reference-Conditioned Structural Vibration Monitoring. Computers. 2026;15(11):742. doi:10.3390/computers15110742

BibTeX
@article{Kordjukovs2026,
  author = {Kordjukovs, Niks and Cigada, Alfredo and Pau, Danilo Pietro},
  title = {VibroAgent: Multiresolution Soft-Token Fusion with Gemma 4 for Reference-Conditioned Structural Vibration Monitoring},
  journal = {Computers},
  year = {2026},
  volume = {15},
  number = {11},
  pages = {742},
  doi = {10.3390/computers15110742}
}
Kostenlos starten →