EUVIMEDEUVIMED entdecken
Crossref

Physics-Informed CycleGAN Framework Combining GNN and Transformer for Domain Adaptation

Shang-Jun CHEN, Chuan-Chuan Hou, Stefano Mariani

e-Journal of Nondestructive Testing · 2026-08 · 31 (8)

DOI: 10.58286/33665 ↗

Kurz gefasst

Fachartikel in der Zeitschrift e-Journal of Nondestructive Testing, 2026-08. Verfasst von Shang-Jun CHEN und 2 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

In this study, a physics-informed Cycle-Consistent Generative Adversarial Network (CycleGAN) framework is proposed for vibration-based structural health monitoring of structures subjected to lateral impacts. The proposed method aims to translate vibration data between the healthy and damaged state domains. Within the CycleGAN, a Graph Neural Network block is employed to model the spatial topology of the sensor network: a weighted graph is constructed according to the physical distances between sensors using a Gaussian radial basis function. This enables the network to capture correlations and structural response propagation characteristics among the multi-channel sensor data. A Transformer block is also incorporated to model long-time sequence data, enhancing the ability to capture vibration decay patterns without altering the adversarial–cyclic training structure. During training, in addition to the adversarial, identity, and cycle-consistency losses, a multi-resolution Short-Time Fourier Transform spectral loss and an adaptive frequency-band regularization loss are integrated within the domain adaptation framework. These physics-based loss functions enforce coherence in both the time and frequency domains between generated and real structural responses. Results obtained for concrete-filled steel tubular (CFST) structures, based on seven-channel acceleration recordings sampled at 25 kHz, demonstrate that the proposed framework effectively enhances cross-domain stability in healthy–damage signal translation and suppresses abnormal frequency peaks.

Zitieren

APA

CHEN, S., Hou, C., & Mariani, S. (2026). Physics-Informed CycleGAN Framework Combining GNN and Transformer for Domain Adaptation. e-Journal of Nondestructive Testing, 31(8). https://doi.org/10.58286/33665

Vancouver

CHEN S, Hou C, Mariani S. Physics-Informed CycleGAN Framework Combining GNN and Transformer for Domain Adaptation. e-Journal of Nondestructive Testing. 2026;31(8). doi:10.58286/33665

BibTeX
@article{CHEN2026,
  author = {CHEN, Shang-Jun and Hou, Chuan-Chuan and Mariani, Stefano},
  title = {Physics-Informed CycleGAN Framework Combining GNN and Transformer for Domain Adaptation},
  journal = {e-Journal of Nondestructive Testing},
  year = {2026},
  volume = {31},
  number = {8},
  doi = {10.58286/33665}
}
Kostenlos starten →