Physics-Informed CycleGAN Framework Combining GNN and Transformer for Domain Adaptation
Shang-Jun Chen, Chuan-Chuan Hou, S. Mariani
TL;DR: Results obtained for concrete-filled steel tubular 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.
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.