Semi-supervised pointwise VIV detection via few-shot and sequential transfer learning

Long-span bridges are particularly vulnerable to vortex-induced vibrations (VIVs) due to their low damping ratios and flexible structural characteristics, making reliable detection a critical task in structural health monitoring (SHM). Traditional threshold-based detection methods often fail to capture complex dynamic behaviors and fluctuating operational conditions. While deep learning models have shown potential, the manual annotation of large-scale SHM datasets remains labor-intensive, costly, and inconsistent. To address this challenge, a semi-supervised learning framework is proposed for pointwise VIV detection, eliminating the need for extensive manual labeling by leveraging pseudo-labeling and sequential transfer learning. The framework adopts a few-shot learning paradigm, in which an initial teacher model is trained using synthetic VIV data and a small set of real non-VIV (NV) samples. This teacher model pseudo-labels large-scale SHM datasets and is iteratively refined through sequential transfer learning. Two regularization methods, a high softmax threshold and label smoothing, were employed during training to enhance robustness against unseen NV patterns and varying environmental conditions. The proposed method is validated using real-world SHM data from the Jindo Bridge, demonstrating stable performance under diverse environmental conditions, including aerodynamic buffeting. By continuously improving without human intervention, the framework achieves classification performance on par with supervised models, offering a scalable and cost-effective solution for near-real-time VIV monitoring in large-scale civil infrastructure systems.