“What You Need is Data!” : Physics-Guided TimeGAN for Data Expansion in Bridge SHM

Traditional Structural Health Monitoring (SHM) technologies have not yet been widely adopted by infrastructure asset managers. However, with the recent advancements in information technology and artificial intelligence (AI), modern SHM has brought greater practical potential. AI-based SHM often relies on big data, but in practice, data availability is a problem, especially for some difficult-to-obtain cases. Theoretically, Generative Adversarial Networks (GANs) can augment data and significantly expand databases. On the other hand, some argue that GAN models can be challenging to train and that the amount of data required to train a GAN might be sufficient to train a diagnostic model. More importantly, the data generated by GANs may not truly capture the underlying physical characteristics. To address these, this paper proposes to use physical laws to guide GANs, with TimeGAN adopted as the base model due to its strong performance with time-series data. In this study, the proposed physics-guided TimeGAN (PyTiGAN) is used for data expansion in bridge SHM under the excitation of traffic events. The results demonstrate the effectiveness of the proposed method in expanding bridge SHM datasets from multiple dimensions.