Transformer-based time-series GAN for data augmentation in bridge digital twins

Recent advancements in AI-based Digital Twins (DTs) have substantially influenced bridge monitoring and maintenance, especially through Deep Learning (DL) for sensor-based damage detection. However, the effectiveness of DL models is constrained by the extensive training data they require, which is often costly and timeconsuming to collect in bridge infrastructure contexts. To address this data scarcity, this paper proposes a data augmentation strategy employing a transformer-based time-series Wasserstein generative adversarial network with gradient penalty (TTS-WGAN-GP) to generate synthetic acceleration data. The synthetic data's fidelity is validated through similarity metrics and frequency domain analysis, showing close alignment with real acceleration signals for damage detection. Results demonstrate that this method achieves high-quality synthetic data with superior computational efficiency compared to existing approaches, improving dataset balancing and potentially enhancing the performance of data-driven models in DTs. This approach reduces dependence on extensive data collection, supporting reliable bridge health monitoring applications.