Transformer-based contrastive denoising autoencoder for unsupervised detection of wind-induced vibration changes in cable-supported bridges

ABSTRACT Wind-induced excitations can significantly affect the structural behavior and serviceability of long-span cable-supported bridges, potentially leading to aeroelastic instabilities, traffic disruption, and operational safety concerns. Timely detection of variability in vibration responses is important for ensuring structural safety and response-informed operational assessment. This study proposes an unsupervised deep neural network, termed Transformer-based Contrastive Denoising Autoencoder (T-CDAE), for vibration-based change detection in cable-supported bridges using only measured acceleration responses. The core idea is to learn the baseline vibration behavior of multi-sensor bridge responses under normal operating conditions and to identify deviations from this learned baseline through reconstruction errors combined with contrastive feature alignment for extracting stable latent representations. The Transformer architecture captures spatial correlations among distributed sensors through a self-attention mechanism. A combined reconstruction-contrastive learning framework enhances separability between baseline and out-of-baseline response patterns and also robustness to measurement noise. The main contributions of this study include: (i) a unified artificial neural network for unsupervised vibration-response change detection, (ii) a Transformer-based architecture for capturing global sensor dependencies, and (iii) a contrastive denoising representation learning framework for enhancing feature separability and robustness against noisy data. Field-recorded acceleration responses of the Ting-Kau Bridge are used to evaluate the proposed method, with historically documented typhoon events serving as extreme-event cases for validating its effectiveness. Results demonstrate the reliability of T-CDAE in detecting out-of-baseline vibration-response changes in the monitored bridge subjected to tropical-cyclone-type windstorms.