Bridge damage detection using unsupervised learning with vehicle-induced vibrations

This article proposes a data-driven method for bridge damage detection utilizing a Transformer architecture with a patching mechanism to extract damage-sensitive features from vehicle-induced vibrations (VIVs). The core concept is the implementation of the advanced bridge weigh-in-motion (BWIM) technique to enable automated data collection and processing of VIV signals from continuous monitoring data. First, the BWIM data (including force magnitude, velocity, and wheelbase) is used to selectively sample VIV events, creating the sampled VIV (SVIV) dataset, which is often induced by specific vehicle types. Subsequently, a patching-based time series Transformer model, called PTAD and trained on the SVIV dataset from intact bridge structure, is specifically designed for the unsupervised structural anomaly detection task. Finally, two damage indices (DIs) corresponding to the long-term and short-term patterns are defined based on the reconstruction error of VIV samples by the trained PTAD model. These DIs can enable continuous monitoring and real-time detection of the health status of the target bridge. Both numerical examples and laboratory experiments on a scaled steel bridge demonstrate that the proposed method can accurately and efficiently detect the presence and extent of damage in bridge structures, even with sparse sensor deployment.