Bridge health monitoring through series analysis: Full-scale validation on a Truss Bridge

In this study, bridge health monitoring is approached as an unsupervised semantic segmentation aimed at identifying structural damage by detecting anomalies in time series data. The interaction between passing vehicles and the bridge itself is evaluated using a developed Vehicle-Bridge Interaction (VBI) model. Subsequently, the Matrix Profile is employed to identify potential changes in the acceleration signals obtained from the bridge, indicative of structural damage onset. The proposed methodology undergoes rigorous testing and validation through numerical simulations conducted on the Old ADA Bridge in Japan. Encouraged by promising results from these simulations, the approach is further applied to field test data collected from the same bridge. The significance of this work unfolds through three key contributions: firstly, its successful application to real-world field test data, a privilege shared by only a few methodologies validated through full-scale field tests; secondly, its demonstrated capability to detect damage using a limited dataset; and finally, its elimination of the need for prior supervised learning, enhancing its practical applicability and efficiency. • A defect identification approach based on semantic segmentation is proposed. • Its successful application to real-world field test data is presented. • Its capability to detect damage using a limited dataset is demonstrated.