Dynamic monitoring of bridges subjected to freezing-induced variability by kernel-enhanced deep anomaly detection

Bridges are prominent structural systems in transportation networks, and their dynamic monitoring provides valuable insights into structural health and operational performance. However, variations in bridge dynamic features such as modal frequencies caused by freezing temperatures can lead to false alarms and misdiagnoses. To address the negative consequences of these errors, this paper introduces a dual-phase unsupervised learning method based on Deep Support Vector Data Description (DSVDD) and Polynomial Kernel Distance (PKD) for bridge dynamic monitoring under severe freezing-induced effects. In this method, DSVDD is used to learn compact representations of normal structural behavior, while PKD is employed as a distance-based anomaly scoring mechanism for decision-making. Accordingly, DSVDD utilizes a deep neural network to map normal modal frequencies into a compact feature space, forming a hypersphere that tightly encloses the normal data. Subsequently, PKD measures the distance between unknown test data and the hypersphere center to assign anomaly scores. These scores are then compared with a decision threshold to detect abnormal structural states. The contributions of this study focus on developing a new unsupervised learning approach based on statistical and deep learning paradigms and integrating PKD as a kernel-based anomaly scoring mechanism within the DSVDD framework. Freezing-affected modal frequencies of real-world bridge structures are utilized to validate the effectiveness and practicability of the proposed method along with several comparative studies. Results demonstrate that DSVDD-PKD can effectively handle the complexities associated with bridge dynamic behavior under freezing conditions and mitigate the risks of false alarms and misdiagnoses.