Speaker
Description
Environmental and operational variability, particularly freezing weather, can significantly influence the dynamic characteristics of bridge structures, complicating the identification of true structural anomalies. In this study, we propose an innovative anomaly detection framework, termed DSVDD-PKD, which integrates Deep Support Vector Data Description (DSVDD) with Polynomial Kernel Distance (PKD) to address this challenge. The method is unsupervised and data-driven, leveraging the power of deep neural networks to transform modal frequency data of bridges under normal conditions into a latent feature space. A hypersphere is constructed in this space to tightly enclose the normal data, where PKD is then employed to compute anomaly scores by measuring the generalized kernel distance be-tween test samples and the hypersphere center. Anomalies are detected by comparing these scores to a statistical threshold. The framework is validated on long-term monitoring datasets from two real bridges: Z24 (Switzerland) and KW51 (Belgium), both exposed to freezing conditions. Results reveal that DSVDD-PKD accurately detects structural changes—including damage and retrofit—while effectively mitigating the adverse effects of freezing temperatures. Comparative studies with existing popular approaches demonstrate the superior accuracy and temperature-robustness of the proposed method, achieving 100% detection accuracy with zero false positives and false negatives in both applications. This work contributes a significant advancement to the field of structural health monitoring by offering a scalable, automated solution capable of isolating true structural anomalies from environmental-induced variations.
