Isolation distributional kernels for indirect bridge health monitoring: A full-scale investigation of anomaly detection

Indirect or drive-by structural health monitoring (SHM) is emerging as a viable alternative to conventional sensor-based bridge monitoring by enabling condition assessment using vehicle-mounted measurements. Despite significant progress, the practical implementation of indirect SHM remains challenging due to sensitivity to noise, operational variability, road roughness effects, and the reliance on large volumes of labeled data or baseline measurements. This study proposes an indirect SHM framework based on Isolation Distributional Kernels (IDK) to address these limitations. Unlike point-wise feature-based approaches, IDK operates at the distributional level, enabling robust anomaly detection from short-duration, noisy vehicle response signals with minimal data requirements and linear computational complexity. This work represents the first field-scale application of IDK to indirect bridge SH, bridging recent advances in distribution-based anomaly detection with real-world vehicle–bridge interaction data. The proposed framework departs from conventional deep-learning-driven approaches by eliminating the need for extensive training or large labeled datasets while maintaining high detection accuracy. %The proposed framework is applied to vehicle–bridge interaction data to detect changes in bridge condition without requiring prior damage labels or extensive training. Its performance is evaluated through comparative studies against established data-driven methods, demonstrating superior classification accuracy. Crucially, the framework is validated using field measurements collected from two full-scale bridges - one located in Nara, Japan, and one located in New South Wales, Australia - highlighting its robustness under environmental variability and measurement noise. The results confirm that IDK-based indirect SHM provides a scalable, data-efficient, and physically interpretable solution for bridge condition assessment compared to competing methodologies, offering strong potential for network-level monitoring and practical deployment in real-world bridge management operations.