Unsupervised structural health monitoring using physics-based vibration features derived from an instrumented bridge mockup

Abstract

Structural health monitoring (SHM) of civil infrastructure is often limited by the scarcity of labeled damage data required for supervised machine learning approaches. This study presents an unsupervised SHM framework that combines physics-based vibration features with novelty detection algorithms for damage identification using only healthy-state measurements. The proposed methodology uses condition-level modal and frequency-response quantities to support physical characterization of structural changes, while a separate baseline-compatible trial-level vibration representation provides the input to unsupervised novelty detection. The framework is experimentally validated using an instrumented two-span steel bridge mockup subjected to fifteen independent damage scenarios representing a range of structural modifications and support-condition changes. More than 5000 impact measurements are used to evaluate baseline repeatability, modal behavior, damage-induced vibration changes, and unsupervised novelty detection performance. The results demonstrate that the trial-level vibration representation effectively distinguishes the undamaged baseline condition from the investigated damage scenarios using principal component analysis (PCA)-based squared prediction error and Mahalanobis distance novelty metrics. Sensitivity analyses further demonstrate the robustness of the proposed framework with respect to the retained PCA variance, while statistical evaluation confirms a positive relationship between novelty magnitude and engineering damage severity. In addition, complementary vibration-response descriptor analyses quantify the relative contributions of different sources of dynamic information, and an output-only implementation illustrates that the proposed baseline-trained novelty detection framework can be reformulated using response-only vibration measurements when excitation information is unavailable. The proposed framework provides an experimentally validated and physically interpretable approach for baseline-trained vibration-based SHM and establishes a foundation for extending unsupervised damage detection toward response-only structural monitoring applications.