A dynamic early warning method for bridge damage via data clustering-aided subdomain feature extraction considering nonlinear environmental effects

Modal frequencies are fundamental damage-sensitive features for vibration-based structural health monitoring of bridges, but exhibit nonlinear and non-Gaussian variability under multi-source environmental conditions, which profoundly obscure damage-induced changes and reduce early warning accuracy. To address this limitation, this paper proposes a fully automated data-driven dynamic early warning method based on data clustering-aided subdomain feature extraction. First, variational mode decomposition separates seasonal environmental fluctuations patterns and filters high-frequency noise from raw frequency sequences. Second, Dirichlet process k-means clustering adaptively partitions the preprocessed frequency data into local subdomains, within which the environmental effects approximate linear correlation. Guided by piecewise linearization, slow feature analysis is then applied within each subdomain to extract damage-sensitive slow features, achieving dimensionality reduction and redundancy elimination. Finally, an environmentally robust warning indicator is constructed using the Mahalanobis-squared distance, combined with a dynamic baseline threshold determined via kernel density estimation. Validation using field monitoring data from the Z24 Bridge demonstrates that the method effectively suppresses nonlinear and non-Gaussian variability without requiring direct environmental measurements, achieving significant improvements in robustness, warning timeliness, and automation level for engineering applications.