Bridges are vital infrastructure elements, and ensuring their long-term safety requires accurate and timely damage detection. While machine learning (ML) has shown promise for structural monitoring, many existing approaches rely on complex models, dense sensor networks, or laboratory-scale validations, limiting their transparency and practical deployment. This study develops a sequential ML framework with interpretable outputs, where each stage provides a clear, physically meaningful checkpoint for damage detection, severity assessment, and localization. Two complementary strategies are employed: (i) frequency-based analysis, where modal frequency shifts are processed using k-nearest neighbor (kNN) classifiers to detect, quantify, and locate damage, and (ii) acceleration-based analysis, where transient responses from only three strategically placed sensors are analyzed with gated recurrent units (GRUs) for initial detection, Fourier-transformed features for severity estimation, and wavelet-based spectrograms combined with a lightweight convolutional neural network (CNN) for localization. The framework leverages a hybrid dataset, combining FEM-generated scenarios validated against field measurements, to enhance coverage of potential damage states while maintaining real-world relevance. Applied to the full-scale KW51 bridge, the approach demonstrates a stage-wise interpretable decision process, showing strong potential for reliable damage identification with minimal computational overhead and sparse instrumentation. • Sequential ML framework for digital damage identification. • Interpretable stage-wise prediction of damage detection, severity, and location. • Sparse instrumentation with only three accelerometers. • Combines FEM-generated scenarios and field data for digital damage identification.
Qiu et al. (Sun,) studied this question.