Data-Driven Bridge Scour Monitoring: A Taxonomy and Drive-By Machine Learning Case Study

Bridge scour threatens bridge safety and serviceability, yet the growing application of Machine Learning (ML) has not produced a unified understanding of how sensing sources, learning tasks and validation practices influence monitoring capability. This article addresses this gap through a taxonomy of 53 ML-based bridge-scour studies and a simulation-based railway case study examining the relatively underexplored use of displacement-related drive-by measurements for scour detection and support-level localisation. The dataset comprises 2924 simulated crossings involving four bridge models and three train types, with scour represented by reductions in vertical support stiffness. Full-bridge and support-centred Discrete Wavelet Transform and Autoregressive with exogenous inputs-derived Markov features are used to train Artificial Neural Network, Random Forest (RF) and Support Vector Machine classifiers under nested cross-validation. Batch inputs incorporate information from consecutive crossings. Among the highest-performing detection configurations, rear-mounted RF using 31-crossing batches achieved 0.990 accuracy and F1 scores of 0.994 and 0.977 for the scoured and healthy classes. The best localisation configuration, front-mounted RF using seven-crossing batches, achieved 0.968 accuracy, with both F1 scores approximately 0.97. These simulation-based results do not demonstrate transfer to unseen bridges or field-confirmed scoured conditions but indicate potential for bridge screening and targeted inspection.