Leveraging Bridge Scour Detection by Multi-Model Machine Learning and Drive-By Data

This study provides a detailed computational investigation of railway bridge scour detection using multiple machine learning algorithms and drive-by sensor data. A simulation-based framework generates synthetic train crossings representing track profiles acquired by Rail Infrastructure Alignment Acquisition, a far-from-axle sensing tool installed on in-service trains. Artificial Neural Networks (ANNs), Random Forests (RFs), and Support Vector Machines (SVMs) are trained, tested, and compared for binary bridge-level classification of imposed support-stiffness-reduction conditions using features derived from the Discrete Wavelet Transform and ARX-derived impulse-response Markov parameters, under leakage-free blocked nested cross-validation. RF achieved the highest overall accuracy, while ANN achieved the highest balanced accuracy in the rear-mounted 7-crossing case. Batch-size effects were non-monotonic, with short windows improving some configurations but longer windows reducing healthy-class recognition. Sensor orientation affected classifier performance and feature sensitivity, with neither configuration consistently superior across classifiers and batch sizes. The framework is further examined through class-balancing, feature-importance, and feature-family ablation analyses. The results demonstrate numerical feasibility in the adopted simulation setting and may support future bridge-condition screening after field validation.