Drive-by Monitoring for Scour Damage Classification in Deep Foundations of a Railway Bridge Using Optimized Convolutional Neural Networks

Bridge scour remains a leading cause of structural failures worldwide, demanding efficient and automated damage detection tools to ensure infrastructure resilience. This study proposes a numerical framework for scour damage classification in deep foundations of railway bridges using drive-by acceleration data. A high-fidelity finite element model of the Canelas Railway Bridge is developed and calibrated with field modal parameters to simulate realistic train–bridge interactions under multiple scour depths. Environmental and operational variabilities—such as speed, mass, temperature-dependent stiffness, track irregularities, and measurement noise—are incorporated to reflect real-world conditions. The vertical acceleration responses are processed using a hybrid Convolutional Neural Network-Long Short-Term Memory model optimized via Bayesian optimization for each sensor configuration. The performance is statistically evaluated over multiple runs to assess variability and robustness. Results demonstrate high and consistent classification accuracy across all scenarios and sensor locations, confirming the capability of the proposed driveby deep learning approach as a reliable, non-intrusive, and scalable solution for automated scour detection in railway bridge foundations.