Railway Bridge Scour Detection Using Drive-By Track Profiles and Machine Learning Algorithms

This research introduces a practical, non-intrusive method for detecting scour in railway bridges using drive-by monitoring with in-service trains. The approach is first validated on a case study railway bridge in the UK, where longitudinal track profiles are collected by the Rail Infrastructure Alignment Acquisition (RILA) unit mounted on a 6 -axle locomotive. Measurements belong to the healthy stage of the bridge; scoured stage measurements are synthetically generated using a Vehicle-Bridge Interaction model. For both scoured and healthy stages of the bridge, a Cross-Entropy algorithm is then employed to optimize pier stiffnesses by minimizing the discrepancies between measured and theoretical track profiles. The theoretical track profiles combine bridge displacements from a Finite Element model with Moving Reference Influence Lines and rail irregularities. Differences in optimized pier stiffness between the healthy and damaged states are interpreted as quantitative indicators of scour. Building on this methodology, the study develops a Machine Learning framework that generalizes the scourdetection concept to a broader set of bridges and operating conditions. Synthetic RILA-measured track profiles are generated for multiple bridge and scour scenarios and used to train classifiers based on Artificial Neural Networks, Random Forests, and Support Vector Machines. These models are tasked with identifying scour and, in extended configurations, localizing the scoured supports using features derived from simulated track profiles. The proposed framework shows that indirect, train-based measurements can provide valuable early indicators of scour, supporting risk-informed management of railway bridge networks.

Keywords: Bridge scour, drive-by monitoring, RILA, Moving Reference Influence Lines, Cross-Entropy optimization, Vehicle-Bridge Interaction, Discrete Wavelet Transformation (DWT), Auto-Regressive Models with Exogenous Output (ARX), Markov Parameters, Machine Learning, Artificial Neural Networks, Random Forest, Support Vector Machines.