Drive-By Early Damage Detection in Railway Bridges using Wavelets and Autoencoders

This paper presents an innovative AI-driven drive-by methodology for unsupervised damage detection on a Warren truss bridge.The methodology employs acceleration data collected from eight sensors mounted on a LAAGRSS-type freight wagon.Wavelet scattering coefficients derived from these acceleration signals serve as input features for the model.Autoencoders, trained on baseline condition data, are utilized to reconstruct these coefficients, with the absolute reconstruction error acting as a damage-sensitive feature.Environmental and operational variations are mitigated through normalization, excluding high-variability components.A three-level data fusion approach, based on the Mahalanobis distance, generates a highly sensitive damage indicator.This indicator accurately detects all simulated damage scenarios, including those in their early stages, without misclassification.The study demonstrates the efficacy of the proposed methodology also for distinguishing between different damage types.Future work will focus on experimental validation and enhancement of the methodology for assessing damage severity.