Steel bridge structural damage detection using Ground-Based Radar Interferometry vibration measurements and deep learning Convolutional Neural Networks

Abstract

[EN] This paper introduces a new, data-driven, vibration-based, damage detection strategy realized on an on-purpose built, Bailey type, steel bridge model (6.12 m x 1.80 m, scale 1:2.5) as part of the research work undertaken in the School of Rural, Surveying and Geoinformatics Engineering, NTUA, Greece. Vibrations of the bridge model in a “healthy” and damaged condition were recorded using a Ground-Based Radar Interferometer (GBRI). Structural damage was deliberately induced on the bridge model by removing a number of carefully selected structural parts, whilst bridge excitation was achieved using a vibration generation apparatus. This system employs an in-house built in trolley system capable of realizing preset dynamic load scenarios. The damage detection approach developed relies on the transformation of GBRI vibration measurements to Continuous Wavelet Transform (CWT) scalogram images. The latter are then used to apply alternate pattern recognition techniques; particularly, a class of pre-trained Deep Learning Convolutional Neural Networks (CNNs) through the application of Transfer Learning technique. The classification results of the bridge health status reach an accuracy of the order of 90%, suggesting the effectiveness and the high potential of the proposed approach.