Damage detection of a cable-stayed footbridge using multiple damage modelling and neural networks

Abstract

This paper presents a routine for damage detection based on multiple damage modelling and artificial neural networks. The routine aims to identify the existence, location, and the most damaged elements. The first step is to define "critical" regions based on stresses and deformations of the updated numerical model of the structure, as well as the probable order of occurrence in the case of multiple damage by simulating different scenarios. The second step is to train a neural network using the natural frequencies of the intact and damaged numerical models as input data and vectors that indicate the position and magnitude of the damage as expected output. The feed-forward backpropagation neural network was adopted. The approach was used to evaluate a cable-stayed footbridge. A new dynamic test was performed and a vector composed of the identified frequencies was input into the network, which indicated the existence of possible damage. The provided damage scenario was applied to the updated model, along with the prestressing forces of the stays obtained experimentally. The maximum difference between the frequencies of the updated and damaged numerical model with the damage vector provided by the network and the frequencies experimentally identified in the new test was 1.80%.