Application of neural networks to the prediction of the compressive capacity of corroded steel plates

Abstract

The application of artificial neural network approaches has been successful in solving complex civil engineering problems, such as damage detection and structural member capacity prediction. Within the context of the present study, corrosion has become the main factor limiting the safety and load-carrying capacity of aging steel bridge girders. Corrosion damage is often most severe near girder ends in simple-span bridges due to deck joint leakage and the pooling of water and de-icing salts. In addition to empirical methods, Finite Element (FE) analysis is typically used to evaluate the residual bearing capacity of corroded steel girders. However, it is prohibitively challenging and time-consuming to create an accurate FE model of a corroded girder due to the irregular nature of corrosion damage. Resultantly, corrosion damage is often reduced to uniform section loss, which leads to unreliable estimates of a girder’s residual bearing capacity. Researchers have proposed methodologies for modeling irregular corrosion damage, but these approaches require a high level of expertise. A comprehensive method is therefore required to efficiently estimate the residual bearing capacity of a corroded steel girder. This paper proposes the use of neural networks to predict the residual bearing capacity of corroded steel plate models as a first step in estimating the residual bearing capacity of an in-service girder. Neural networks are constructed and trained on a database built from FE analysis performed on steel plate models with realistic representations of corrosion damage. This study assesses the ability of neural networks to estimate the compressive capacity of corroded steel plates since plate girders are one of the most prevalent girder forms in steel bridges. Three types of neural networks are trained to predict the compressive capacity of corroded plate models, including a multilayer perceptron (MLP), a convolutional neural network (CNN), and a hybrid MLP-CNN model. The average mean absolute percentage errors (MAPE) for the three models are 20.65%, 11.46%, and 9.64%, respectively. The results of this study demonstrate the potential of using neural networks to predict the compressive capacity of corroded plates efficiently and accurately, which could facilitate proactive maintenance decision-making for aging bridges.