Convolutional neural network for highway bridge indirect structural health monitoring

Abstract

This work addresses the challenge of detecting early-stage scour damage, a hidden, leading cause of roadway bridge failures, using indirect monitoring data. The main contribution is the systematic investigation of the Convolutional Neural Network (CNN) capability to detect low-intensity damage, representing the initial stages of foundation scour, under a comprehensive set of Environmental and Operational Variabilities (EOVs). The dataset is generated from a numerical finite element code (VBI-2D) simulating the vehicle-structure dynamic interaction. Damage is modeled as a reduction in foundation support stiffness at the midspan. The EOVs incorporated include variations in vehicle properties, speed fluctuations, and stochastic road surface irregularities (ISO Class 'A'). Additionally, the study evaluates the damage-classification performance for different sensor locations along the vehicle and quantifies the uncertainties arising from both data variability and the stochastic nature of the learning algorithm. The optimal CNN architecture accuracy demonstrates the feasibility of using supervised CNNs with indirect monitoring data for early scour detection in highway bridges.