Bridge Damage Identification Using Artificial Neural Networks

Abstract

An objective, data-driven approach to evaluate the performance of bridges for developing a structural health monitoring system is introduced as bridge behavior. A method of identifying structural damage through the evaluation of response data from an instrumented bridge is proposed. Strains during operational traffic events at the Powder Mill Bridge in Barre, Massachusetts, are recorded at many locations on the bridge. Bridge behavior is defined as each sensor location’s range of expected peak strain during a traffic event based on all other sensor locations’ strains measured at that instance in time. Artificial neural networks (ANNs) are trained with operational bridge response data in a bootstrapping scheme to generate a probabilistic model of bridge behavior. When tested against new data, the ANN-learned model of predicted bridge behavior is proven effective and applicable to varying traffic events with unknown loading conditions. A method for long-term performance assessment using the expected bridge behavior is proposed. Structural damage can impact bridge behavior and thus bridge performance. The effects of structural damage are extracted from simulated HS2 design truck runs on a calibrated finite-element model (FEM) and are applied to operational strain data to assess the damage identification method. When assessed, the damage identification method is effective at detecting the presence of damage, with no Type I or Type II errors when using a Wilcoxon rank-sum test of an appropriate significance level. Damage is effectively localized for most types of simulated damage.