FE-ANN based modeling of 3D simple reinforced concrete girders for objective structural health evaluation

Abstract

The structural deterioration of aging infrastructure systems is becoming an increasingly important issue worldwide. To compound the issue, economic strains limit the resources available for repair or replacement of such systems. Over the past several decades, structural health monitoring (SHM) has proved to be a cost-effective method for detection and evaluation of damage in structures. Visual inspection and condition rating is one of the most commonly applied SHM techniques, but the effectiveness of this method suffers due to its reliance on the availability and experience of qualified personnel performing largely qualitative damage evaluations. The artificial neural network (ANN) approach presented in this study attempts to augment visual inspection methods by developing a crack-induced damage quantification model for reinforced concrete bridge girders that requires only the results of limited field measurements to operate. Simply-supported three-dimensional reinforced concrete T-beams with varying geometric, material, and cracking properties were modeled using Abaqus finite element (FE) analysis software. Up to five cracks were considered in each beam, and the ratios of stiffness between cracked and healthy beams with the same geometric and material parameters were measured at nine equidistant nodes along the beam. Two feedforward ANNs utilizing backpropagation learning algorithms were then trained on the FE model database with beam properties serving as inputs for both neural networks. The outputs for the first network consisted of the nodal stiffness ratios, and the sole output for the second ANN was a health index parameter, computed by normalizing the area under the stiffness ratio profile over the span length of the beam. The ANNs achieved excellent prediction accuracies with coefficients of determination (R²) exceeding 0.99 for both networks. Additional FE models were created to further assess the networks’ prediction capabilities on data not utilized in the training process. The ANNs displayed good prediction accuracies (R² > 0.8) even when predicting damage levels in beams with geometric, material, and cracking parameters dissimilar from those found in the training database. A touch-enabled user interface was developed to allow the ANN models to be utilized for on-site damage evaluations. The results of this study indicate that application of ANNs with FE modeling shows great promise in SHM for damage evaluation.