ABSTRACT
Traditional single-model structural system identification (St-Id) techniques are widely employed to assess the health of structures by addressing inverse problems related to damage detection. However, in high-dimensional scenarios, these methods often become computationally demanding and prone to ill-posedness, complicating result interpretation. To address these challenges and improve the robustness of data interpretation, this study explores the use of Multi-Layer Perceptron Artificial Neural Networks (MLP ANNs) pre-trained on datasets derived from Finite Element (FE) model simulations, framing the damage assessment task as a model class selection problem. In this approach, ANNs are designed to learn complex input-output mappings, with multi-layer architectures enabling compact representations of high-dimensional data while preserving essential information. The network k operates as a classifier, identifying the most appropriate model class among the considered classes based on input data. The effectiveness of this methodology is evaluated using a numerical model of a representative bridge span, with parametric analyses conducted under simulated damage scenarios. The study focuses on two primary damage scenarios—stiffness reduction and tendon prestress losses at various locations. The analysis emphasizes selecting the optimal model class by comprehensively evaluating the predictive performance of the ANN, which demonstrates high accuracy and reliability in identifying damage. To construct robust training, validation, and testing datasets, diverse data sources such as modal displacements, frequencies, and static rotations are utilized. The accuracy of the ANN-based approach is then compared to the Bayesian Information Criterion (BIC), highlighting the advantages of multi-feature approaches and the significance of data fusion in achieving precise and reliable damage identification.
