Nowadays, traditional model-driven SHM-based structural system identification (St-Id) techniques are often effective for assessing the health status of structures by solving inverse problems related to damage detection. However, in high-dimensional problems, St-Id methods can be computationally intensive and may suffer from ill-posedness, leading to challenges in interpreting the results. To enhance the robustness of data interpretation, this study investigates using a Deep Neural Network (DNN) that is pre-trained with data derived from a calibrated Finite Element (FE) model to streamline the model class selection process. In this context, DNNS can learn complex mappings between inputs and outputs, since layers in DNNs can learn compact representations of high-dimensional data, can effectively reduce dimensionality while preserving essential information, can be updated continuously as new data becomes available, can be designed to learn multiple model classes and select the most appropriate one based on input data. The approach effectiveness is tested through a numerical model of a representative bridge span, with a parametric analysis conducted under simulated damage scenarios. The analysis focuses on selecting a model class guided by a comprehensive evaluation of the DNN's predictive performance across two main damage scenarios, stiffness reduction and tendon prestress losses in different locations, that consistently exhibits the highest accuracy and reliability in identifying damage. Diverse data sources, including modal displacements, frequencies, and static rotations, are used to build comprehensive training and test datasets. Hence, the study underscores the potential of multi-feature approaches and the importance of data fusion for achieving precise damage identification.
Keywords: - Model class selection, Deep Neural networks, Prestressed concrete bridge, Data fusion, System identification.