A substantial portion of structural health monitoring applications is recently focusing on bridges, given the growing need for real-time assessment of complex and aging infrastructures. One of the most challenging aspects is to ensure an effective management of a large bridge network and to organize properly scheduled monitoring activities. Conventional machine learning models need sufficient labeled data, especially related to damage conditions, and require separate training for each bridge, which is often impractical for multiple structures. To mitigate these issues, Transfer Learning allows one to leverage health-state information across a network of bridges by transferring knowledge from a source labeled domain to an unknown target domain. In this paper, a transfer learning-based methodology is proposed via the implementation of domain adversarial neural networks to build a domain-invariant space where the classifier can successfully generalize. Of specific interest to this paper is the case where an archetypal finite element model is used as the source domain to improve the ability to identify damage on real monitored bridges (target domains) belonging to the same structural typology. To this aim, a simplified model, selected as the source, is used to generate labeled data from multiple health-state classes (archetypal finite element model), while a most accurate model is employed as the target domain, so as a proxy of the “real” bridge. Natural frequencies are extracted from specific simulated damage scenarios, differing in typologies and localization. A discussion on the obtained results, focusing on the advantages and limitations of the proposed procedure, underlines the possibility of performing model-class selection by gaining knowledge from a numerical model via adversarial transfer learning.