PSI - Issue 84

Available online at www.sciencedirect.com

ScienceDirect

Procedia Structural Integrity 84 (2026) 481–488

© 2026 The Authors. Published by ELSEVIER B.V. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0) Peer-review under responsibility of the scientific committee of the Conference Abstract Structural Health Monitoring of bridge infrastructures faces major challenges due to the scarcity of labeled damage data and the strong heterogeneity of monitoring information across different structures and operational conditions. These limitations significantly reduce the effectiveness of conventional supervised learning techniques when applied to real-world bridge networks. In this context, transfer learning strategies have emerged as promising solutions to exploit information from data-rich source domains and transfer it to unlabeled target structures. In this context, this study presents a preliminary sensitivity analysis aimed at supporting the design of damage-tailored domain adversarial neural networks (DANNs) for transfer learning in bridge monitoring. The idea is to identify and rank the most informative input features to feed parallel DANNs, each one associated to a specific damage class. An archetypal simplified finite element model is adopted as the labeled source domain to simulate multiple damage scenarios, while a more detailed FE model of the same bridge acts as a proxy for the real, unlabeled target structure. A One-vs-Rest (OvR) sensitivity analysis is performed on a feature set comprising modal frequencies and nodal rotations to select scenario-dependent feature subsets. To effectively classify damage in the target domain, scenario dependent input features are selected to ensure sensitivity to specific damage mechanisms. The results show that different damage classes activate distinct feature patterns, highlighting the limitations of a global feature set for multi-scenario damage classification, while also providing the basis to design parallel DANNs, perform data fusion, learn domain-invariant representations tailored to a single damage class and ensure effective and interpretable transfer learning. III Fabre Conference: Existing Bridges, Viaducts, and Tunnels: Research, Innovation, and Applications A Preliminary Sensitivity Analysis Toward Damage-Tailored Domain-Adversarial Neural Networks for Transfer Learning in Bridge Monitoring Valentina Giglioni a, * , Francesco Mariani a , Ilaria Venanzi a , Filippo Ubertini a a Department of Civil and Environmental Engineering, University of Perugia. Via G. Duranti 93, Perugia - 06125, Italy

* Corresponding author. E-mail address: valentina.giglioni@unipg.it

2452-3216 © 2026 The Authors. Published by ELSEVIER B.V. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0) Peer-review under responsibility of the scientific committee of the Conference 10.1016/j.prostr.2026.06.062