PSI - Issue 84

Valentina Giglioni et al. / Procedia Structural Integrity 84 (2026) 481–488

482

Keywords: Structural Health Monitoring; Bridge damage detection; Damage sensitivity; Machine Learning

1. Introduction The continuous monitoring of bridge infrastructures has become a critical priority worldwide, driven by the aging of existing assets, increasing traffic demands, and the growing exposure to extreme environmental events. In this context, Structural Health Monitoring (SHM) has emerged as a key tool to support condition assessment, early damage detection, and informed decision-making throughout the life cycle of civil structures (Figueiredo et al. (2022)). In particular, the accurate modelling and monitoring of post-tensioned bridge systems play a crucial role in ensuring a reliable understanding of structural behaviour and in enabling effective damage identification and long-term performance assessment (Sconocchia et al. 2024). Over the last two decades, SHM has progressively evolved toward data-driven approaches, largely enabled by advances in sensing technologies and computational capabilities (Svendsen et al. (2023), Sclafani et al. (2025)). Recently, Artificial Intelligence (AI) and Machine Learning (ML) methods have demonstrated remarkable potential in extracting meaningful patterns from large volumes of monitoring data, improving damage detection, localization, and classification accuracy (e.g., Giglioni et al. (2022), Sun et al. (2023), Cianci et al. (2026)). Despite these advances, the practical deployment of AI-based SHM systems at the scale of large bridge networks remains challenging. One of the main limitations lies in the scarcity or complete absence of labeled data for real damaged conditions, since damage events are rare and often unobserved. Moreover, monitoring data collected from different bridges or under varying environmental and operational conditions typically exhibit significant distributional discrepancies, which severely degrade the performance of models trained on a single structure or dataset. In order to address the issues described above, several Transfer Learning (TL) approaches have been proposed in recent years to enable knowledge transfer between different structures, operational conditions, or numerical and experimental domains (e.g., Pan et al. (2023), Giglioni et al. (2025)). TL is particularly attractive for SHM applications, as it allows models trained on data-rich source domains to be adapted to data-scarce target structures. However, in most real-world SHM scenarios, target labels are unavailable, making supervised TL approaches impractical. A possible TL approach relies on the use of Domain-Adversarial Neural Networks (DANNs), which adopt an adversarial training strategy in which a feature extractor is simultaneously optimized to minimize classification loss on the source domain and to confuse a domain discriminator that attempts to distinguish between source and target data (Wang et al. (2023), Jin et al. (2025)). As a result, the learned feature representations become domain-invariant, enabling effective damage classification on the target structure despite the absence of labels. Beyond the advantage of enabling health-state classification on the target bridge without requiring labeled target data, DANN-based architectures allow the simultaneous exploitation of several heterogeneous features. Nevertheless, a major challenge remains: real bridges may experience multiple damage scenarios, including conditions that are not explicitly represented in the source training dataset. In such cases, a single global classifier may fail to correctly identify or even recognize unseen damage patterns (Singhal et al. (2023)). To address this limitation, the presented framework aims to develop parallel DANN architectures, each specifically tailored to transfer a given damage scenario. In fact, scenario-dependent feature selection plays a crucial role, as different damage mechanisms may be sensitive to different subsets of structural features. Consequently, it becomes essential to analyze the sensitivity of features extracted from the source domain, to improve the effectiveness and robustness of the damage classifiers embedded within the DANN framework. Preliminary analyses are therefore presented in this paper to identify the most informative features for each damage scenario, to enhance classification performance and improve the interpretability and transferability of the SHM system. As an application case study, the source domain is defined by a simplified numerical model developed in SAP 2000 environment (Computers and Structures, 2018), representing a four-span continuous bridge composed of four double T girders, a concrete slab, and three piers with pier caps. A total of 17 distinct structural health scenarios, including both undamaged and damaged configurations, are simulated to construct a comprehensive training dataset for the TL procedure. For each scenario, an extensive set of features, such as nodal rotations measured at multiple locations and