PSI - Issue 84

Elisa Tomassini et al. / Procedia Structural Integrity 84 (2026) 288–295

289

with notable cases of collapse highlighting the consequences of inadequate monitoring and lack of maintenance (American Society of Civil Engineers ASCE, 2025, Cuerva Navas et al. , 2021, Tan et al. , 2020, Gkoumas et al. , 2019). These challenges have driven the widespread adoption of Structural Health Monitoring (SHM) systems, particularly vibration-based approaches, which enable continuous, non-intrusive assessment of in-service structures. However, while large-scale SHM deployments are increasingly feasible, translating raw monitoring data into reliable, network level damage information remains an open research problem (Andreose et al. , 2024, Tomassini et al. , 2024, Magalhães et al. 2008). Within SHM, damage assessment traditionally encompasses detection, localization, and quantification. Data-driven techniques, especially unsupervised learning methods, have proven effective for damage detection under limited prior knowledge, but they are generally insufficient for localization and quantification (Rai et al. , 2025, Tomassini et al. , 2025). Supervised approaches and model-based system identification can address these higher-level tasks but are hindered by the scarcity of labeled damage data and high computational costs. To overcome data limitations, Population-Based SHM (PBSHM) has recently emerged as a promising paradigm, leveraging shared information across populations of similar structures (Bull et al. , 2021, Gosliga et al. ,2021, Gardner et al. , 2021, Jian et al .). Transfer Learning (TL) techniques have therefore been introduced to adapt models trained on well-instrumented structures to underrepresented targets, improving generalization and scalability across heterogeneous bridge networks. This paper proposes a model-driven framework for network-scale damage identification based on transferable surrogate models (SMs) constituted by a feedforward neural network (FNN). FNN-based SMs replace high-fidelity finite element models (FEMs), enabling efficient damage identification (López-Cuervo et al. , 2025). Leveraging their intrinsic compatibility with TL, models trained on a source bridge can be fine-tuned for similar target structures using limited additional data. The proposed approach integrates PBSHM principles into a continuous, regression-based damage identification strategy that does not rely on predefined damage classes. The methodology is validated on two real-world bridges, demonstrating the effective transferability of surrogate models between similar structures and providing a scalable solution for bridge-network SHM. The paper is organized as follows. Section 2 presents the proposed methodology and the architecture of the FNN. Section 3 applies the approach to two real-world bridges and discusses the results. Finally, Section 4 summarizes the In this study, knowledge transfer is performed from a SM developed for a source structure to a target structure for which no SM is initially available. The objective is to exploit the knowledge embedded in the source SM to efficiently construct a new SM for the target structure. This strategy is particularly advantageous because generating a sufficiently rich training dataset for an SM can be computationally demanding, especially for complex structures exhibiting nonlinear behavior, interface effects, or requiring highly refined finite element discretization. In such cases, both direct FEM updating and the generation of dedicated training datasets become prohibitively expensive. Starting from a pre trained SM of a structurally similar source system significantly reduces the amount of data required for the target, thereby lowering computational costs and development time. The FNN architecture, illustrated in Fig. 2, adopts a modular configuration designed to simultaneously predict natural frequencies and mode shapes while accounting for their distinct characteristics. Structural damage is modelled as variations in the elastic moduli , =1,…, , over N predefined regions of the structure, and is represented through stiffness multipliers . The network begins with an input layer followed by several fully connected layers with hyperbolic tangent activation functions, which capture the nonlinear relationship between damage parameters and modal features. The network input consists of a vector of non-dimensional damage-sensitive parameters ∈ℝ , whose components are defined as: = 1 1 2 √ , =1,…, , (1) where 1 is the fundamental frequency of the FEM, the average span length, the average mass density, and and the average cross-sectional area and moment of inertia of the deck. The architecture then separates into different branches: one branch predicts the natural frequencies, while the other ones consist of independent subnetworks, each main conclusions. 2. Methodology