PSI - Issue 78

Pooria Mesbahi et al. / Procedia Structural Integrity 78 (2026) 1839–1846

1840

structural capacity. However, obtaining accurate estimates of these quantities remains a major challenge, especially for non-instrumented structures. Recent advances have highlighted the potential of integrating probabilistic methods, structural reliability tools, and machine learning techniques within a Bayesian framework to improve the assessment of infrastructural systems under uncertainty. Previous research has demonstrated the potential of Bayesian networks (BNs) for modeling and updating structural vulnerability and degradation models, including seismic fragility Wang et al. (2020), underwater scour Maroni et al. (2021), and post-earthquake damage Yue et al. (2010). BNs o ff er a flexible and interpretable means of fusing heterogeneous sources of information, including sensor data, numerical simulations, and expert judgment, to perform reliability assessments and probabilistic inference Straub and Der Kiureghian (2012); Yue et al. (2010). Building on this foundation, recent studies have explored the integration of model updating techniques into BNs for network level vulnerability assessment Wang et al. (2020), and have proposed methods for extrapolating seismic demand to unmonitored locations Yue et al. (2010). In this context, the present work introduces a novel transfer-learning methodology based on a BN architecture, aiming to estimate the seismic capacity of reinforced concrete bridges across a network. By leveraging data from monitored bridges and transferring learned patterns to similar but unmonitored ones, the proposed approach enhances capacity estimation capabilities where direct measurements are lacking. This method contributes to the growing body of research that employs probabilistic and data-driven techniques for structural assessment and aims to facilitate informed post-earthquake decision-making. After an earthquake, prioritizing rehabilitation of damaged infrastructure is critical, with highway bridges being among the most strategic assets. Nevertheless, equipping all assets with sensors is impractical. Therefore, structural vulnerability is commonly assessed via fragility curves, which estimate the probability of exceeding specific damage states. In this context, it looks reasonable to leverage the information of proximate and typologically similar monitored structures to enhance the vulnerability assessment of non-monitored assets to this end, building on our previous work on seismic demand transfer learning Mesbahi et al. (2024), this paper focuses on transfer learning for seismic capacity. The proposed framework, illustrated in Fig.1, integrates nonlinear finite element model updating (FEMU) of monitored assets, using accelerometer data to update probabilistic distributions of key model parameters. For non monitored bridges, these updated distributions are inferred via a Bayesian Network capturing dependencies between monitored and non-monitored assets. Posterior parameter distributions are then used to compute pushover curves, providing seismic capacity estimates with reduced uncertainty bounds. 2. Proposed methodology and framework

Monitored

Sensor data (acceleration measurements)

Posterior distribution of the model parameters for B1

FEM

FEMU via DE-MC

Accelerogram of the earthquake

B1

Transfer learning via BN

Updated pushover curve

Updated pushover curve

B2

Posterior distribution of the model parameters for B2

FEM

Non-monitored

Fig. 1: General proposed framework for post-earthquake seimic capacity updating