PSI - Issue 78

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

1845

Fig. 6: Posterior distributions of model parameters before and after BN updating for Bridge2

For simplicity, a Gaussian distribution is fitted to the posterior, although samples could also be drawn directly from the discretized or more accurately fitted distributions. Random samples are then generated from these distributions, and POA is performed. With the probabiluty distributions of all parameters before and after updating, POA is conducted using 1,000 samples for both B1 and B2. As expected, due to the greater pier height of B2, the lateral shear capacity (peak shear force) is lower compared to B1 curve. In both bridges, the uncertainty bounds narrow after updating, improving confidence in decision-making. This methodology allows the seismic capacity changes of individual assets to be clearly observed and enables comparative risk assessment across multiple assets. In the present case, the seismic capacity is significantly reduced for B1, while it remains almost unchanged for B2 (based on mean value).

1000 1500 2000 2500 3000 3500 4000 4500 5000 Shear base [kN]

1000 1500 2000 2500 3000 3500 4000 4500 5000 Shear base [kN]

Pre-FEMU Mean Pre-FEMU Lower Bound Pre-FEMU Upper Bound

Pre-BN Mean Pre-BN Lower Bound Pre-BN Upper Bound

Pre-FEMU Range Post-FEMU Mean

Pre-BN Range Post-BN Mean

Post-FEMU Mean - 1SD Post-FEMU Mean + 1SD Post-FEMU ±1SD

Post-BN Mean - 1SD Post-BN Mean + 1SD Post-BN ±1SD

0 500

0 500

0

50

100 150 200 250 300 350 Displacement [mm]

0

50

100 150 200 250 300 350 Displacement [mm]

(a) Bridge1 (Bayesian FEMU is used for updating model parameters)

(b) Bridge2 (BN is used for updating model parameters)

Fig. 7: POA results before and after updating model parameters

4. Conclusions

This paper proposes a probabilistic framework that integrates Bayesian Networks (BN) with Bayesian finite ele ment model updating (FEMU) to enable transfer learning for post-earthquake assessment of bridge networks, explic itly accounting for structural nonlinearities. The results demonstrate that the proposed approach e ff ectively updates uncertain model parameters, initially defined with non-informative uniform priors, toward their true values through probabilistic FEMU. The updated knowledge is then transferred to non-monitored bridges using BN, which captures the probabilistic dependencies among random variables. Compared with deterministic approaches, this methodology provides a more robust treatment of uncertainty and achieves higher confidence in model validation. Furthermore, the use of DE-MC enhances the exploration of the parameter space relative to classical MCMC, improving convergence and reliability.