PSI - Issue 84

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

292

Fig. 3. Experimental and calibrated modes of the (a) source bridge and (b) target bridge.

3.2. Training of the FNN on the source domain and transfer learning To develop the SM for predicting the dynamic modal properties of the source bridge, a synthetic training dataset was generated using the calibrated high-fidelity FEM. The bridge deck was discretized into 20 control regions, uniformly distributed along the five spans, with each span divided into four segments of equal length, as depicted in Fig. 2. Each region was associated with a stiffness multiplier , =1,…,20 , scaling the nominal Young’s modulus of the deck. The stiffness multipliers were independently sampled within the interval [0.80,1.05] : the lower bound corresponds to a maximum 20% global stiffness reduction, while the upper bound exceeds the undamaged condition by 5% to account for benign environmental-induced fluctuations during the continuous damage assessment. An efficient Latin Hypercube Sampling (LHS) strategy was adopted to explore the parameter space, generating a source