PSI - Issue 84

Valentina Giglioni et al. / Procedia Structural Integrity 84 (2026) 481–488 485 measurements are considered as damage-sensitive features, including six natural frequencies (whose reference values are reported in Table 2) and three rotational components under dead loads, namely 1 , 2 , and 3 , extracted at the nodes shown in Fig. 2a. The measurement locations are selected at one-quarter of the span length ( /4 ) for each span, as these positions are considered sufficient to capture meaningful variations in the structural response associated with damage, while limiting the number of monitoring points. The quantities 1 , 2 , and 3 represent rotations about the -, -, and -axes, respectively.

Fig. 2. (a) selected nodes to measure rotations; (b) general view of the bridge model in the SAP environment

Table 2. Calibration results.

Modes

Natural frequency [Hz] SAP model

Natural frequency [Hz] ABAQUS model

Estimation error [%]

MAC

Mode 1 Mode 2 Mode 3 Mode 4 Mode 5 Mode 6

2.314

2.303 3.121 3.161 3.294 3.813 4.036

0.48 0.99 1.23

0.92 0.76 0.80 0.98 0.89 0.92

3.09

3.122 3.439 3.591 3.897

4.4

5.82 3.44

3. Overview of the proposed framework To enable damage identification in the target structure without requiring labeled target data, a DANN framework is adopted. Conventional domain adaptation approaches typically employ a single multi-class DANN, trained to simultaneously transfer knowledge and classify multiple damage scenarios using a high-dimensional feature set. However, in practical SHM applications, this strategy presents two major limitations. First, damage classes that are absent or poorly represented in the source domain are difficult to transfer reliably to the target domain, leading to