PSI - Issue 78

Lorenzo Principi et al. / Procedia Structural Integrity 78 (2026) 1681–1688

1686

Table 7. Topology and hyperparameters of the optimal ANN model selected for CER through GS-5fFCV.

Topology

Optimization Algorithms Hyperparameters

Input Layer Neurons Output Layer Neurons Hidden Layers Number Hidden Layers Neurons

27

Complexity-Penalty Max iterations number

0.40

5 1

69

Batch size (*)

- -

27

Initial learning rate (*)

Activation function

Hyperbolic Tangent

Optimization algorithm

L-BFGS

(*) Only for ADAM and SGD algorithms

3.3.2. Performance Analysis In Step 2 of Phase III, the optimal ANN is tested for performance, as shown in Figure 3 and Table 8. The ANN achieves an F1 - score of 0.71, showing good results across all classes. Table 8. Classification Report for CER. Class Precision Recall F1-score Samples L - - - 0 ML 0.86 0.86 0.86 7 M 0.69 0.77 0.73 26 MH 0.72 0.75 0.73 44 H 0.68 0.56 0.61 27 Metric Training Samples Test Samples F1-Score 0.79 417 0.71 104 Figure 3. Confusion Matrix for CER. 4. Case Study The ANN developed in Chapter 3 is used to predict the CER along a road axis crossing a seismically active region with non - uniform hazard distribution. The dataset comprises 95 bridges of the Italian State Highway 77 (SS77), which is 95km long and connects the Apennine Mountain region to the Adriatic coast (Figure 4).

(a) (b) Figure 4. Location of the SS77 in Italy (in blue) (a), and its position within the regional road network of the Marche region (in red and black, respectively) (b). Bridge lengths range from 7 to 543 m, with 75% under 200 m. Maximum spans range from 5.5 to 77 m (average 29.5 m), and most bridges have 1 - 5 spans (average 4). Traffic volumes (ADT/ADTT) increase from inland to coast (Figure 5a–b). P.R.C. decks (bridges built between 1945 and 1980 with Bridge Design Code B) are most common,