PSI - Issue 78

Franco Ciminelli et al. / Procedia Structural Integrity 78 (2026) 921–928

925

Precision = TP (True Positive) TP (True Positive)+FP (False Positive) Accuracy= TP + TTNP ++ TFPN +(TFrNue(FNaelgsaetNiveeg)ative) Recall = TP T + P FN F1-score =2 ∙ Precision ∙ Recall Precision + Recall

(1)

(2)

(3)

(4)

These metrics were computed on both training and validation sets to evaluate generalization performance and identify signs of overfitting or underfitting. The selection of the best model was based on three main criteria: • high performance on validation metrics; • low gap between training and validation performance; • robustness across larger datasets. This analysis enabled the comparison of the tested configurations and the identification of the most robust one. In particular, the stratified dataset without SMOTE, with a size of 25'000 bridges, was selected for final training, as it offered the best trade-off between accuracy, generalization capability, and computational efficiency, and thus was adopted as the most effective configuration for both methodology and dataset size. The final model was trained using the previously selected 25'000-bridge dataset and the best-performing algorithm identified during the pre-training phase. This time, AutoGluon-Tabular was run in its most advanced mode, best_quality , which activates an extended training pipeline with optimized hyperparameter tuning and full exploitation of ensemble strategies. The model was evaluated on the 20% held-out test set using the same metrics defined in Section 2.2 (Accuracy, Precision, Recall, and F1-score). The trained model was saved so it can be used for future analyses or applications. 2.4. Experimental validation The experimental validation was conducted on a sample of 100 bridges, with data sourced from the ANSFISA database. Seismic characterization was carried out in accordance with the parameters defined in Table 1, which includes 17 descriptors categorized under the risk components: hazard, vulnerability, and exposure. With regard to hazard parameters (1–3), 29% of the bridges are located in areas with peak ground acceleration (a_ g) between 0.15 g and 0.25 g, 28% between 0.10 g and 0.15 g, 20% between 0.05 g and 0.10 g, and another 20% in areas with ag ≥ 0.25 g, while only 3% are situated in areas with ag < 0.05 g (parameter 1). All bridges fall under topographic category T1–T2–T3 (100%) (parameter 2). Regarding subsoil category, 52% are classified as C–D–E and 48% as A–B (parameter 3). Concerning vulnerability parameters (4–10), 78% of the bridges have a maximum span length (L3) greater than 20 m, while 22% have L3 ≤ 20 m (parameter 4). Static schemes are isostatic in 90% of cases and hyperstatic in 10% (parameter 5). In terms of structural material, 49% are made of prestressed reinforced concrete, 29% of ordinary reinforced concrete, 18% of steel, and 4% of masonry (parameter 6). A total of 77% of the structures have multiple spans, while 23% are single-span bridges (parameter 7). Seismic vulnerability factors are present in 36% of cases and absent in 64% (parameter 9). Regarding seismic design, 89% of the bridges were not designed according to seismic criteria, whereas 11% were (parameter 10). As for the level of seismic degradation (parameter 8), 36% of the bridges are classified as low, 35% as medium-low, 15% as medium, 5% as medium-high, and 9% as high. With respect to exposure parameters (11–16), 76% of the bridges have an average daily traffic (ADT) ≤ 10'000 vehicles, 21% between 10'000 and 25'000, and 3% ≥ 25'000 (parameter 11). The average span length (Lm) is ≤ 20 m 2.3. Training and model evaluation