PSI - Issue 78

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

924

Table 1. Parameters used for CoA-S assessment, with associated risk components and corresponding numerical codings (* categories are defined according to Italian technical standard of construction (NTC-2018))

N° Parameter description

Risk Component

Parameter Option

1 Expected peak ground acceleration, return period Tr = 475 y (a_g)

Hazard

[1: a_g ≥ 0,25 g; 2: 0,15 g ≤ a_g < 0,25 g; 3: 0,10 g ≤ a_g < 0,15 g; 4: 0,05 g ≤ a_g; 5: a_g < 0,05 g]

2 Topographical category* 3 Subsoil category* 4 Max span length (L3)

Hazard Hazard

[1: T1-T2-T3; 2: T4] [1: A-B; 2: C-D-E]

Vulnerability Vulnerability Vulnerability Vulnerability Vulnerability Vulnerability Vulnerability

[1: L3 ≤ 20 m; 2: L3 > 20 m] [1: Isostatic; 2: Hyperstatic]

5 Static scheme 6 Deck material 7 Number of spans

[1: RC; 2: PRC; 3: Steel; 7: Masonry]

[1: single; 2: multiple]

8 Level of degradation (LoD) 9 Seismic vulnerability factors

[1: High; 2: Medium-high; 3: Medium; 4: Medium-low; 5: Low]

[1: available; 2: not available] [1: seismic; 2: not seismic]

10 Design according to seismic criteria 11 Average daily Traffic (TGM)

Exposure Exposure Exposure Exposure Exposure Exposure

[1: TGM ≥ 25000; 2: 10000 < TGM < 25000; 3: TGM ≤ 10000] [1: Lm > 50 m; 2: 20 m < Lm ≤ 50 m; 3: Lm ≤ 20 m]

12 Average span length (Lm)

13 Road alternatives 14 Bypassed entity

[1: available; 2: not available] [1: High; 2: Medium; 3: Low]

15 Frequency of pedestrian traffic

[1: yes; 2: no]

16 Strategic function

[1: strategic; 2: not strategic]

17 Seismic Attention Class (CoA-S)

Overall

[1: High; 2: Medium-high; 3: Medium; 4: Medium-low; 5: Low]

2.2. Pre-training In order to reduce the computational burden of the final training phase and identify the most effective configuration, a preliminary analysis was conducted using learning curves . The goal was to determine the optimal dataset size and the best-performing machine learning algorithms among those included in the AutoGluon (Tabular) framework, excluding models that — despite comparable or lower performance — were computationally expensive in terms of processing time, such as ANNs. Three model configurations were tested, each combining a different sampling strategy (stratified without SMOTE, stratified with SMOTE, and balanced) for the prediction of the CoA-S. All available features were used except for the Level of Degradation (LoD, parameter 8 in Table 1). SMOTE (Synthetic Minority Over-sampling Technique was applied only to stratified datasets, with the aim of rebalancing the underrepresented classes ("Low" and "Medium Low") by generating synthetic samples (Chawla et al., 2002) in a controlled manner to avoid distortion and improve generalization. All models were trained using 80% of the data for training and 20% for testing, leveraging the AutoGluon-Tabular framework in the medium_quality_faster_train configuration. This setup ensures a balanced trade-off between computational efficiency and predictive accuracy. AutoGluon training strategy integrates repeated − bagging with multi-layer stack ensembling. In repeated k-fold bagging, the dataset is split multiple times into folds. For each iteration, models are trained on −1 folds and validated on the remaining one — known as the Out-Of-Fold (OOF) set. Repeating the partitioning process increases model robustness and reduces overfitting by exposing models to diverse data splits. The OOF predictions generated during each iteration are averaged and then used as additional input features in the subsequent stacking layers. This enriched representation enables the ensemble to leverage complementary strengths from different models, enhancing overall performance and stability. The learning curves shown in Fig. 2 illustrate the evolution of the main metrics — Precision (1), Accuracy (2), Recall (3), and F1-score (4) — during both the training phase (solid lines) and the test (validation) phase (dashed lines). The main performance metrics used to evaluate the models are defined as follows: