PSI - Issue 84

Mirko Calò et al. / Procedia Structural Integrity 84 (2026) 392–400

395

representing different model realizations of the same taxonomy branch. The LM1 is the demand model for traffic load analysis and the associated intensity measure, IM, is the amplification factor, α , applied to both uniformly distributed and concentrated loads of LM1. Push-down displacement-control analysis are carried out by scaling the LM1 model loads through α . The analysis is concluded when either the prescribed traffic load level is reached without incurring failure, or the ultimate flexural or shear capacity of the structure is attained. The ultimate flexural capacity is assumed to be reached when either the top concrete fiber or bottom steel fiber attains its corresponding ultimate strain. Instead, the ultimate shear capacity is defined by the capacity model adopted in the Italian building code (MIT, 2018). The outcome of the analysis, for both failure mechanisms, is classified as either “no collapse” or “collapse” at the corresponding α . In the final stage, the XGBoost ML algorithm is trained through a supervised classification problem. The goal is to predict, starting from input variables (i.e., superstructure geometric and mechanical characteristics, and intensity level of the traffic load), if the superstructure exceeds (“collapse”) or not (“no collapse”) the LS (output variable). As with any ML algorithm, accurate hyperparameter tuning is required, in this case implemented through a grid search approach with f-fold cross-validation strategy. The performance of the XGBoost algorithm is evaluated considering the confusion matrix and related metrics (i.e., Accuracy, Precision, Recall, and F1-score). Furthermore, the use of eXplainability approaches as Shapley Additive exPlanation SHAP (Lundberg and Lee, 2017), is proposed to investigate the impact of each input feature on the prediction of the model.

(model features and IM) ( ailure No failure)

Definition of

Modelling of

branch

variables variables parameters

Grid search with cross validation

and E model generation

Maximum span length ( ) Superstructure defectiveness level ( )

Fig. 1. Framework.

Table 1. Proposed taxonomy for multi-span simply supported PC girder bridges (the identification codes assigned to the parameters are indicated in brackets). Category Classification Associated parameter Maximum span length ( L_MAX ) Short span ( L ) Maximum span length ( L MAX ) Medium span ( M ) Long span ( H ) Superstructure defectiveness level ( SDL ) Low ( L ) Equivalent percentage reduction of the corroded steel area (ξ) Medium-Low ( ML ) Medium ( M ) Medium-High ( MH ) High ( H )