PSI - Issue 62

Lorenzo Principi et al. / Procedia Structural Integrity 62 (2024) 89–96 Principi L./ Structural Integrity Procedia 00 (2019) 000 – 000

95

7

the Medium-Low and Medium Class, for which the F1-Score is 0.55 and 0.61, respectively. In summary, the results in predicting the LoD show a significant dependency on the considered class. The model needs improvement to achieve a more uniform performance across all classes, aiming for at least 0.60 on average. In contrast, the predictions for the Structural and Foundational Attention Class are generally satisfactory, except for the Medium-Low and Medium Classes. As for the LoD, improvements are required to attain a consistent performance across all classes, but the overall outcome is more favorable in this case.

a

b

Fig. 2. (a) Classification Matrix for LoD; (b) Classification Matrix for Structural and Foundational Class.

Table 6. Classification Report

Classification Report for LoD Class

Classification Report for Structural and Foundational Class

Precision Recall F1-score Samples

Precision Recall F1-score Samples

Low

0.33 0.68 0.72 0.31 0.58 0.51

0.17 0.75 0.62 0.56 0.50

0.22 0.71 0.67 0.55 0.38

12 28 21 16

Low

1.00 0.60 0.65 0.79 0.77 0.77

1.00 0.50 0.58 0.83 0.94

1.00 0.55 0.61 0.83 0.86

5 6

Medium-Low

Medium-Low

Medium

Medium

19 29 16 75 75

Medium-High 0.53

Medium-High 0.83

High

8

High

Accuracy F1-Score

85 85

Accuracy F1-Score

6. Conclusions In recent decades, the global concern regarding the condition of existing bridges has heightened due to an increasing number of failures. In response to this issue, Italy has implemented the New Italian Guidelines, which introduce an innovative multi-level and multi-risk approach to comprehensively assess the safety of existing bridges. The assessment is divided into six levels, this paper specifically concentrates on the first three levels. This work adopted an ANN to predict the Level of Defectiveness and the Structural and Foundational Attention Class. Data from 423 bridges were collected using Level 0 Reports, followed by preprocessing the information based on the nature of each feature. A subset of effective predictors was then selected to create the Input Features set. The model was developed and optimized identifying the best combination of hyperparameters. The performance of the model was evaluated on a case study. Results are promising, considering the minimal data used. The Level of Defectiveness was adequately predicted, although the performance was highly dependent on the specific class. On the other hand, the Structural and Foundational Attention Class was effectively assessed, achieving good result. However, improvements are necessary, particularly regarding the Level of Defectiveness. In the future, this can be achieved by using more samples and exploring different architectures and sets of Input Features. This tool will help in rationalizing inspections and provide information on the structural and foundational risk of bridges, supporting territorial-level management and planning.