PSI - Issue 78

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

1683

3. Artificial Neural Network for Seismic Risk The framework shown in Figure 1 was applied to develop an ANN model for predicting the seismic risk (CER) of bridges. All analyses were performed in Python (v3.11.2150.0, 64 - bit, released 08/02/2023, Rossum et al., 2010) using the scikit - learn library (v1.2.0, released 12/2022, Pedregosa et al., 2010).

3.1. Phase I: Data Selection and Data Preprocessing

3.1.1. Data Selection Based on Principi et al. (2025), a set of input features was selected to predict the CER. Two additional variables are introduced: Peak Ground Acceleration (PGA) with a 475-year return period (rock and stiff soil conditions) and a binary indicator of Seismic Design compliance. The complete list of input features is shown in Table 1.

Table 1. Input features and descriptions (V = Vulnerability, E = Exposure, H = Hazard according to IG criteria).

Input Features CER Features Description Max span length V Indicates the length of the longest span Number of spans V Indicates the number of spans of the structure Static scheme V Indicates the static scheme of the bridge Deck material V Indicates the deck material typology Design code class V Class of design code (1) Bridge age class V

Range in which the date of construction of the bridge falls into (≤1945, 1945 - 1980, ≥ 1980) (1) V Whether designed according to post-2003 Italian standards (O.P.C.M. No. 3274 – March 20, 2003)

Seismic Design Obstacle type Alternative routes

E Indicates the category of obstacles the bridge allows to cross E Presence of other roads to service bridge traffic E Indicates the Average Daily volume of Traffic (ADT) E Indicates the Average Daily Trucks volume of Traffic (ADTT)

ADT

ADTT

H Maximum allowable mass on the bridge (60 tons, ≤ 44 tons, ≤ 26 tons, ≤ 8 tons, ≤ 3,5 tons) (1) H Peak Ground Acceleration (PGA) (a g <0.05g, 0.05g ≤ a g <0.10g, 0.10g ≤ a g < 0.15g, 0.15g ≤ a g <0.25g, a g ≥ 0.25 g) (1)

Load Limit

PGA

Total length

-

Indicates the length of the structure

(1) Ranges and classes as standardized in MIT (2020). The dataset used to train the algorithm includes 521 highway bridges located across several Italian regions. Table 2 reports key statistics for quantitative features, while Table 3 summarizes the categorical data.

Table 2. Statistical report for the quantitative input features.

Total Length [m]

Max Span Length [m]

Number of Spans

ADT

ADTT Integer

Format

Float64 129.34 189.53

Float64

Integer

Integer 11109 12516

Mean Value

25.47 15.67

5 6 1

656 771

Standard Deviation Minimum Value Maximum Value

1.00

0.90

0

0

1943.00

167.00

56

70199

5547