PSI - Issue 44

Xuguang Wang et al. / Procedia Structural Integrity 44 (2023) 1736–1743 X. Wang et al. / Structural Integrity Procedia 00 (2022) 000–000

1742

7

(21)

where 1 = 0.05716 , 2 = 1.6620 , 3 = 0.9727 , 4 = 1.0788 , 5 = 1.2587 , and 6 = − 0.1883 . The targets and estimations from the training set and test set are plotted in Fig. 3. The 2 scores are all higher than 0.8, which indicate the established equations are acceptable.

a. Mean drift ratio in the DLS ( ) b. CoV of drift ratio in the DLS ( )

c. Mean drift ratio in the ULS ( ) d. CoV of drift ratio in the ULS ( ) Fig. 3. Scatter plots of the regression models for predicting the mean drift ratios and their .

3.3. Seismic fragility curve generation With the classification and regression models generated via genetic programming, the main parameters for generating the fragility curve can be quickly estimated for a given set of input parameters. For instance, a set of input parameters is randomly selected from the generated dataset: = 2 m , ℎ = 4 m , = 0.4 m , = 400 MPa , = 35 MPa , = 0.75% , = 0.04% , and = 2% . Two sets of and 2 , as listed in Table 3, are calculated by substituting and from the generated dataset (analytical model) and the established equations (Eqs. (17), (20), and (21)) into Eq. (1). The fragility curves generated by fitting lognormal distribution with and 2 and the one directly generated from the Monte Carlo simulations are compared in Fig. 4. The 3 curves are well matched, showing that the regression model provides good accuracy with a much lower computational cost. Table 3. Calculated and 2 based on and from different sources From the generated dataset From the equations 2 2 0.2642 0.2615 -1.3642 0.06614 0.2558 0.2609 -1.3963 0.06586