PSI - Issue 44

Valentina Giglioni et al. / Procedia Structural Integrity 44 (2023) 1948–1955 Valentina Giglioni et al./ Structural Integrity Procedia 00 (2022) 000–000

1952

5

4.2. FE model calibration A simplified model of the Z24 bridge, shown in Fig. 2 a), was built in SAP software and afterwards calibrated based on structural dynamics information stemming from experimental measurements. Specifically, the deck cross-section is modelled with 20 beam finite elements and pinned restraints at the connections deck-abutments. Moreover, as suggested by Figueiredo et al. (2018), the thickness of the girder plate has been slightly increased at the end of the main span. The supporting piers are modelled with 30 beam elements and connected to the soil with fixed restraints. Additional to the dead load, a linear vertical mass and 20 concentrated translational masses are considered to justify traffic intensity, the presence of barriers, guardrails and sidewalks or any external load induced on the bridge. Based on the continuous monitoring data, dynamic system identification is carried out by applying the Stochastic Subspace Identification (SSI) technique within MOSS (García-Macías and Ubertini (2020)), a software implemented by the SHM group of the Department of Civil and Environmental Engineering of the University of Perugia, leading to the results reported in Table 1. To calibrate the computational model, the concrete’s elastic modulus, the mass values and the thickness of the girder plate are adopted as calibrating parameters in order to minimize the difference between the four experimental and estimated natural frequencies. As inferred from Table 1, the calibrated FE model provide accurate estimates, yielding errors below 1% in most cases.

Table 1. Comparison between experimental natural frequencies and those estimated by the calibrated FE model Mode Experimental [Hz] FE model [Hz] Error [%] 1 3.85 3.82 0.78 2 4.9 4.91 0.20 3 10.36 10.26 0.97 4 12.46 12.91 3.61

While the first and the fourth are symmetric bending modes, the second is a lateral mode and the third one is a non symmetric bending mode. 4.3. Data generation The calibrated FE model is deployed to generate a sufficient number of accelerations data to train and test the autoencoder network. Firstly, 10 minutes-long random forces simulating white noise with a sampling frequency of 50 Hz are simultaneously assigned to different nodes of the deck along the vertical and transversal direction. To prove the model’s accuracy in reproducing reliable bridge responses during operational and environmental conditions, the virtual and experimental measurements extracted by three sensors located along the center line of the bridge are compared in terms of RMS acceleration. Table 2 shows similar values with the same order of magnitude between the real and the FE model.

Table 2. RMS acceleration obtained from the FE model and real measurements Sensors FE model response Real measurements S6 0.011 0.017 S14 0.0059 0.0032 S16 0.0020 0.0032

For the purpose of testing the ML algorithm for post-earthquake assessment, a realistic earthquake-induced scenario is simulated through the FE model by reducing by 48% and 20%, respectively, the elastic modulus of the concrete at the bottom of the piers and in proximity of the connections with the deck, as depicted in Fig. 2 c). Damaged accelerations are recorded by using four sensors: S1-T and S2-L measure the transversal and longitudinal accelerations