PSI - Issue 78

Ivan Roselli et al. / Procedia Structural Integrity 78 (2026) 128–136

131

The test setup is illustrated in Fig. 1b. The instrumentation utilized to acquire the vibration data comprised a 3D optical motion capture system based on Vicon technology (De Canio et al. (2013)). Vibration data were recorded at 200 fps by 13 near – infrared (NIR) cameras equipped with 5 Mpixel sensors tracking the displacements of 68 markers placed at the most relevant positions of the RCF specimen. The accuracy of the recorded markers displacements is in the order of 0.03 mm. The markers are visible like small white balls (diameter of 25 mm) in Fig. 1b. More detailed information on this shaking table campaign are available in Cataldo et al. (2023). 3. Damage assessment The structural damage can be quantified in several ways. A widely accepted definition of the damage states (DSs) was provided by Taghavi et al. (2003), according to several literature studies (Verderame et al. (2019)), and taking into account the damage observed in post-even surveys (Masi et al. (2014)). However, the use of DS grades is a quite imprecise and qualitative classification method, albeit very effective under many points of view. Many damage indices have also been developed by several authors. Most of them take into account energy dissipation parameters, which are quite difficult to be obtained from experimental measurements and are generally theoretically calculated from numerical models (ref). In the context of Structural Health Monitoring (SHM) most methods are based on the dynamic identification achieved by the analysis of ambient vibration data recorded in several positions of the structure (Pandey et al. (1994)). In particular, they often focus on the detection of anomalies in terms of decay of the modal frequencies (Ren et al. (2002)), which imply a loss of stiffness in the structure. This is potentially due to the loosening of connections, the development of cracks and deterioration phenomena in the material. Thus, once that the effect of changes in the environmental conditions (e.g. cyclic variations of the air temperature etc.) can be neglected (Azzara et al. (2018)), which is substantially true in the controlled environment of a shaking table laboratory, the decay of structural stiffness is generally a widely accepted indicator of the state of damage of the structure. In the present study the following formulation of the Damage Index (DI) was used (Mendes et al. (2014)): = 1 − ( 0 ) 2 (1) where f is the modal frequency of the damaged structure and f 0 is the modal frequency in undamaged conditions. As usual, the first modal frequency was considered, as the first mode of vibration is the most significant for the global behavior of the structure so it is representative of the overall state of damage of the entire building. 4. Artificial Intelligence procedure Fig. 2 summarizes the workflow of the implemented AI procedure. The simulations were performed using Python, with various packages such as Numpy, Pandas, Sklearn, Keras, and Matlab 2015b. Before the WNDC vibration data are given in input to the implemented AI procedure based on CVAE model, a pre-processing is performed for data preparation,. The pre-processing steps include (Palumbo et al. (2022)): signals selection, data alignment, time segmentation, train-test random partition and normalization. Signal selection is performed to discard bad quality signals that might invalidate the following processing steps. In particular, the data are processed to build the necessary features by selecting a certain number of measurement points based on their standard deviation. The ones with a standard deviation between 0.2 and 0.3 were chosen. With these limits 17 measurement points were selected. Then data are aligned in order to have the same number of samples for each experiment. As a general rule, the time samples length must be a compromise between the longest possible sequence of input data, in order to have the best probability of catching the real behavior in the pattern extracted from data (Römgens et al. (2024)), and the highest number of samples obtainable from each time history to train the CVAE.