PSI - Issue 78

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

129

1. Introduction Shaking table testing is widely considered the most representative and accountable technique of laboratory testing for experimental studies on the seismic response of structures. In particular, 6-DOF (Degree-Of-Freedom) shaking table facilities are able to reproduce the actual ground motion of an earthquake, as the seismic input time histories can be provided in all three components recorded at a given seismic station. In addition, large facilities can test real scale specimens of structures, so as to avoid the problems related to the application of scaling factors to respect the needed similitude rules (Xue et al. (2021)). In practice, this is generally accepted even for slightly scaled specimens, such as 2/3-scaled specimens (Occhipinti et al. (2025)). Once a representative seismic test can be performed, the dynamic response of the specimen must be caught through the collection of appropriate and accurate measurements. In particular, acceleration and displacements are generally recorded at the most relevant positions of the specimen. The recent but already quite consolidated advances in video based measurement systems do not represent a novelty anymore in the collection of displacements data for the analysis of the specimen deformations and of the crack pattern (Roselli et al. (2015); Liu et al. (2020)). Furthermore, video based acquisitions of dynamic identification tests are by now usable also for detecting the damage evolution of the specimen in terms modal frequency decay through modal analysis methods (De Canio et al. (2011)). The present work explores the potentialities of applying a procedure based on artificial intelligence (AI) techniques to the analysis of vibration data in dynamic identification tests as an alternative to the consolidated methods based on the modal analysis of the studied structure. In fact, both Deep Learning (DL) and Machine Learning (ML) techniques applied to the detection of the structural damage have already been explored in recent studies by Cha et al. (2024), Malekloo et al. (2022) and Cross et al. (2022). In the present study a specific DL method called Convolutional Variational Auto-Encoder (CVAE) was considered. CVAE has already been applied to detect and classify damage scenarios, but only very recently Ma et al. (2020) and Pollastro et al. (2023) investigated damage identification by CVAE at the scale of the single structure in laboratory experiments. In De Angelis et al. (2024) the CVAE application on-the-field to real buildings was explored. The proposed AI procedure comprises a training step on the ambient vibration data recorded at the studied structure in the initial conditions, e.g. in undamaged conditions. Once the algorithm is sufficiently trained, it is supposed to recognize the different conditions of damage, by analyzing the ambient vibration data of the structure having a certain level of damage. The considered metrics were the Mean Square Error (MSE) and the Original to Reconstructed Signal Ratio (ORSR) as proposed in Pollastro et al. (2023) and De Angelis et al. (2024), since they demonstrated that such two metrics have a good potential to distinguish data of a damaged structure from its undamaged conditions. More specifically, they coupled CVAE with a one-class support vector machine (OC-SVM) to classify data. Then, in the ORSR-MSE plane they found an effective graphical representation of the OC-SVM to discriminate data of the undamaged case from data of damaged scenarios. However, they just classified the studied cases in undamaged or not, without proposing a parameter to quantify different levels of damage. The further step forward we propose is to consider the centroid of each damage case cluster in the ORSR-MSE plane as representative of a given level of damage. The decay of the first modal frequency of the structure was considered to define the level of damage. In particular, it was used to formulate a damage index (DI) for each damage case. Then, the distance of each DI cluster centroid fromthe undamaged case cluster centroid (DI centroid distance) was explored as a parameter to quantify the damage. In order to validate the proposed procedure, it was applied to a shaking table campaign of a typical reinforced concrete frame. 2. Shaking table experiments The shaking table experiment was carried out on a reinforced concrete frame (RCF) specimen. It was a 2/3-scaled two-story frame designed according to the Italian standards and practice typical of the 1960s and 1970s when many buildings were constructed in Italy in the aftermath of World-War-Two destruction. Nowadays, it still represents one of the most common building typologies in Italy. The specimen had one-way ribbed slabs typically lightened by