Optimization of CVAE parameters for the damage assessment of historic masonry buildings

Ivan Roselli, Domenico Palumbo

Abstract

Artificial Intelligence (AI) based on a deep learning procedure through Convolutional Variational Autoencoders (CVAEs) have been previously explored to analyze vibration data with the aim of assessing the state of damage of historic masonry structures subjected to seismic shakes. Nonetheless, CVAE application can be optimized by investigating the effect of different sizes of the used latent space and the time sequence length. Therefore, in the present work the optimal size of the two fundamental parameters of CVAE in the analysis of white-noise vibration data as dynamic characterization tool of shaking table tests of a rubble masonry prototype was investigated. The optimization process aimed at finding the best compromised size that leads to the maximum reconstruction capacity of the input and, hence, to the maximum utility for classification tasks. The results show that the used indicators, Mean Squared Error (MSE) and the Original to Reconstructed Signal Ratio (ORSR), lead to a bell-shaped optimization space, for the latent spatial dimension and the length of the time sequence, and therefore to the identification of a maximum point. The above optimization process was applied with remarkable results to vibration data of shaking table tests of a building prototype in historic rubble masonry typical of Central Italy.

Download

IMEKO-Metroarchaeo-2025-029.pdf

DOI

10.21014/tc26-2025.029

IMEKO TC

TC26 - Metrology for Cultural Heritage

Back to the proceedings