Speaker
Description
The use of Artificial intelligence (AI) has been recently applied for improving structural design and rehabilitation strategies, as well as structural monitoring of concrete buildings. In particular, Convolutional Variational Autoencoders (CVAEs) are advanced Machine Learning (ML) models used in AI applications that can be very effective in the analysis of vibration data to assess the damage conditions in concrete structures after earthquakes. The CVAE-based procedure is trained with the response of the undamaged structure and accuracy in the reconstruction of the ambient vibration response of the damaged structure was measured by the Mean Squared Error (MSE) and the Original to Reconstructed Signal Ratio (ORSR). The performance of CVAEs can be improved by optimizing two key parameters: the size of the latent space and the time sequence length of the processed data. This study investigated the optimal value for these two parameters when using a CVAE-based process to analyse white-noise vibration data from shaking table tests of a concrete frame specimen. The goal of this optimization methodology was to find a balance that maximizes the model’s ability to reconstruct the original input data, which is crucial for effective classification tasks of damage. The experimental results showed that the optimal values for both the latent space and the time sequence length can be found by Pareto method or by finding the maximum point of a bell-shaped surface generated to maximize the linear regression with a consolidated damage index based on the decay of the modal frequencies. This optimization process was successfully applied to vibration data from shaking table tests on a reinforced concrete building specimen, yielding significant results.