Uso de autocodificadores variacionais para a detecção de danos estruturais

Abstract

The state of integrity of a structure, as well as its monitoring, have constantly been the subject of study, given global technological advances in this area. Thus, the concept of Structural Health Monitoring (SHM) emerges, as a set of techniques and equipment used to investigate the state of integrity of structures. In this context, with the rapid evolution of computational and information technology, artificial intelligence algorithms focused on deep learning have been gaining more and more space when it comes to analyzing structural behaviors based on their dynamic responses. Within this scope, the present work is dedicated to the evaluation of a deep learning algorithm called Variational Autoencoder (VAE) when used as a feature extractor from dynamic data. In this way, the ability of VAE models to provide sets of variables that enable the detection of behavior considered abnormal for a given structure is evaluated. The methodology now developed consists of reconstructing dynamic signals using VAE models. From the reconstruction errors, two metrics are used to discriminate different damage scenarios in the analyzed structures: the MSE (Mean Square Error) and a combination of the MSE with the Mahalanobis Distance. Two real experimental applications are used to validate the proposed approach. Initially, a small-scale structure: a flat portico tested at the UFJF Images and Signals Laboratory; then a full-scale structure: the classic Z24 bridge in Switzerland. The results obtained for the frame were quite satisfactory, with the methodology capable of distinguishing the different structural behaviors with an accuracy varying close to 99%. However, in the case of the Z24 bridge, there were cases in which the classification was not completely adequate, which indicates the difficulty of working directly with raw vibration data to detect abnormal behavior in structures.