Inteligência artificial aplicada ao monitoramento de estruturas: detecção de alterações mecânico-estruturais baseada no uso de redes neurais autocodificadoras esparsas para a caracterização de respostas dinâmicas

Abstract

The evolution of computational technologies has brought remarkable advances in Structural Health Monitoring (SHM) systems, making the development of more automated strategies for detecting anomalous structural behaviors, through the adequate characterization of dynamic responses, plays an important role in this field of research. In the last few years, artificial intelligence algorithms focused on deep learning have become increasingly popular as tools for detecting changes in structures. The concept of deep learning has its origins in the theory of neural networks and comprises several approaches mainly formulated to deal with non-linear and high-dimensional problems. Although the basic foundations of such algorithms were established in the 1960s, their use in SHM strategies is still relatively new. In this context, the present work evaluates the deep learning algorithm called Sparse AutoEncoder (SAE) when used as an extractor of parameters from dynamic data. More specifically, the ability of SAE models to provide sets of representative variables is studied to detect abnormal structural behaviors through dynamic signal processing directly in the time domain. Firstly, a sensitivity analysis is performed, aiming to verify the influence of the variables that define the functioning of the SAE exert on its ability to represent vibrational responses. Then, in order to investigate whether the parameters modeled by SAE allow different structural states to be distinguished, a supervised strategy that uses SAE as a feature selector and the Support Vector Machine (SVM) algorithm as a classifier is applied to a numerical beam model and a monitored bridge in João Pessoa, Brasil. Finally, in an attempt to explore the capacity of the SAE in situations closer to those found in actual SHM systems, an unsupervised methodology for detecting structural alterations is proposed. The fundamental idea is to characterize the structural dynamic responses via SAE models and, subsequently, to detect the onset of abnormal behavior through the well-known T 2 Hotelling’s control chart, calculated with SAE extracted features. An automated criterion for models’ definition to be applied in SHM problems is also presented, based on the association of the model data reconstruction error and a proposed generalization index of the points T 2 . Most studies that have been carried out in the area classically define the SAE models most suitable as those with the lowest reconstruction error of dynamic signals, which does not guarantee a good performance in terms of structural alteration detection, as found here. The unsupervised approach is analyzed and exemplified in three cases: a laboratory steel frame, the classic case of the Z24 bridge, and an instrumented tower in Italy. It is noteworthy that the influence of temperature was also evaluated for the last two mentioned structures. In all cases studied, whether, in the supervised or unsupervised approach, satisfactory results were achieved, even under the influence of temperature variations, making it clear that the SAE presents itself as an effective and robust tool when applied to SHM problems.