Aplicação de autocodificadores convolucionais para identificação de danos em estruturas

Abstract

Continuous monitoring of structural integrity has been highlighted due to recent incidents that highlight the importance of ensuring safety and predicting the dynamic behavior of structures. In general, this monitoring is performed by means of sensors attached to the analyzed structures, which provide data for specific algorithms capable of identifying their integrity status. In this context, this work focuses on the application of technologies such as machine learning and artificial intelligence, with an emphasis on convolutional autoencoders, to assess the integrity of structures through the analysis of the collected data. To this end, four analysis strategies based on the reconstruction errors of dynamic signals and the Mahalanobis distance of latent spaces, obtained by models based on convolutional autoencoders, are used. The proposed methodology was validated using two data sets. The first set, a double-clamped frame, tested at the Signals and Images Laboratory of the University of Juiz de Fora, represents a structure with controlled damage scenarios. On the other hand, the second application, the Z-24 bridge, represents a real structure subject to variations of external factors. The results obtained demonstrated the effectiveness of the strategies in controlled environments and real structures, since in both cases, it was possible to detect structural changes in a satisfactory manner, indicating that the CAE approach can be effective in real structures.