Geração de assinaturas de impedância para o monitoramento de integridade estrutural Utilizando Deep Learning

Abstract

Generative adversarial networks (GANs) are capable of producing artificial data that are the same as real data, and have been successfully applied to various imaging tasks as a useful tool for data augmentation. In this work, we seek to develop a GAN-based signal generator to learn from electromechanical impedance signals and generate realistic and reliable representative data. The proposed approach will be designed to produce realistic synthesized signals and the generated signals can be used as real data for other applications in structural failure diagnosis. In order to evaluate the performance of the generator model, a set of quality assessments of the generated samples is introduced in the literature review. Finally, after modeling two case studies, it was possible to obtain an effective GAN in the generation of electromechanical impedance signals very similar to the experimental samples collected and inserted in the training network and obtaining a behavior of representative signals of temperature compensation.