Sobre o monitoramento de saúde estrutural utilizando análise modal: uma abordagem de aprendizado de máquina e aprendizado profundo

Abstract

With technological advances, civil structures became taller, slender, and lighter. With bolder structural designs, newer and classical materials are often subject to dynamic loads arising from gusts of wind, earthquakes, and human occupation. To ensure reliability and safety in structures under environmental and occupational conditions, it is necessary to implement a Structural Health Monitoring system to verify the structural integrity and mitigate future failures. One component of the SHM is damage detection, localization, and measurement by human inspection. However, this process is susceptible to errors due to human presence itself, such as incorrect measurement or pathology identification. The damage quantification and localization can be determined by computational simulations, making it impossible for real-time monitoring due to the time required to execute such simulations and obtain answers close to reality. Because of mentioned factors, Machine Learning (ML) and Deep Learning (DL) techniques, combined with Modal Analysis, can be used for such tasks. They enable fast and real-time detection when subjected to the proper dataset and calibration of the model’s hyper-parameters. Thus, this undergraduate thesis aims to develop a computational tool capable of performing damage detection, localization, and quantification based on structural modal parameters and using Artificial Neural Networks (ANN) and Random Forests (RF). Several damaged and undamaged steel trusses and frames were submitted to Modal Analysis and simulated in the implemented software to determine the modal parameters and create input datasets for the aforementioned statistical models. The damage detection capability of each model was analyzed to find a model that generalizes two-dimensional steel frames and trusses under the boundary conditions used in the numeric simulations. Results show that RF is better than ANN in both structural models in the classification and regression tasks.