Eventos Anais de eventos
COBEM 2021
26th International Congress of Mechanical Engineering
APPLICATION OF ARTIFICIAL NEURAL NETWORKS IN DAMAGE DETECTION
Submission Author:
Matheus Janczkowski Fogaça , SC , Brazil
Co-Authors:
Matheus Janczkowski Fogaça, Eduardo Lenz Cardoso, Ricardo De Medeiros
Presenter: Matheus Janczkowski Fogaça
doi://10.26678/ABCM.COBEM2021.COB2021-0519
Abstract
Composite materials are a recurrent choice in Engineering, since their strength to weight ratio is steadily increasing and there is the possibility to tailor the material properties during construction. Therefore, it is fundamental to develop methodologies to detect damage in these materials. A powerful tool for this task is the vibration analysis, whose outputs may be interpreted with Artificial Neural Network (ANN) to identify and classify whether a sample is damaged or not. This work aims to evaluate and to investigate the relationship between the performance and the parameters of a feed-forward ANN to be applied in damage detection of composite structures. To this end, topology, activation functions, data compressing methods, optimizers, testing, training, and validation data sets are analyzed. Vibration data from Glass Fiber Reinforced Plastic (GFRP) composite beam are used. However, it is not feasible to directly use large dimensional data, i.e., Frequency Response Function (FRF), which could demand more complex models and high computational cost. Therefore, different techniques for dimension reduction, like Principal Component Analysis (PCA), Dislocated Series (DS), and Linear Discriminant Analysis (LDA), are investigated. Julia language, especially the Flux library, is employed to build the ANN, validate, train, and test the models, which classify the samples as intact or damaged. Different damage patterns are generated to evaluate the performance of the methodology. Thus, every aspect of the machine learning process is compared to find which combination of parameters is the best for this application. The preliminary results show that the ANN allied with the vibration analysis in the frequency domain can be a powerful tool to detect damage in Engineering situations before catastrophic failure and without the necessity of expensive assays.
Keywords
machine learning, Artificial neural networks, Composite Materials, Dataset Analysis, Optimization
