Eventos Anais de eventos
COBEM 2021
26th International Congress of Mechanical Engineering
Development of a defect detection methodology in beams using modal analysis data in optimized artificial neural networks.
Submission Author:
Leonardo Tavares , PA
Co-Authors:
Leonardo Tavares, Wesley Homem, Javier Ferreira, Camila Santos, Eduardo Mendes, LEONARDO DANTAS RODRIGUES
Presenter: Wesley Homem
doi://10.26678/ABCM.COBEM2021.COB2021-1835
Abstract
The constant search for the optimization of structural designs aiming at material savings and lower weights make it even more important to improve their monitoring systems regarding the appearance of damages that may lead to their failure. In this way, many researchers have dedicated themselves to the development of techniques for Structural Health Monitoring (SHM) in real time. Such monitoring generates a large number of information that needs to be properly treated in order to provide accurate diagnoses regarding the existence of damage, its location and severity. Advanced artificial intelligence techniques are being increasingly used in these monitoring methodologies, with emphasis on artificial neural networks (ANN). This work deals with the development of a defect detection system in beams using a trained and validated backpropagation ANN with data from numerical modal analysis using the finite element method. To guarantee a large volume of data to improve the effectiveness of ANN, a python code was developed that communicates directly with Ansys APDL, automatically generating several models of beams with cracks with different positions and depths. The generated and input data are sent to ANN also programmed in Python, being divided into groups for training, validation and testing. And then, with the data prepared, a grid search is used, in the neural network with three hidden layers, in order to optimize its design by finding the number of neurons that obtained the best result within a predetermined range for the hidden layers , under these conditions, the value of twenty neurons demonstrated the best efficiency. For this optimized ANN, the accuracy of the tool to determine the location of the defects with a depth of more than 40% was 83%, which can be considered satisfactory.
Keywords
Structural Health Monitoring, Artificial neural networks, Python, Damage Detection
