Eventos Anais de eventos
DINAME2019
DINAME2019
Artificial Neural Network Application for Structural Damage Diagnosis from Vibration Measurements
Submission Author:
Calebe Paiva Gomes de Souza , PI , Brazil
Co-Authors:
Calebe Paiva Gomes de Souza, Paulo Roberto Gardel Kurka, Romulo Lins, José Medeiros de Araújo Júnior
Presenter: Calebe Paiva Gomes de Souza
doi://10.26678/ABCM.DINAME2019.DIN2019-0093
Abstract
The structural damage diagnosis aims to verify, in the first stage, if damages exist in the structure and if they can compromise its operation. During the second stage, it is mandatory to localize the damage and determine its severity level. For developing an efficient diagnosis, one possibility is to use vibration measurements, because a damage modifies the modal parameters of structures. However, the efficient damage detection still remains a challenge, and then, modern techniques have been proposed in order to improve the initial diagnosis, such as the use of Artificial Neural Network (ANN). In this paper, a ANN-type Multi-Layer Perceptron (MLP) will be used, being that its architecture will be formed in three layers (input layer, hidden layer and output layer). Moreover, the samples to feed the MLP will be obtained from a numerical modeling of a simply supported beam by using Finite Element Method, which will supply dynamic response of the structure in two main cases: undamaged structure and with various damage scenarios. From the vibration response, some samples will be used for training the ANN by using backpropagation algorithm and Gradient Descent Method in a supervised learning, once for each damage will be known input data (modal parameters) and the expected results of damage location and its severity. Many ANN topologies are analyzed, varying both the number of hidden neurons and the activation function. The best topology will be obtained from the analysis of statistical parameters. In addition to the ``tanh'' activation function, the ``ISRU'' function, which has never been used to diagnosis of structures based on vibration measurements, will also be verified. After training procedure, the validation set will be inputted into the trained ANN in order to prove that it can detect, localize and quantify damages arbitrarily with accuracy and reliability.
Keywords
Vibration-based damage diagnosis, Pattern recognition and classification, Finite Element Method, artificial neural network, Multilayer Perceptron
