Eventos Anais de eventos
COBEM 2023
27th International Congress of Mechanical Engineering
WELD BEAD GEOMETRY PREDICTION USING ARTIFICIAL NEURAL NETWORKS AND GENETIC ALGORITHM HYBRID MODEL IN GMAW
Submission Author:
Caio Augusto Mascarenhas Dias Filho , AM , Brazil
Co-Authors:
Caio Augusto Mascarenhas Dias Filho, Marcus Eduardo Nascimento Sena, Dimas Alencar Neves, Sergio Brandi, Rubelmar Maia de Azevedo Cruz Neto
Presenter: João D'Anuzio Lima de Azevedo
doi://10.26678/ABCM.COBEM2023.COB2023-2140
Abstract
The capability to predict the weld bead geometry through the welding process parameters is of great interest, as it is directly related to both aesthetic and quality constraints, since parameters such as bead width, reinforcement and penetration define the mechanical strength of a welded joint. Within this context, this research focuses on using Artificial Neural Networks (ANN) to predict the geometric shape of the weld bead through the input parameters of Gas Metal Arc Welding (GMAW) process. Comparing a Backpropagation Neural Network (BPNN) and a Neural Network with Genetic Algorithm (GA-ANN) hybrid model, developed to overcome the main weakness of an ANN model, overfitting. For this research, data from 22 test specimens were used, employing different thicknesses of ASTM A36 steel plates, applying the GMAW process with a 1.2 mm diameter ER70S-6 wire and 75%Ar-25%CO2 gas mixture as shielding gas. For each test specimen, the arc voltage, welding current, and welding speed were varied, and the resulting bead geometry was obtained from the macrographs. The results showed that both predictive models developed can predict the weld bead geometry through geometric points accurately. However, as expected, the results for the GA-ANN hybrid model had a better performance on avoiding overfitting, predicting the weld bead geometry with greater confidence.
Keywords
Gas Metal Arc Welding (GMAW), Weld Bead Geometry Prediction, Genetic Algorithm (GA), Genetic Algorithm Artificial Neural Network (GA-ANN), Artificial neural network (ANN)
