Aplicação de redes neurais artificiais na predição de perfurações durante a soldagem de juntas tubulares em um Chassi Baja

Abstract

The quality of welded joints in an automobile chassis is essential to ensure safety and performance. In order to reduce cost, meet deadlines and ensure maximum quality, knowledge about adjustment settings and their adjustment is fundamental for operations. The objective of this study is to train an artificial neural network for prediction of burn-through in welded joints using the GMAW process with ER70S-6 electrode in SAE 1020 steel tubes of a Baja SAE chassis. Electrode wire diameter, type of shielding gas, source voltage and feed speed were used as predictor variables. Two weld beads are welded to each specimen, following the configurations determined by a DOE (Design of Experiments). Following the weldings, it was observed which bodies showed burn-through in the base metal. After the observations, a PMC network was trained in Python and the model's responses were compared to the tests. The model presented an accuracy of 78% and a sensitivity of 76% for data, which was not trained, revealing the potential of using artificial neural networks to predict burn-through in welded joints.