Analysis and estimation of extension of the fusion zone and sub-regions of the heat affected zone through neural networks in weld beads produced by GMAW process

Abstract

This study aims to evaluate the use of sensor fusion, based on neural networks, to estimate the extension of the fusion zone and of the sub-regions of heat affected zone in GMAW processes. Three samples were produced by depositing weld beads composed by AWS ER70S-6 wire on SAE 1020 steel plates varying the welding voltage, travel speed, and wire-feeding speed. Furthermore, the deposition parameters were varied during the deposition in each of the samples. The infrared radiation profiles emitted from the samples were gathered using the thermographic camera. The final microstructure was characterized by Optical Microscopy, Scanning Electron Microscopy and X-ray Diffraction. Finally, estimation models of the microstructure of weld beads are developed. The results obtained showed a gradual reduction in the fractions of the predominant phases of the fusion zone until the existence of only primary and perlite ferrite as the distance from the melting line increases. With the data obtained by thermography, microscopy and the welding parameters acquired by the interface, 6 network models (A to F) were proposed testing different configurations. The Model F with 17 and 15 neurons in the 1st and 2nd hidden layers trained with Bayesian Regularization and 4 training cycles showed good results in comparison with the other models (average error of 12.9%) and was satisfactory in terms of practicality in estimate of the microstructure of the welded region.