Machine learning models for determination of weldbead shape parameters for gas metal arc welded T-joints -- A comparative study

TL;DR

This study compares statistical and neural network models for predicting weld bead shape parameters in GMAW, finding that multiple linear regression outperforms neural networks in accuracy and error metrics.

Contribution

It introduces empirical models for weld bead shape prediction using MLR and ANN, with a detailed comparison highlighting the superior performance of MLR.

Findings

MLR models outperform ANN in prediction accuracy

Welding parameters significantly influence bead shape

Predictive models aid in numerical analysis of welding

Abstract

The shape of a weld bead is critical in assessing the quality of the welded joint. In particular, this has a major impact in the accuracy of the results obtained from a numerical analysis. This study focuses on the statistical design techniques and the artificial neural networks, to predict the weld bead shape parameters of shielded Gas Metal Arc Welded (GMAW) fillet joints. Extensive testing was carried out on low carbon mild steel plates of thicknesses ranging from 3mm to 10mm. Welding voltage, welding current, and moving heat source speed were considered as the welding parameters. Three types of multiple linear regression models (MLR) were created to establish an empirical equation for defining GMAW bead shape parameters considering interactive and higher order terms. Additionally, artificial neural network (ANN) models were created based on similar scheme, and the relevance of specific features was investigated using SHapley Additive exPlanations (SHAP). The results reveal that MLR-based approach performs better than the ANN based models in terms of predictability and error assessment. This study shows the usefulness of the predictive tools to aid numerical analysis of welding.