Optimised machine learning classification model to detect void formations in friction stir welding

Abstract

Friction stir welding is a novel solid-state joining process which has been a major area of research due to its applications in joining lightweight alloys. Although this welding process has the inherent advantage of joining alloys at significantly lower temperatures when compared to conventional fusion welding processes, thereby reducing energy consumption and thermal residual stresses, the overall final strength and feasibility of the joint has been found to be highly sensitive to the choice of welding process parameters. Defects such as void formations in friction stir welds are extremely detrimental to the strength of the weld joint as it could lead to catastrophic failures, expensive breakdowns in components, rejection of parts and even loss of lives (Yusof and Jamaluddin, 2014) [1]. These voids in the weld act as discontinuities for the joint as a result of which, the joint strength is severely weakened. Hence, it is of utmost importance to avoid the welding conditions that contribute to the formations of the void defects in structurally critical metallic joints. This study proposes an improved classification model based on a hyperparameter tuned XGBoost classifier algorithm to predict the probability of void formations with raw welding process parameters as inputs. The data set consisted of 108 entries of experimental data collected from past literatures on void formation for the friction stir welding of three aluminium alloys, AA2024, AA2219, and AA6061. While each alloy has unique chemical composition and mechanical properties, the study later identifies that the choice of alloy is amongst the weaker contributors to the presence of void defects. The proposed tuned model could successfully classify the joints containing voids with an accuracy of 90% while both the decision-tree based model and the neural-network based model could classify with an accuracy of 83.3%. © 2023