Force data-driven machine learning for defects in friction stir welding

Abstract

This study proposes a strategy for developing force-data-driven machine learning models to precisely predict defects and their types in friction stir welding (FSW). The characteristics of the three component forces in FSW, including traverse force (Fx), lateral force (Fy), and plunge force (FZ) are studied. The change in the force wave corresponded well with the variation in the defect. Fyavg had the best correlation with the characteristics of tunnel defects, whereas some other time-frequency features had negligible effects on the defect variation. The machine learning models built with the input of 15 force features could detect defects with an accuracy of 95.8% and classify them into tunnels and porosities with an accuracy of 98.0%. The abnormal increase in Fyavg, caused by the buildup of redundant material transported to the retreating side, was the main characteristic of force change when a defect was formed.