Volumetric Defect Detection in Friction Stir Welding Through Convolutional Neural Networks Generalized Across Multiple Aluminum-Alloys and Sheet Thicknesses

Friction Stir Welding (FSW) is a solid-state welding process, which has strongly impacted welding technology, particularly for aluminum alloy applications. Due to its high-quality welds in all aluminum alloys, comparatively low specific heat input at high energy efficiency and ecological friendliness, FSW is used in a rapidly growing number of safety critical applications. Currently destructive and non-destructive testing methods are added as a separate process step to verify weld seam quality, adding complexity, cost, and time to the production. Imperfections are detected late in the production process and require costly rework or discarding of the assembly. Several studies have shown the possibility of using Deep Neural Networks to evaluated data recorded during the FSW process. Analyzed data includes thermal measurements, acoustic measurements, image or video data and most commonly the comparably large and distinctive process feedback forces. This study is a continuation of efforts by the research group. In this study Convolutional Neural Networks (CNN) based on the DenseNet architecture were successfully trained to classify FSW process force recordings supplemented with weld meta-data to detect volumetric subsurface defects. The data-sets were generated while welding different aluminum alloys in multiple sheet thicknesses over a wide range of spindle rotational speeds and feedrates. The networks classification accuracy as well as the ability to generalize across the different welded aluminum alloys, sheet thicknesses and corresponding welding tools was evaluated. Achieving a classification accuracy of 98.37%, the development aims to provide a reliable and cost-effective quality monitoring solution with a wide range of applicability to replace the required expensive and time intensive ultrasonic, x-ray or macro-section weld seam testing.