Prediction of electron beam weld quality from weld bead surface using clustering and support vector regression

Destructive manual experiments are primarily used for quality assessment of electron beam welded components, which consume the resources and time significantly. Vision-based and other NDT-based quality monitoring in EBW is rarely reported, and consequently, more investigation is required. An inexpensive quality evaluation technique for the welded joints has been proposed using the weld bead surface attributes. These attributes have a good correlation with quality indicators like aspect ratio and mechanical properties of the joint. Here, electron beam welding of CuCrZr alloy was conducted to get varying weld profiles, and the surface weld attributes were measured. The model correlating top and bottom bead widths with the process variables was established using support vector regression. The predicted unlabelled weld attributes were employed as inputs for the unsupervised clustering algorithms, namely fuzzy C-mean clustering and density-based clustering. The clustering algorithms provided different clusters of the joints, namely poor, fair and good. FCM was found to be more precise for the clustering of welding data. Poor category of the joints was represented by that with partial penetration and lower aspect ratio, and the corresponding process variables’ combination could be avoided in the next iteration through the feedback. The fair category represented the joints with a high aspect ratio and keyhole-based profiles with moderate strength and ductility. The good category signified the weld joints having the maximum joint strength and ductility, and an acceptable aspect ratio with the conduction mode of welding. Keyhole-based weld profile could be obtained using high beam power (5.5 kW to 6.3 kW) with high welding speed (1000 mm/min and above). A combination of 0.1 mm oscillation amplitude and 900 Hz oscillation frequency was found to be the most favourable one, in this study.