Post-Process Semantic Segmentation of Ultrasonic M-scans from Resistance Spot Weld Process Monitoring

Abstract

combinations into modern vehicles. These new challenges necessitate new solutions. Resistance spot welding (RSW), a key joining process in several industries including automotive manufacturing, has become increasingly challenging in its own right, especially due to the use of novel advanced materials. Consequently, there is an increased need for nondestructive evaluation (NDE) systems that are high-resolution, fast, accurate, and fully automated – and the same needs apply to the interpretation of their outputs. Such systems also align with smart factory and Industry 4.0 ideals. In this work, we developed a semantic segmentation approach using deep learning (DL) for the automated characterization of ultrasonic M-scan data from RSW process monitoring. The approach segments nugget and stack regions in the data, and the segmentation mask outputs can be used downstream for further weld process characterization e.g. computation of nugget penetration into each sheet, nugget growth rate, nugget solidification rate, etc. Our approach yielded overall intersection-over-union (IoU) of 0.978 which shows a very strong ability to localize the nugget and stack regions in the M-scans, and an F1 score of 0.995 which indicates that it rarely misjudges nugget or stack region presence. Our work demonstrates that semantic segmentation using DL can be an extremely powerful approach for RSW ultrasonic NDE data interpretation. The approach can be used to similar effect on NDE data from other modalities and from other use cases. Our work provides the capability for immediate, accurate, and comprehensive post-process inspection for RSW which can be used for quality control as well as post-process feedback, and it represents another major advancement toward Industry/NDE 4.0.