Generation of Ultrasonic Images from Resistance Spot Welding Using Deep Learning
Standing
Undergraduate
Type of Proposal
Oral Research Presentation
Challenges Theme
Open Challenge
Faculty Sponsor
Dr Roman Gr. Maev
Proposal
By running an electric current through electrodes that firmly clamp down on a stack of metal sheets, a resistance spot welding (RSW) system joins them together. Nondestructive evaluation (NDE) of these joints is vital for various sheet assembly industries including automotive manufacturing. At the Institute for Diagnostic Imaging Research, we use a sophisticated inspection approach involving an inline ultrasonic system which creates 2D signatures of the weld process. Thus, our approach creates an ultrasonic signature of every weld created by a given robot on which our system is installed, providing complete traceability and quality control. Development of artificial intelligence (AI) systems for automating NDE data analysis is key for progression to Industry/NDE 4.0. Such AI systems, especially those developed using state-of-the-art deep learning techniques, require significant datasets, which are costly and time-consuming to develop and annotate for AI development. Variational autoencoding and generative adversarial networks are two examples of contemporary AI-based techniques that can be used to learn from existing ultrasonic images from real RSWs, in a way such that artificial ultrasonic image samples can be generated, relieving the burden of dataset development. These AI approaches can even allow the conditional generation of artificial images from welds with never-before-seen weld parameters, materials, and sheet thicknesses. Successful development of a conditional generative AI approach would be groundbreaking in the NDE community and could be used in other NDE applications and on data from other NDE modalities as well.
Grand Challenges
Viable, Healthy and Safe Communities
Generation of Ultrasonic Images from Resistance Spot Welding Using Deep Learning
By running an electric current through electrodes that firmly clamp down on a stack of metal sheets, a resistance spot welding (RSW) system joins them together. Nondestructive evaluation (NDE) of these joints is vital for various sheet assembly industries including automotive manufacturing. At the Institute for Diagnostic Imaging Research, we use a sophisticated inspection approach involving an inline ultrasonic system which creates 2D signatures of the weld process. Thus, our approach creates an ultrasonic signature of every weld created by a given robot on which our system is installed, providing complete traceability and quality control. Development of artificial intelligence (AI) systems for automating NDE data analysis is key for progression to Industry/NDE 4.0. Such AI systems, especially those developed using state-of-the-art deep learning techniques, require significant datasets, which are costly and time-consuming to develop and annotate for AI development. Variational autoencoding and generative adversarial networks are two examples of contemporary AI-based techniques that can be used to learn from existing ultrasonic images from real RSWs, in a way such that artificial ultrasonic image samples can be generated, relieving the burden of dataset development. These AI approaches can even allow the conditional generation of artificial images from welds with never-before-seen weld parameters, materials, and sheet thicknesses. Successful development of a conditional generative AI approach would be groundbreaking in the NDE community and could be used in other NDE applications and on data from other NDE modalities as well.