Post-process Semantic Segmentation of Ultrasonic M-scans 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
Ultrasonic imaging has allowed for post-process non-destructive evaluation (NDE) of resistance spot welds (RSW). Resistance spot welding is a widely accepted technology to join multiple metal sheets together. Deep learning has become state-of-the-art for many tasks in computer vision which has led to its increasing use in NDE. The objective of this work, conducted at the Institute for Diagnostic Imaging Research, is to develop an AI system that can characterize the top and bottom of the weld stack, and the top and bottom of the molten nugget in an ultrasonic image which will help determine valuable information such as the penetration of nugget into each sheet, size of the nugget (diameter), etc. In the end this valuable information will help determine the quality of the weld. Semantic segmentation was the approach we used to classify each pixel in the ultrasonic image. We used lightweight, customized versions of image segmentation architectures such as U-Net, DeepLab, and SWIN U-Net to create our system. The usage of three different architectures allowed us to compare models and select the best performing one. The results of each model were promising, reaching a combined intersection-over-union for nugget and stack regions >0.975.
Grand Challenges
Viable, Healthy and Safe Communities
Post-process Semantic Segmentation of Ultrasonic M-scans from Resistance Spot Welding Using Deep Learning
Ultrasonic imaging has allowed for post-process non-destructive evaluation (NDE) of resistance spot welds (RSW). Resistance spot welding is a widely accepted technology to join multiple metal sheets together. Deep learning has become state-of-the-art for many tasks in computer vision which has led to its increasing use in NDE. The objective of this work, conducted at the Institute for Diagnostic Imaging Research, is to develop an AI system that can characterize the top and bottom of the weld stack, and the top and bottom of the molten nugget in an ultrasonic image which will help determine valuable information such as the penetration of nugget into each sheet, size of the nugget (diameter), etc. In the end this valuable information will help determine the quality of the weld. Semantic segmentation was the approach we used to classify each pixel in the ultrasonic image. We used lightweight, customized versions of image segmentation architectures such as U-Net, DeepLab, and SWIN U-Net to create our system. The usage of three different architectures allowed us to compare models and select the best performing one. The results of each model were promising, reaching a combined intersection-over-union for nugget and stack regions >0.975.