Real-time In-process Ultrasonic M-scan Segmentation Using Deep Learning For Adaptive Resistance Spot Welding
Standing
Undergraduate
Type of Proposal
Oral Research Presentation
Challenges Theme
Open Challenge
Faculty Sponsor
Dr. Roman Gr. Maev
Proposal
Advances in ultrasonic imaging techniques allow for the real-time non-destructive evaluation (NDE) of resistance spot welds (RSW) as they progress. In real-time RSW ultrasonic NDE data analysis, the most important features to characterize are the existence and positions of key interfaces – the top and bottom of the welded stack, and the top and bottom of the molten nugget – which allow for the estimation of resultant weld nugget size, position, and penetration into the welded stack. Deep learning has established the state of the art for many tasks in computer vision and natural language processing, and consequently it has seen increased use for related tasks in NDE (e.g., feature extraction, signal processing, sequence processing, etc.). The objective of this proposed work is to develop an AI system that characterizes ultrasonic RSW NDE data in real-time such that it can be used to provide closed-loop feedback to a weld controller in an adaptive welding system. Preliminary work shows the feasibility of a UNet-based convolutional LSTM to conduct semantic segmentation of ultrasonic data. The resultant image masks from segmentation allow for the calculation of total nugget penetration and the estimation of lateral nugget shape. These automated measurements can be fed back to a weld controller to allow weld schedule adaptation, moving automotive manufacturing another step closer towards the ultimate goal of zero-defect manufacturing.
Grand Challenges
Viable, Healthy and Safe Communities
Real-time In-process Ultrasonic M-scan Segmentation Using Deep Learning For Adaptive Resistance Spot Welding
Advances in ultrasonic imaging techniques allow for the real-time non-destructive evaluation (NDE) of resistance spot welds (RSW) as they progress. In real-time RSW ultrasonic NDE data analysis, the most important features to characterize are the existence and positions of key interfaces – the top and bottom of the welded stack, and the top and bottom of the molten nugget – which allow for the estimation of resultant weld nugget size, position, and penetration into the welded stack. Deep learning has established the state of the art for many tasks in computer vision and natural language processing, and consequently it has seen increased use for related tasks in NDE (e.g., feature extraction, signal processing, sequence processing, etc.). The objective of this proposed work is to develop an AI system that characterizes ultrasonic RSW NDE data in real-time such that it can be used to provide closed-loop feedback to a weld controller in an adaptive welding system. Preliminary work shows the feasibility of a UNet-based convolutional LSTM to conduct semantic segmentation of ultrasonic data. The resultant image masks from segmentation allow for the calculation of total nugget penetration and the estimation of lateral nugget shape. These automated measurements can be fed back to a weld controller to allow weld schedule adaptation, moving automotive manufacturing another step closer towards the ultimate goal of zero-defect manufacturing.