PNNL

Abstract

Machine learning (ML) has shown huge potential in automated data analyses for ultrasonic nondestructive examination/evaluation (NDE) in the context of welding inspections. To assess the impact of ML on the reliability of ultrasonic NDE, the factors that influence ML performance need to be identified. In this work, we use a convolution neural network (CNN) model as a prototypic algorithm for the classification of ultrasonic B-scan images of welding defects. Ultrasonic data are collected from four stainless-steel specimens with weldment and two types of defects: saw cuts and thermal fatigue cracks. B-scan images were acquired using shear wave transducers on four stainless steel specimens with welding defects. A CNN model was built and trained with different flaw data for the B-scan image classification. Different training and testing combinations using the data from the four specimens were studied. The model trained with saw cut data showed good generalization for both saw cut and thermal fatigue crack (TFC) data. However, when using TFC data for training the ML model, poor performance was observed on test data from both saw cuts and TFC flaws. Low accuracy was also observed if the flaw was inside the weldment or had a small size. Therefore, flaw type, size, and location were seen to be critical factors affecting the ML model performance. The model trained by TFC was retrained with additional high-quality saw cut data, and the retrained model showed good performance on other saw cut data. Finally, preprocessing of the B-scan images was also found to impact the ML performance.