Characterization and classification of flaws in PAUT using a convolutional neural network

Phased Array Ultrasonic Testing (PAUT) is widely used in non destructive testing (NDT) for defect detection and characterization. However, interpreting PAUT data requires significant expertise, and results may vary depending on the inspector, especially when the weld geometry introduces artifacts that add complexity to the analysis. To address these challenges, this study employs a Faster R-CNN architecture to automate flaw detection and sizing. The impact of incorporating contextual information into the dataset is investigated, comparing raw images, with overlayed geometry, and a geometry subtraction techniques. The results demonstrate that enhancing the context improves accuracy and is necessary when adressing a challenging dataset.