Ultrasonic Flaw Classification with Self-Supervised Learning Using a DINO Framework

Deep Learning (DL) algorithms have been successfully applied in many ultrasound-based Non-Destructive Testing (NDT) applications for classification, characterization, and detection tasks. However, a significant challenge remains in the availability of labeled data. This work proposes the use of DINO (Self Distillation with No Labels), a self-supervised learning algorithm, to improve the accuracy and efficiency of ultrasonic NDT flaw classification. By utilizing unlabeled ultrasonic signal data, DINO addresses the limitation of well-labeled field data. We implement DINO with a ResNet-18 backbone and test its performance on experimental ultrasonic NDT data. Our results demonstrate that DINO achieves 94.7% accuracy using a linear classifier and 96.3% with an Support Vector Classification, matching the performance of supervised algorithms.