Automatic Feature Extraction Based on Meta-Learning for Ultrasonic Flaw Classification

Abstract

Deep Learning based classifiers such as Convolutional Neural Networks (CNN) have been successfully used in Ultrasonic non-destructive testing applications. For example, CNN performs well for flaw classification when trained and tested with a fixed number of flaw types. However, CNN architectures built for the purpose of classification of ultrasonic flaws need to redesigned or retrained in order for the new flaw types to be recognized by CNN. Meta Learning provides a solution to this issue as it can generalize feature extraction and compare the features with the template features of respective flaw type. In this work, we introduce a new Siamese Network based encoder and classifier architecture to generalize feature extraction and ensure robust classification of flaw types. Preliminary results for One-Shot One-Way classification demonstrate an accuracy of above 90%, confirming that Meta Learning can be successfully adopted for Ultrasonic flaw detection tasks.