Automated detection and characterisation of defects from multiview ultrasonic imaging

This paper explores how different off the shelf machine learning (ML) models may be applied to the problem of immersion inspection with ultrasonic phased arrays. The paper addresses the problem of training set availability through the application of a well validated forward model of the system. This allow training sets of arbitrary size with arbitrary defect locations and types to be generated. This is generally not possible with purely experimental datasets. The availability of these large datasets allows relative assessment of the capabilities and limitations of a range of ML models. In doing this we show that shallow learning has similar performance to deep learning for defect detection and significant additional benefits. The paper then explores the problem of defect characterisation and show that deep learning models can offer excellent performance. The same models can also be readily re-purposed to different challenges, such as determining a defect's type.

Automated summary

This paper explores the application of different off the shelf machine learning models for ultrasonic phased array immersion inspection. By using a well validated forward model, training sets with arbitrary sizes and defect locations can be generated, allowing relative assessment of the capabilities of various ML models. The paper shows that shallow learning performs similarly to deep learning in defect detection, and deep learning models offer excellent performance in defect characterisation as well as potential for repurposing to different challenges.