• Tunukovic, Vedran
• Dobie, Gordon
• Ohare, Tom
• Mohseni, Ehsan
• Lawley, Alistair
• Mcknight, Shaun
• Macleod, Charles N.
• Pierce, Stephen

Abstract

The growing interest in applying Machine Learning (ML) techniques in Non-Destructive Testing (NDT) to assist expert detection and analysis is facing many unique challenges. This research seeks to create an object detection network that would automatically generate bounding boxes around various defects found in Carbon Fibre Reinforced Polymers (CFRPs) through which the quantitative defect size information can be inferred. CFRPs are structurally anisotropic resulting in complex physical interactions between the emitted acoustic waves and the material structure when Ultrasonic Testing (UT) is deployed. Therefore, the structural noise makes the detection of various types of defects, such as porosities, delaminations and inclusions, that are frequently observed in CFRPs [1] even a more challenging task. In order to take a supervised learning approach in the detection of defects, a training dataset must be produced and labelled. Extensive automatic methods for data collection exist, however, in many cases labelling is done manually, which requires extensive use of expert time. Therefore, a method for automatically labelling simple defects could potentially be useful for accelerating the ground truth creation and allowing experts to focus on the detection of more complex defects.

Topics

• impedance spectroscopy
• polymer
• Carbon
• inclusion
• anisotropic
• ultrasonic
• machine learning