Automated Defect Detection for Non-Destructive Evaluation by Radar Imaging and Machine Learning

This work provides a concept for a fully automated, non-invasive evaluation of dielectric objects by use of synthetic aperture radar (SAR) imaging combined with machine learning for application in non-destructive evaluation (NDE). For demonstration purposes we investigated a set of polyvinylchloride (PVC) test objects. A millimetre-wave imaging system is used to scan the objects. By employing subsurface SAR imaging techniques, high-resolution images from the objects' insides can be obtained. These images are used as a dataset for a convolutional neural network. By training the network on several typical defect structures such as cracks, delaminations or contaminations, the proposed procedure enables an automated decision whether an object under test is intact or defective, independent of the kind of defect.