Application of a U-Net Convolutional Neural Network to Ultrasonic Wavefield Measurements for Defect Characterization

Abstract

Recent advances in nondestructive evaluation (NDE) techniques have sought to improve the testing speed and accuracy of automatic flaw detection algorithms to minimize the costly downtime of removing in-service parts and components for testing and maintenance. Acoustic wavenumber spectroscopy (AWS) is a rapid NDE technique that utilizes steady-state ultrasonic excitation and laser Doppler vibrometer (LDV) measurements to identify component flaws orders of magnitude faster than traditional time-of-flight ultrasonic NDE techniques. However, current AWS technology is limited when applied to larger domains, such as rooms and larger structures, due to increased processing needs, and it is limited in accuracy and spatial resolution when applied to smaller defects on the order of one wavelength in size as well as defects on the edges of the structure. This paper presents the novel application of a U-Net style convolutional neural network (CNN) to improve the processing speed and spatial resolution of current AWS technology by performing semantic segmentation on simulated ultrasonic wavefield images of a steady-state, single-tone excitation of an aluminum plate. The well-adopted ResNet architecture, which was pre-trained on the large and openly available ImageNet dataset, was trained by transfer learning on the augmented wavefield dataset for the purpose of defect localization and characterization in aluminum plates. Finally, the performance of the CNN processing time and spatial resolution accuracy were shown to improve upon the processing methods of current AWS technology.