A supervised learning approach for prediction of x-ray computed tomography data from ultrasonic testing data

Abstract

Quantitative characterization of impact damage in polymer matrix composites (PMCs) with ultrasonic inspection is desired to enable improved prediction of damage evolution for lifing of composite structures. Post-processing of single-sided pulse-echo Ultrasonic Testing (UT) data produces 2D C-scan images that indicate the presence and 2D extent of delaminations, with very high depth resolution of the first reflector, while further damage below the first reflector is hidden. X-ray Computed Tomography (XCT) characterizes internal damage with a 3D voxel-based representation. Delaminations, matrix cracks, and surface-breaking cracks can be clearly visible in some XCT reconstructions. Modern damage evolution models take as input full 3D damage after impact and predict the growth of damage after loading, and need as accurate a representation of the full 3D damage as possible. This work discusses development of an approach for full 3D damage characterization using the desirable aspects of UT and XCT data. Machine learning models were developed to take as input a collection of UT pulse-echo scans of an impacted PMC panel and predict as output the results of an XCT scan in the form of a 3D voxel-based representation of damage. The models were trained on UT and XCT data from previous impacted PMC panels. The approach, including UT and XCT inspection data collection, feature extraction, training of the models, and evaluation of the models on new UT data is presented. The accuracy of the damage characterization results and challenges with this approach will be discussed.