Uni and Multimodal Data Augmentation using Generative Adversarial Networks for Enhanced Multi Failure Classification in Turbine Engine Blades

Master thesis (2024)

Authors

P.M. IJzermans Aerospace Engineering

Contributors

M. Lourenço Baptista Air Transport & Operations - Aerospace Engineering (mentor)
I.I. de Pater Air Transport & Operations - Aerospace Engineering (mentor)
A. Bombelli Air Transport & Operations - Aerospace Engineering (graduation committee member)
R. Merino Martinez Aircraft Noise and Climate Effects - Aerospace Engineering (graduation committee member)
M. Deuten Royal Netherlands Aerospace Centre NLR (graduation committee member)
Faculty
Aerospace Engineering
More Info
expand_more

Published Date

16-09-2024

Language

English

Reuse Rights

Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.

Faculty

Aerospace Engineering

Abstract

Inherent subjectivity, inefficiencies, and the substantial cost related to human-based visual inspection of high-pressure turbine (HPT) blades has driven research in alternative automated techniques. The
combination of computer vision (CV) and deep learning (DL) provides a compelling alternative. However, as DL models increase in capability, their large parameter spaces require substantial data to
achieve robust training. Therefore, this study explores the use of two Auxiliary Classifier Generative Adversarial Networks (AC-GANs) to augment proprietary datasets for two failure modes: (a) a 3-
channel red-green-blue (RGB) model for obstructed holes, and (b) a 4-channel RGB plus depth (RGBD) model for foreign object damage, with the depth reconstructed using monocular depth estimation. A Differential Evolution Optimizer (DEO) was used for hyperparameter optimization of a ResNet-18 classification target model. In the obstructed hole dataset, GAN-based augmentation showed significantly improved accuracy, recall, F1, and AUC-ROC (p < 0.05), competing with traditional methods using less augmentation. In the foreign object damage dataset, the inclusion of depth information significantly enhanced accuracy, precision, F1, and AUC-ROC performance (p < 0.05). However, the RGBD augmentation mainly resulted in a trade-off between precision and recall, without statistically significant differences (p > 0.05).