Enhancement of defect data sets using a Generative Adversarial Network (GAN) for surface inspection of aerospace components

During aircraft maintenance, components are visually inspected for defects. There is no photographic documentation of the defects. In the future, sensor technology will be used to inspect components for defects in order to increase the quality of the diagnostic data, to digitise it more and to speed up the process. A promising approach for defect detection in image recordings are neural networks, which, however, require a large amount of training data. Existing 2D defect datasets, which could be used for training an AI for this application by means of transfer learning, often show surface defects only in small numbers, underrepresented compared to other defects or with low variance. The aim of this work is to design and apply a GAN to existing datasets to increase quantity and variance.

Your subtasks

  • Research on defect datasets, synthetic training data, neural networks (GANs)
  • Design and implementation of a GAN to extend existing data sets
  • Investigation regarding the successful expansion, quality of the generated training data, optimisation potentials

Your profile

  • You are studying mechanical engineering, mechatronics or a comparable subject
  • You have knowledge, skills or an interest in digitalisation, AI
  • You have programming skills (Python)

Organisational

  • Processing mainly possible in the home office

 

If you are interested, please contact: Falko Kähler, M.Sc. | f.kaehler@tuhh.de

Leave a Reply

Your email address will not be published. Required fields are marked *