E4-a3 - Investigation of a Deep Learning Methodology for Automatic Detection and Characterization of Crack-Type Defects in Ultrasonic Non-Destructive Testing
- Event
- 2025 ICU PADERBORN - 9th International Congress on Ultrasonics
2025-09-21 - 2025-09-25
Paderborn - Band
- Lectures
- Chapter
- E4-a - Machine Learning and Statistical Methods in Acoustics
- Author(s)
- P. S. Sheehan, R. Miorelli, S. Robert, B. Chapuis - Université Paris Saclay, Palaiseau (France), S. Chatillon - EDF – Direction de la Qualité Industrielle, Saint-Denis (France)
- Pages
- 295 - 298
- DOI
- 10.5162/Ultrasonic2025/E4-a3
- ISBN
- 978-3-910600-08-9
- Price
- free
Abstract
The paper introduces a supervised simulation-based deep-learning pipeline for characterising welding defects with ultrasonic arrays in Non-Destructive Testing and Evaluation. Synthetic, multimodal TFM images generated with CIVA form the training and validation sets. The pipeline trains several state-of-the-art models using automated hyperparameter optimization. The trained models are then applied to experimental data collected under conditions mirroring the simulations. We compare regression performance and outline future directions.