Optimisation de la précision et de la robustesse de la mesure d’épaisseur des matériaux corrodés par l’intégration des réseaux de neurones Transformer

Abstract

Accurate evaluation of corroded material thickness is a major challenge in nondestructive evaluation (NDE) due to the complexity of ultrasonic signals and overlapping echoes. Given the limitations of Convolutional Neural Network (CNN) architectures found in the literature, this work explores the integration of Transformer Neural Networks (TNN) to improve the resolution and reliability of thickness measurements. Trained and validated on a dataset combining finite element simulations and experimental data from measurements on a corroded block, an initial hybrid CNN-TNN architecture was developed. Validated in a published, peer reviewed paper, this approach demonstrated that incorporating a Transformer module introduces a global learning context, enabling superior performance compared to existing CNN models. However, this method exhibited high computational complexity and a risk of overfitting. To address these limitations, an optimized version, Lite CNN-TNN, was designed by integrating a convolutional layer in parallel with the Multi-Head Self-Attention (MHSA) mechanism. This hybridization leveraged the CNN’s ability to extract local features while preserving the Transformer’s capacity to capture global interactions. This approach improved learning stability while significantly reducing the model’s parameter count. Furthermore, the simulation model was enhanced by incorporating realistic signals from experimental echoes, adjusting probe size, and refining the labeling methodology to better reflect real-world conditions. The learning process was also optimized through tailored data augmentation techniques. The results confirm the effectiveness of this approach : the Lite CNN-TNN model achieved a success rate of 98.84% on the experimental dataset and 97.40% on the simulated dataset, surpassing the best CNN model from the literature, which reached 98.70% and 95.70% respectively, while being 24 times faster.

Date	12 May 2025
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Pierre Bélanger (Supervisor)