Reconstruction de profils de corrosion à partir d’images ultrasonores à l’aide de réseaux antagonistes génératifs conditionnels (cGAN)

Abstract

Corrosion mapping using ultrasonic testing is widely employed in industry but remains a challenging task due to the complexity of interpreting the acquired data. To address this, a method based on a conditional Generative Adversarial Network (cGAN), trained on both simulated and experimental data, is proposed to interpret ultrasonic images and generate schematic corrosion profiles that assist inspectors in data interpretation. Simulated datasets were generated using a GPU-accelerated finite element solver, incorporating realistic corrosion patterns, probe characteristics, and noise. Experimental data were acquired with high positional accuracy using a robotic arm and ultrasonic phased array probe. A U-Net generator is employed in the cGAN architecture to translate ultrasonic images into schematic corrosion representations, while a convolutional discriminator ensures realism in the generated outputs. Several training strategies, including simulated-only, mixed, and transfer learning approaches, were evaluated and compared in this paper. Performance was assessed using metrics such as Mean Squared Error, Structural Similarity Index, Pearson correlation, and Hausdorff distance. Additionally, the proposed model is compared with the traditional maximum-intensity method, demonstrating a substantial improvement in thickness estimation and corrosion profile accuracy. Results indicate that the Total Focusing Method (TFM) inputs consistently yield superior reconstruction quality, and that transfer learning substantially improves the model’s performance, particularly when the inputs are lower-resolution B-Scan. The proposed approach demonstrates the potential of AI-driven ultrasonic imaging to accurately reconstruct complex corrosion profiles, providing a framework for enhanced nondestructive evaluation in industrial applications.

Date	11 Dec 2025
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Pierre Bélanger (Supervisor)