Détection de défauts dans les scans sectoriels avec des architectures d’apprentissage profond avancées

Abstract

Phased Array Ultrasonic Testing (PAUT) has become a widely adopted technique in nondestructive testing (NDT) due to its precision and efficiency. However, the interpretation of PAUT data remains highly dependent on the inspector’s expertise, and the presence of geometric features within materials generates artifacts that complicate analysis and increase interpretation time. Although artificial intelligence has demonstrated excellent performance in object detection across various fields, its application to ultrasonic data remains relatively limited, mainly due to data confidentiality constraints. This study investigates the use of deep learning for the automated detection of weld defects using PAUT data. Two architectures are compared : an enhanced Faster R-CNN and the latest versions of YOLO (v5 and v8). To explore the influence of contextual information on model performance, three distinct data configurations are analyzed : raw PAUT images used as a reference, PAUT images with a schematic overlay indicating weld geometry, and PAUT images preprocessed by median geometry subtraction to remove structural artifacts and highlight defect regions. Training and validation are conducted on industrial weld specimens containing various defect types, including lack of fusion, porosity, and cracks. The integration of components such as ROI Align, EfficientNet-B5, and a feature pyramid network, combined with K-means–based anchor box optimization, enables the enhanced Faster R-CNN to achieve the best performance, reaching a mean Average Precision (mAP) of 58.05% when combined with median geometry subtraction, outperforming YOLO models (up to 57.03%). Median geometry subtraction improves mAP by 5 to 7 percentage points while reducing false positives. Conversely, the addition of the overlay does not improve performance and sometimes hinders detection, despite its alignment with standard inspection practices.

Date	28 Dec 2025
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Pierre Bélanger (Supervisor)