Recent advances in deep learning, particularly transformers and large vision models, have transformed natural language processing and computer vision. However, direct application of such models to automated analysis of Nondestructive Testing (NDT) data faces major challenges: limited labeled datasets, rigid code-based workflows, and the need for explainable, auditable results. In Phased Array Ultrasonic Testing (PAUT), the variability introduced by inspection geometry, technique parameters, and benign reflections (weld cap, back-wall, etc.) further complicates model training, making end-to-end deep learning impractical.

To address this, we propose an alternative approach that integrates code-mandated rules for detection, sizing, grouping, and classification with lightweight machine learning. Rules automatically generate a list of candidate indications, and features such as amplitude, position, and shape are extracted to train simple models to disqualify irrelevant geometric reflections. This hybrid method ensures full compliance with inspection standards, while offering explainable results aligned with human reasoning. Validation on field data from wind turbine weld inspections shows that the approach achieves performance comparable to certified inspectors, with reduced data requirements and improved robustness across inspection conditions.

Presenter: Mohammad Marvasti, Ph. D., CSWIP Phased Array Level 2