DiffectNet: diffusion-enabled conditional target generation of internal defects in ultrasonic non-destructive testing

Abstract

Ultrasonic testing has been widely adopted as a non-destructive evalua- tion technique for detecting defect-related anomalies across various indus- trial fields. While several previous deep learning-based studies have shown promising results in addressing the inherent limitations of ultrasonic non- destructive testing, a critical challenge remains in acquiring diverse and large- scale datasets, hindering both detection performance and generalization. In this study, we propose a deep learning approach to generate synthetic defect cases tailored to phased array ultrasonic testing (PAUT) systems. Specifi- cally, we introduce a DiffectNet, a diffusion-enabled conditional target gen- eration network that can produce high-fidelity and defect-aware ultrasonic images. Both qualitative and quantitative evaluations demonstrate the su- perior generative performance of the proposed approach compared to exist- ing methods, achieving a 77% improvement in Fŕechet inception distance, a 98% improvement in kernel inception distance, and a 26% improvement learned perceptual image patch similarity error, respectively. Furthermore, we highlight the potential advantage of our approach as a neural augmenta- tion method, which can enhance model performance and generalizability for unseen defect scenarios. This study offers a promising solution to the practical challenge of limited data availability and further contributes to advancing data-driven ultrasonic non-destructive testing methods.