Real-Time PAUT Defect Classification with Class-Weighted Knowledge Distillation
Minsu Jeon, R. Guyon, C. Fisher, D. Mun, Jaebeom Lee
TL;DR: This work proposes a PAUT defect-classification framework that minimizes false negatives by applying class-weighted Knowledge distillation (KD) to transfer a Vision Transformer (ViT) teacher’s diagnostic capability to an extremely lightweight linear student.
Phased array ultrasonic testing (PAUT) provides high-resolution subsurface imaging through electronic beam steering and focusing and is widely used for internal defect diagnosis in structures. However, effectively interpreting PAUT data requires understanding complex spatial-temporal patterns, and deep learning has recently been introduced to address this challenge. While deep learning–based automatic interpretation has improved performance [1], most prior approaches rely on heavy CNN models that inadequately capture long-range dependencies and incur high computational cost, limiting real-time operation and on-device deployment. This work addresses these limitations by proposing a PAUT defect-classification framework that minimizes false negatives by applying class-weighted Knowledge distillation (KD) to transfer a Vision Transformer (ViT) teacher’s diagnostic capability to an extremely lightweight linear student. The proposed approach preserves high reliability while achieving real-time performance. The proposed method has three stages. First, we construct a dataset. PAUT responses are sensitive to beam angle, which can obscure or weaken defect echoes, so we pair left and right acquisitions and use them together as a two-channel input. Next, we train a ViT as the teacher model with class-weighted cross-entropy so that the model becomes more sensitive to defect classes. Finally, we transfer the teacher’s knowledge to a lightweight linear student through class-weighted KD. In this study, the lightweight linear student model consists of one convolutional layer and one linear layer. Standard KD combines a KL divergence term between teacher and student distributions with a ground-truth cross-entropy term, and the ground-truth term can reintroduce false alarms even when the teacher suppresses them. To avoid this issue and to preserve recall on defect classes, we apply the same class weights to the ground-truth term during distillation. To validate the performance of the proposed method, we compared the student model trained with the proposed scheme against ResNet and ViT models. On 937 test samples, all three models achieved an accuracy above 90%. However, while both ResNet and ViT produced two false negatives, the proposed method resulted in zero false negatives. In addition, the inference throughput was 227 images/s for ResNet and 135 images/s for ViT, whereas the proposed method achieved 2,482 images/s, reducing the inference time to less than one tenth of that of the baseline models. Per-sample inference time was about 0.0004 s, confirming real-time operation. The target computing platform for this work is an embedded inspection system under development at CEA, intended for integration with PAUT sensing hardware. While the proposed framework is ultimately envisioned to be deployed at the edge, directly on PAUT systems, the present study focuses on algorithmic feasibility rather than full system-level deployment. References [1] Guyon, R., Newson, M., Fisher, C., Miorelli, R., & Roué, D. (2025, July). Towards embedded AI models for welding defect detection in pipes. In 2025 IEEE Sensors Applications Symposium (SAS) (pp. 1-6). IEEE.