Attention-enhanced semi-supervised graph neural network for binary imaging of subsurface defects with laser ultrasonics

This paper proposes a deep learning-based ultrasonic imaging method specifically designed to classify signals located at defect boundaries, enabling accurate reconstruction of C-scan images of subsurface defects. The IMNS-GAAN algorithm, which integrates graph attention network within the generative adversarial network framework, incorporates imbalanced and negative-sampling graph attention adversarial modules, effectively addressing signal misclassification within defect areas due to imbalanced dataset and enhancing the accuracy of signal classification at defect boundaries. Laser ultrasonic experiments performed on artificial subsurface defects fabricated from additively manufactured AlSi10Mg and TC4 specimens demonstrate the efficiency of the proposed IMNS-GAAN algorithm in accurately reconstructing C-scan images. Compared to traditional amplitude imaging in time and frequency domains, the IMNS-GAAN algorithm depicts defect boundaries more distinctly than other three graph neural network-based methods: GraphSAGE-MEAN, GraphSAGE-LSTM, and GAT, achieving a defect size measurement error of 7% and an image similarity score of 98.46%. These results demonstrate the effectiveness and accuracy of the proposed IMNS-GAAN algorithm in precisely imaging subsurface defects under the laser-ultrasonic testing conditions with multi-mode coupling.