Lamb-Wave-Based Multi-Defect Reconstruction via DCPSUnet Network

Abstract Ultrasonic guided waves have gained significant attention in non-destructive testing of thin plates, owing to their high efficiency in defect detection. However, factors such as mode conversion and dispersion effects can worsen signal aliasing of proximal defects, which compromises detection accuracy and reliability. Conventional physical-driven imaging techniques rely on idealized assumptions about complex physical processes. It often fails to effectively resolve such overlaps, especially in scenarios involving multi-defect reconstruction. To address this limitation, an improved U-Net framework based on dilated convolution and pixel shuffle (DCPSUnet) is proposed for anti-overlapping multi-defect reconstruction. High-fidelity training images are generated using the delay multiply and sum algorithm based on Lamb waves acquired by an ultrasonic array. A dilated convolutional pyramid is introduced into the encoder to perform multi-branch parallel extraction of image features, thereby improving the accuracy of downsampling. After 150 training epochs, the DCPSUnet model achieves a prediction accuracy of 99.2% compared to 83.7% from the classical U-Net. Defect features can be extracted and mapped to the ground truth successfully. The proposed network demonstrates superior performance in resolving overlapping signals and reconstructing multiple defects simultaneously. This approach significantly improves the reliability of non-destructive testing for thin plates by effectively mitigating signal overlapping and enhancing multi-defect reconstruction accuracy, offering a robust solution for complex defect scenarios in industrial applications.