Research on Weld Defects Detection Method Based on Laser Ultrasound and CBAM-CNN

A novel weld defect detection methodology by integrating laser ultrasonic technology with a convolutional neural network (CNN) enhanced through a convolutional block attention module (CBAM), aiming to improve the accuracy and efficiency of defect characterization in industrial non-destructive testing. The approach begins with the establishment of a finite element model using COMSOL Multiphysics to simulate laser-ultrasound interactions with distinct weld defect . The acquired one-dimensional time-domain signals are then transformed into two-dimensional time-frequency representations using continuous wavelet transform (CWT), optimizing the data for deep learning processing. A CBAM-CNN architecture is subsequently developed, leveraging channel-spatial attention mechanisms to enhance feature extraction and suppress noise interference. Experimental results demonstrate exceptional performance, with the model achieving 99.06% mean classification accuracy on pristine test data and maintaining 89.18% accuracy under different Gaussian noise conditions. The model’s robustness is further validated through t-SNE visualizations and confusion matrix analyses, confirming its ability to accurately classify porosity defects across varying sizes and noise levels. These findings highlight the method’s potential for automated quality inspection in manufacturing, particularly for high-precision weld defect detection.