Radiographic detection of welding defects based on neural ordinary differential equations

Aiming at the problems of strong subjectivity, high work load, repetitive tasks, and low efficiency in defects evaluation in the process of radiographic testing (RT), a neural memory ordinary differential equations based residual network model was proposed in this paper, for the objective, accurate, and intelligent classification of weld defects. The data set of defect images containing 7 types of defects including crack, lack of fusion, incomplete penetration, concave, undercut, slag inclusion, and porosity was collected to ensure the diversity of defects, and the corresponding image preprocessing and expansion was carried out. The typical artificial neural networks, ResNet18, ResNet34, ResNet50, and ResNet101, were iteratively trained. The ResNet34 model got the highest accuracy and was selected as the backbone network to avoid under-learning or overfitting. Then, the nmODE-ResNet was built to improve the classification performance by exploring the excellent nonlinear mapping ability of the nmODE module. The experimental results show that, compared to ResNet34, nmODE-ResNet can significantly improve the accuracy of defect classification in the seven-classification task on RT images of welds’ defects by 1.56% without increasing the number of parameters, and the performance is comparable to that of qualified inspectors.