Predictive analysis and feature extraction of weld penetration in P-GMAW based on interpretable deep learning

Abstract

Deep learning technologies have been widely used to predict and control weld penetration based on visual sensing in GMAW. However, the black-box nature of these models makes it difficult to analyze the correlation between visual information of weld pool and penetration. This study investigates the correlation between surface wave behavior of weld pool and back side penetration using an interpretable deep learning method based on ResNet18 + Grad-CAM. Reflected laser streak images of P-GMAW weld pool captured by laser vision are used for this purpose. The research findings indicate that the dynamic features of reflected laser streak images can reveal surface wave texture information, including oscillation patterns and intensity, as well as the amount of weld deposition in the GMAW weld pool. This morphology of the laser streak can reflect the oscillation patterns and intensity the surface of weld pool. Meanwhile, positional information reveals the characteristics of weld deposition on the surface. The presence of the second free surface at the bottom of weld pool does not significantly affect the oscillation pattern of weld pool. Nevertheless, it causes changes in the intensity of the surface oscillation and the volume of the weld deposit. The ResNet18 convolutional neural network model that considers both the underlying information and the fluctuating texture feature information outperforms the neural network model that only considers the fluctuating texture feature information in terms of accuracy and generalization ability.