Prediction of variable-groove weld penetration using texture features of infrared thermal images and machine learning methods

When workpieces with varying grooves are welded, changes in the groove angle of the base metal can affect the butt weld penetration, thus affecting the final weld forming quality. Therefore, it is particularly important to be able to diagnose variable-groove weld penetration in real time. In this study, a novel variable-groove weld-penetration diagnosis technique using the texture features of infrared thermal images and machine learning methods is proposed. First, an improved local binary pattern algorithm based on the window standard deviation is introduced for texture analysis of welding-temperature field images. Second, the feasibility of using the images' texture features for predicting variable-groove weld penetration is demonstrated. Finally, a predictive machine-learning-based model for predicting variable-groove weld penetration is developed, and the effects of the size and location of the monitored area on the model's performance are quantified. The experimental results suggest that the proposed predictive model is feasible and offers a new and effective method for real-time diagnosis of variable-groove weld penetration. The research results of this paper enrich the application of optical inspection in the field of welding processing.