Predicting the weld width from high-speed successive images of the weld zone using different machine learning algorithms during laser welding.

Abstract

The dynamic behavior of the keyhole and molten pool is associated with the quality of weld seam. In this study, an on-line visual monitoring system is devised to photograph the keyhole and molten pool during external magnetic field assisted laser welding on the AISI 2205 duplex stainless steel plates. Seven features are defined to describe the morphology of the keyhole and molten pool. Then, the principal component analysis (PCA) algorithm is applied to reduce the dimensions of these features to obtain different number of principal components (PCs). Three different machine learning algorithms, i.e. the back propagation neural network (BPNN), the radial based function neural network (RBFNN) and the support vector regression (SVR), are utilized to fit the relationship between the chosen PCs and the weld width. Finally, the global and local prediction accuracy of these three machine learning algorithms are compared under different number of PCs. Results illustrated that data dimensionality reduction is helpful to improve the modeling efficiency. Machine learning algorithms can be exploited to predict the weld quality during laser welding with high accuracy. Among them, the BPNN model performs best and SVR model performs better than RBFNN model in this research. This work aims to model the relation between the features in weld zone and the weld quality with different machine learning algorithms, and provides a guideline of model selection for laser welding on-line monitoring and a necessary foundation for realizing intelligent welding with advanced algorithm.