Penetration-state recognition in magnetic field-assisted molten pool oscillation based on adaptive variational mode decomposition of arc voltage and hybrid deep learning

Abstract

Alternating cusp-shaped magnetic field, which can be used to effectively control welding quality, can significantly enhance the regular molten-pool oscillation signal during tungsten inert gas (TIG) welding; however, the nonlinear arc voltage signal causes the accuracy of penetration-state recognition to be very low. A novel method of performing penetration-state recognition that utilizes magnetic field-assisted molten-pool oscillation based on adaptive variational mode decomposition (VMD) of the arc voltage and hybrid deep learning is proposed in this paper. A subtractive averaging-based optimizer (SABO)-VMD algorithm was selected to preprocess the arc voltage signals, in which the adaptive bandwidth optimization mechanism can dynamically adjust the parameters according to the signal characteristics to achieve the global optimal solution, thereby enhancing the quality of the signal decomposition. After the hyperparameters of the convolutional neural network (CNN) and support vector machine (SVM) were optimized by the rime optimization algorithm (RIME) and the grid search algorithm, respectively, the CNN-SVM classification algorithm was constructed by combining the powerful feature-extraction capabilities of the CNN and the efficient classification performance of the SVM. The nonlinear components of the arc voltage signal were separated by the SABO-VMD algorithm to obtain multiple intrinsic mode functions (IMFs) with different frequencies and amplitudes; this was done so that the eigenvector of the molten-pool penetration state could be extracted from the IMF with the lowest envelope entropy. Then, the CNN-SVM classification algorithm was used to recognize the penetration-state. The results show that the proposed method is robust and that its recognition accuracy can reach 95 % for various welding speeds.