Defect monitoring of high-power laser-arc hybrid welding process based on an improved channel attention convolutional neural network

High-power laser-arc hybrid welding (HLAHW) has emerged as a prominent method for manufacturing medium-thick-walled steel components in modern industrial manufacturing. The process is prone to instability and defects because the high-energy–density heat source of HLAHW has a strong interaction with the material. Monitoring the welding process is beneficial for adjusting the process parameters in time and is a valuable guide to reducing the occurrence of welding defects. This study proposes an improved convolutional neural network (CNN) with the separable channel attention mechanism for the defect monitoring of the HLAHW process. First, a top vision platform is used to acquire the images of the HLAHW process. Next, in order to improve the monitoring accuracy of the top vision HLAHW process, a mixed pooling attention mechanism (MPAM) module is designed to calibrate HLAHW feature maps adaptively. Then, several modules are embedded in the CNN, named mixed pooling attention mechanism network (MPAMnet), to focus on features in different stages of the network. Four new welding experiments are used to test the performance of the network. The experimental results reveal that the MPAMnet outperforms other popular CNNs, achieving the highest accuracy of 95.02% on the test set, which includes well-formed welds, incomplete penetration, root humps, and surface collapse, with a processing time of 2.6 ms per image.