As a reliable technology for material joint processing, laser-MIG hybrid welding (MIG welding, melt inert-gas welding) has been applied to various fields of the product manufacturing industry for decades. Due to its characteristics such as deep penetration, high welding speed, and high-quality shaping, laser-MIG hybrid welding has become the research focus. However, all kinds of defects troubling many scholars often occur in laser-MIG hybrid welding, and root hump is one of the common defects. Unlike instantaneous defects such as undercut and non-penetration, root hump defects are caused by the accumulation of molten metal flowing to the end of the pool over a a period of time. During the formation of the root hump, the weld quality is continuously affected by it. When the molten metal has solidified to form a hump, the new molten metal will continue to accumulate in the next position to form a new hump, resulting in the periodic occurrence of the root hump within a certain range. This study presents an online detection of root hump based on invariable moment characteristics of the tail molten pool, which can detect accurately root hump defect in the strong noise environment of laser-MIG hybrid welding. We hope that our innovative approach could provide the basis for the online detection of defects in laser-MIG hybrid welding.

The laser-MIG hybrid welding process detection system is established by a high-speed camera, six-axis robot, arc welding machine, high-power fiber laser, and image processing computer. During laser-MIG hybrid welding, the images of the molten pool outlines are collected by the high-speed camera. To reduce the gray difference between the two sides of the molten pool when the arc is retracted or released, the multi-scale Retinex (MSR) enhancement method based on Retinex theory is used. After threshold segmentation and morphological processing, the binary images of the tail molten pool are obtained. Whereafter, the four kinds of invariant moments of the tail molten pool images are calculated. For suppressing the interference of local noise caused by random error on the tail molten pool invariant moments, the moving average method is adopted to reduce the influence of noise. The one-dimensional convolution neural network model using the improved dynamic learning rate algorithm is established, and the moving average values of the four normalized invariant moments from the tail molten pool images are used as input. The model is successful to realize the online detection of hump defects at the root of the weldment based on images of the weldment surface during laser-MIG hybrid welding.

According to the comparison of the moving average values of the four normalized invariant moments from the tail molten pool images between root hump and full penetration samples, the moving average values of root hump samples are higher than those of full penetration samples. The values of the root hump are almost higher than the specific moving average value, and the full penetration is lower than it (Fig.5). The occurrence of the root hump defect in the welding process can be preliminarily judged by the moving average values of the invariant moment. To accurately detect the root hump defects in the laser-MIG hybrid welding process, the one-dimensional convolution neural network model using the improved dynamic learning rate algorithm is established. The best accuracy of training set from training samples is 99.73%, and the best accuracy of the validation set from training samples even reaches 99.88% (Fig.8). A continuous weld bead, whose the first half of the weld bead has root hump and the second half is normal, is used to verify the reliability of the model. The samples are detected as root hump defect samples in the first 3604.5 ms. The false detection occurs in 2750-2900 ms. The reason for false detection is that this position is close to the boundary between the root hump area and no defect area. At this time, the moving average values of invariant moment decrease. In the latter part, the detection result alternates between 0 and 1 in 3950-4100 ms (Fig.9). A weak hump on the back of the weld bead leads to this false detection. Although the model has some detection errors, it can still accurately detect most root hump defects with 94.7% accuracy (Table 2).

This study adopts invariable moment characteristics of the tail molten pool to detect root hump in laser-MIG hybrid welding. Aiming at the problem of uneven illumination on both sides of the molten pool, the MSR enhancement method based on Retinex theory is adopted to reduce the gray difference on both sides of the molten pool. The moving average values of the four normalized invariant moments from the tail molten pool images coming from the image process can be used to judge the occurrence of the root hump defects. It is observed that the moving average values of the root hump samples are higher than those of the full penetration samples. A one-dimensional convolution neural network model with an improved dynamic adjusting learning rate algorithm is established to detect the root hump defects. The experimental result shows that the accuracies of the training set and the verification set can reach 99.73% and 99.88% respectively. The model is applied to detect root hump defects in continuous weld bead, whose accuracy reaches 94.7%. The root hump defects in laser-MIG hybrid welding are detected accurately, which provids a new idea for the realization of welding status and welding quality detection in laser-MIG hybrid welding.