Recognition Of Molten Pool Morphology In Real Time And Prediction Of Weld Appearance During High-power Disk Laser Welding

Laser welding is an important technique of bonding techniques among different metal or nonmetal materials, which has advantages of high laser power, high welding speed, good welding quality and less laser affected zones. The variation of molten pools morphology is an important phenomenon during high power laser deep penetration welding, and attracts more and more attentions. In recent years, the main research of molten pool during laser weelding was concentrated on the infrared characteristics of molten pools and the characteristics and behavior of keyhole. The morphology of the molten pool varied intensively during laser welding, and it is difficult to be directly measured, therefore the research progress of the molten pool morphology is limited.The features of molten pool morphology varies along with the welding state, and different welding states produce different features of molten pool morphology. In this paper, dynamic morphology information of the molten pools was obtained by high-speed photography technology during high power disk laser welding, and the molten pools images were segmented by the combination of the histogram piecewise linear stretch enhancement algorithm, LoG edge detection, morphological operations and operational method based on templates. The model of the reconstruction of the3D surface was siplified and the minimized optimum method was applied to get the reconstruction solution, then the results of reconstruction of the molen pools under different welding speed were analyzed. The morphology of molten pools was indirectly measured by the defined2D features of the shadow of the molten pool casted by the auxiliary diode illuminant. The extracted2D features were analyzed by statistic methods and fitting methods. Finally, the prediction model of weld appearance with the extracted2D features based on BP and RBF neural networks were established, and the BP neural model was also improved by genetic algorithm to avoid the local optimum of training weights of BP model. The main work in this study is as following:1. Exploration and analysis of molten pool morphology monitoring experimental system during the high-power disk laser welding The experimental system was composed of the laser processing equipment and a high-speed camera with a narrow band filter. A scheme for lOkW bead-on-plate disk laser welding was designed. An auxiliary diode laser light was applied to illuminate the molten pools and form their shadows. We carried out several groups of experiment at different welding speed, while other welding parameters maintained the same, and got corresponding image sequences.2. Considering the features of the obtained images, an integrated segmentation method combined by the the histogram piecewise linear stretch enhancement algorithm, LoG edge detection, morphological operations and operational method based on templates, was proposed. The molten pool and its shadow were extracted out successfully. The algorithm has the function of batch processing without human intervention, and automatically batch molten pools segmentation. It provides a foundation for automated detection of molten pool shape during laser welding.3. With the extracted molten pools from the images, the reflection model of the molten pool in our work was analyzed and the SFS (Shape from Shade) technology was applied to reconstruct the3D surface of the molten pools with the minimized optimum numerical computing methods. The morphology of molten pools under different welding speed were reconstructed and analyzed.4. The morphology of molten pools was difficult to be measured as its violent variation during hight power laser welding. The3D information of a molten pool by the reconstruction algorithm gives the morphology of the molten pool in detail, but the reconstruction algorithm need large computations and long time to get the solutions. In our work, An auxiliary diode laser light was applied to illuminate the molten pools and form their shadows, and the morphology of molten pools could be researched indirectly by the morphology of the casting shadows.In this paper, four features inlucding the area feature A of molten pools shadows, the tilt feature T, the width feature W and the length feature D were difined and analyzed with the weld appearance. With the correlation analyses method and fitting method, the analyses are as follows:(1) The variation of weld width is correlated to the variation of feature A of molten pools shadow, and the welding quality was highly affected by the variation of feature A during laser welding. The feature A, feature W and feature D can represent the volume of the molten pool. The variation of volume of molten pool is due to the variation of the energy from laser beam to welding material. The feature T gave the clues of formation of the hump and accumulation in the welds. The changes of the welding quality can be reflected by these nonlinear factors together, which reflected the stability of the welding process, and provided a new way of online monitoring of welding process.(2) The four features of the molten pools shadows were linear fitted and10times nonlinear fitted and the spans were formed by the two fitted curves. The parameter L, D and the production of L and D were defined to analyze the spans in quantitive way. It is found that the welding quality can be related with the four extracted features, and it provided a new way of online monitoring of welding process.5. The weld width prediction models based on molten pools morphology features with BP and RBF neural networks were established. With the comparision of the prediction errors, it is found that BP neural network is better than RBF neural network when the training set and validation set were large. The genetic algorithm was introduced to improve the BP neural network, in order to help the BP model which is highly depending on the initializing values get the global optimum than the local optimum of the training weights. A prediction of weld appearance which includes the weld heitht and weld width with BP neural network improved by the genetic algorithm was also established. With the validation set, the meodel can effectively predict the weld appearance under different welding speeds.