Analysis Of The Dynamic Characteristics Of Metal Vapor Image During High-power Disk Laser Welding

The emergence of high-power lasers opened up a new field of laser welding and promoted the development of deep penetration welding based on the keyhole effect. As one of the key issues encountered in high-power deep penetration, laser-induced plasma phenomena has received a lot of attention lately. Over the last decade, the study mainly focused on the on-line monitoring of plasma information, the analysis of its feature, and the control of it. Although a limited degree of automation of laser welding has been achieved in a few laser-processing fields, laser welding process is a complex process with multi-variable, which is non-linear, time-varying and vulnerable to interference. These result in the difficulties in detection and the complexity of the equipment, and greatly limit the automation and promotion of laser welding.The plasma formed in the high-power disk laser welding process mainly consisted of metal vapor, and characteristics of laser-induced metal vapor image varied with the different state of laser welding. This paper achieved the transient information of the metal vapor during high-power disk laser welding process by high-speed photography, developed color space clustering segmentation algorithms based on metal vapor image to extract the characteristics, including the metal vapor area, height, intensity, swing angle and the spatters area, combined modern detection analysis means of welding samples, and analyzed the relationship between the metal vapor characteristics and the laser welding process quality. Two welded seam width prediction models based on BP and RBF were established respectively. Moreover, a support vector regression algorithm was introduced to predicting the welded seam width, establishing a welded seam width prediction model based on support vector regression. Through these studies, the main achievements are obtained as following:1. Exploration and analysis of metal vapor monitoring experimental system during the high-power disk laser welding processThe experimental system was composed of the laser processing equipment and a high-speed camera with a combination filter. A scheme for10kW bead-on-plate disk laser welding was designed. We carried out several groups of experiment at different welding speed, while other welding parameters maintained the same, and got corresponding image sequences.2. The proposing of color space clustering segmentation algorithm based on metal vapor imageMetal vapor color images contained more information than the gray images. Two clustering image segmentation algorithm based on color space were proposed in the paper, achieving the segmentation of the metal vapor image in color space directly.(1) Because the traditional clustering segmentation algorithm based on K-means was of high time complexity and poor classification performance when processing the metal vapor images, an improved segmentation algorithm based on K-means was proposed. On the one hand, introduce of the distance parameter d to enhance the cohesion of each sample within a class, in order to improve the accuracy of segmentation. On the other hand, reduce the number of samples. When processing samples, first filter the samples which were in the background region, and then classify the samples in foreground region, in order to reduce the time complexity. The experimental results showed that the improved segmentation algorithm can effectively improve the accuracy of segmentation, and decrease the time complexity.(2) However, the time complexity of the improved segmentation algorithm based on K-means was still high, and the selection of the initial cluster centers greatly impacted on the clustering results. Therefore, a probabilistic clustering method based on minimum error probability criterion was applied. It adopted Bayers model, which assigned the sample X to the cluster of the largest posterior probability. In the algorithm, the class conditional probability density function was defined as the reciprocal of the distance between sample X and the cluster center. The experimental results showed that the accuracy of segmentation of both algorithms were high, but the probability clustering method greatly decreased the time complexity.3. Extraction and analysis of the characteristics of the laser-induced metal vapor image and the welded seamCharacteristics of the laser-induced metal vapor image, including the area, height, intensity, centroid, swing angle of metal vapor and the area of spatters, were extracted. These characteristics were analyzed combined with the welding quality, including the surface characteristics of the weld seam, the bead width, the weld penetration and the depth-to-width ratio, in order to find the relationship between the laser-induced metal vapor and the welding quality. 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.4. The establishment of welded seam width prediction models based on metal vapor image features(1) The establishment of welded seam width prediction models based on artificial neural networksThe input signals in the existing models, such as welding current, temperature near the molten pool, welding speed, presented one-dimensional for the welding process, and lost the spatial distribution information of the signals, which reduced the reliability of the welding quality prediction. Welded seam width prediction models based on BP and RBF were established respectively. The input layer was a7-dimension feature vector composed of metal vapor image features, including the average area value of metal vapor and its variance, the average area value of spatters and its variance, the average swing angle value of metal vapor and its variance, and the welded seam width output of the previous sample. The output layer is the welded seam width. The experimental results showed that the model based on BP was better than the model based on RBF, and was suitable for the laser welding process.(2) The introduction of support vector regression algorithm to establish a welded seam width prediction model based on support vector regressionNeural network was instable, over reliant on the learning samples, and the numbers of samples obtained in the laser welding process were limited. The support vector machine method was introduced in the paper, establishing a welded seam width prediction model based on support vector regression. The experimental results showed that the support vector machine’s generalization ability is stronger than BP network in nonlinear regression, and the support vector machine method is suitable for high-power disk laser welding process.