Research Models And Methods Of Identifying Welds Under Strong Arc Environment

With the development of the machine vision technology of computer and the requirement of development of automatic welding, the technology of weld seam image processing and automatic detecting based on the machine vision becomes a very important field of automatic welding tracking system. However, during the actual welding process it is very difficult to acquire clear and stable image of the near seam area due to the disturbance of noise such as arc light, high temperature, workpiece deformation, image transformed error, etc. In order to get appropriate image, it should apply a special filter system, and collect weld image by a image collective card and save them in PC, then get the right seam position through image pre-processing and automatic detection. Seam tracking technology is one of the important research issues in the field of the automation of arc welding process, and accurate seam tracking can improve the weld quality. The technology of seam deviation detection, which is the deviation between the seam center and the arc, is one of the key to realize accurate seam tracking.A vision sensing automatic welding tracking system was designed and setup for this experiment. In order to track the weld seam automatically, the system was composed of PC and other parts including vision detecting system without light source, etc. The traditional method of seam tracking based on machine vision is to catch weld image real time by CCD (Charge Couple Device) vision, and the information between the weld pool center and the seam has been acquired through image processing, whose dispersion reflects the deviation between the arc and the seam at the moment. The procedureof image filter, image strength, image binary processing, image edge detection and position acquisition based on the image processing were analyzed by mathematics,and simulated in system. Also, This paper discussed the binary processing process and the edge detection methods, provided the description of the basic principle and procedure for the operators such as Robert operator, Sobel operator, Prewitt operator, Log operator and Canny operator. Also the characteristics of these operators were introduced.Through the study of method of the traditional seam position detection, a new method called image processing based on the centroid of image was studied and a new technology for obtaining the seam tracking error was presented. The thesis, which is different from the traditional method getting the seam deviation information only by image processing, chooses a weld pool process region which includes the foreside of weld pool and the seam in front of the pool. Then the centroid of weld pool image is made as the characteristic parameter to analyze the seam deviation. And the new method is researched on how to set up the seam deviation measurement visual model by the pool characteristic parameters.To solve problems such as the unclear images caused by strong electric arc interference, a part and multiscale statistical method was proposed to obtain high-dimensional feature space that represents the deviation status during welding process. Based on this, an online prediction model was established. Firstly, the images were categorized according to the zones, namely the electric arc zone, the fusion zone and the weld seam zone. Secondly, both the horizontal and vertical coordinates of the categorized images were reconstructed. Principal Component Analysis (PCA) was used to realize control over the dimensions of high-dimensional feature space, based on which an input vector of the online prediction modelwas calculated. Data of strong interference, electric arc distortion, focal deviation, aperture reflection, sensor movement was collected and samples of model training and testing were obtained. Experimental results show that the adoption of BP network yields higher prediction accuracy than that of the Elman network. Especially, the model that employs multiscale feature space performs much better than the model using single scale feature space.Research findings also suggest that the part and multiscale statistical feature BP network prediction model is effective in preventing the possible interference of image pre-processing on weld seam deviation measurement. Especially, the proposed model has strong recognition ability under the interference of strong electric arc. It can also realize accurate identification of different welding deviation statuses without measuring the geometrical features of the objects to be identified. The part statistical method helps to obtain welding deviation status indirectly. Experimental results prove that the suggested method is effective in building up an online prediction model with high stability and strong generalization.