Recognition Of Penetration Based On Vision Sensing During Arc Welding

TIG Welding Penetration Control is a significant research field of automatic arc welding, and the degree of welding penetration is a pivotal performance index of welding quality control. The main factors affecting welding penetration degree are welding materials, welding current, welding speed and other welding technique parameters, and the direct control factors are welding current and welding speed. Identification detection method of welding penetration state consists of direct detection method and indirect detection method. In order to detect, the direct detection method needs to cut open after finishing welding, and it is a kind of irreversible destructive detection, thus it can not be regarded as the regular detection method in practice. There are a variety of indirect detection methods, and what is described in this dissertation is based on visual sensing, and it is to study the relationship model between the characteristic parameters of front molten pool and welding penetration, and observe these characteristic parameters, to infer the penetration state indirectly.The general method of visual-sensing penetration state monitoring is to use CCD sensor to collect real-time welding pool image, and then use image processing to extract the characteristic parameters of the welding pool, and finally indirectly infer the penetration state according to the relationship model which has been built up between these characteristic parameters and penetration state. This thesis is focused on the extraction of welding pool characteristic parameters and modeling of the relationship between these characteristic parameters and penetration state, and the design of a validate classifier, which is used to identify penetration state corresponding to different characteristic parameters.In the first place, we have design a arc welding testing system based on visual sensing, and perform the research on this platform. The system is composed of a table, CCD sensor, stepper drivers, control system and other devices. Firstly we discuss method of image acquisition under strong arc welding, with the help of filtering system. Since the image obtained through this method high reflection of light, we can develop a kind of image filtering techniques combining blank domain filtering with frequency domain filtering, to remove the noise during the extraction of welding pool characteristic. Besides, make use of the techniques of image filtering to eliminate image noises and extract the characteristic parameters of the welding pool. While welding, during the transition process from incomplete penetration to complete penetration, surface of the welding pool changes intensely. Welding pool width can be divided into internal welding width and external welding width. During the changing process of the welding pool surface, on the inter edge and outer edge of the welding pool, there exists phenomenon of liquid metal mutual reflection and specular reflection and lights reflected into the camera from different part of the welding pool are not the same. Therefore, according to the specular reflection of the welding pool, we can identify and calculate the internal edge and the external edge.Finally, we use TIG overlay welding run-time image to extract geometric parameters of the welding pool, and take three characteristic parameters, including external welding width, external welding width and the deviation value of external welding width between adjacent images, as the input of the neural network, and three states, including incomplete penetration, complete penetration and over penetration, as the output. Then, RBF and BP neural network can be built up respectively. And the penetration degree can be predicted according to the neural network. Comparing the actual penetration degree and the prediction, we can analyze the accuracy rate of the penetration classification identification from different neural networks.