Design and Experiment of Visual Model for Detecting Weld Position

A visual model for detecting the weld position is studied theoretically and experimentally. This model is developed by a BP neural network trained by the elastically gradient descending arithmetic. The weld pool images are caught by a vision sensor during the arc welding process. A location of the welding pool image is chosen as a special processed image whose characteristic corresponds with the welding pool image centroid. This special image is processed by the median filtering and the image gray transform so as to sharpen the weld character in the welding pool images. The difference, the displacement and the moving velocity of the image centroid are regarded as weld pool characteristic parameters. These parameters are applied as the input variables of the BP neural network. The offset between the weld position and the welding arc is used as the output variable of the BP neural network. This BP neural network model is trained by the elastically gradient descending arithmetic and the network weights are calculated. The established visual model based on the BP neural network has certain perceiving and adapting ability to the welding environment. The generality of the visual model is tested and the experimental results show that this visual model has the excellent generality and accuracy. The weld position can be detected accurately through the weld pool characteristic parameters.