A Study of Vision-Based Measurement of Weld Joint Shape Incorporating the Neural Network

Abstract

Visual sensing of the surface geometry is often necessary to inspect and evaluate the quality of welded joints as well as to sense the transient distortion of a structure during welding for the feedback of its current geometry. This investigation presents a simple and non-contact digitization method of the vision-based system for measuring the three-dimensional surface geometry of the object distorted by welding. Its basic principles are based on the equation derived from the geometric optics, for which the illumination of the laser beam was controlled in the form of the projected plane. This method utilized a 10 mW He-Ne laser for the structured light and a charge coupled device (CCD) camera as the vision sensor. When the laser stripe is projected on to the weldment, a minute deviation from the perfect plane existing on the specimen surface causes a distortion of the stripe. The shape and amount of the weldment distortion can be then calculated by analysing the distorted laser stripe. In this study, a neural network was proposed and implemented for recognizing the laser stripe features from the image plane. A calibration scheme of corresponding an image to the world position was also adopted for determining the sectional features of the welding distortion. The feasibility of determining the welding distortion by the proposed vision-based system was demonstrated through the experiments with various types of specimen.