Visual Sensing and Defect Detection of Gas Tungsten Arc Welding

Weld imperfections or defects such as incomplete penetration and lack of fusion are critical issues that affect the integration of welding components. The molten weld pool geometry is the major source of information related to the formation of these defects. In this dissertation, a new visual sensing system has been designed and set up to obtain weld pool images during GTAW. The weld pool dynamical behavior can be monitored using both active and passive vision method with the interference of arc light in the image significantly reduced through the narrow band pass filter and laser based auxiliary light source.Computer vision algorithms based on passive vision images were developed to measure the 3D weld pool surface geometry in real time. Specifically, a new method based on the reversed electrode image (REI) was developed to calculate weld pool surface height in real time. Meanwhile, the 2D weld pool boundary was extracted with landmarks detection algorithms. The method was verified with bead-on-plate and butt-joint welding experiments.Supervised machine learning was used to develop the capability to predict, in real-time, the incomplete penetration on thin SS304 plate with the key features extracted from weld pool images. An integrated self-adaptive close loop control system consisting the non-contact visual sensor, machine learning based defect predictor, and welding power source was developed for real-time welding penetration control for bead on plate welding. Moreover, the data driven methods were first applied to detect incomplete penetration and LOF in multi-pass U groove welding. New features extracted from reversed electrode image played the most important role to predict these defects. Finally, real time welding experiments were conducted to verify the feasibility of the developed models.