Learning and online prediction of weld quality in robotic GMAW

Abstract

We have come a long way in the automation of robotic welding. But with the decline in qualified welders, there has been a steadily growing demand for robotic solutions with capabilities that match that of human welders. Among these capabilities, a major missing element is the ability of the robotic systems to extract actionable information about the quality of their welds, in real time. While this is a challenging problem, we have been able to show that we can extract this information through-the-arc that captures the controlled dynamics leading to defect occurrence and at the same time identify the variation in the process leading to that defect. Our methodology involves a unique approach to gathering synchronized measurements of the welding process mapped to an objective evaluation of the resulting weld bead at a resolution of 1 mm. Demonstrating a system that can achieve this was at the heart of this work, after which we gathered measurements on voltage, current, weld bead profile, visual images, welding torch pose and radiographic images. These measurements were gathered under experiments involving induced porosity and geometric defects under process variations such as the presence of primer and oil, loss of shielding gas, erratic wire feeding, out of joint and the presence of gap between plates. From these measurements, we created a dataset and identified discriminative features that are indicative of weld quality deviations from process variations. Particularly, we proposed a specialized feature extraction methodology, "Bag-of-Pulses" as informed by the domain expertise knowledge based on "Bag-of-Words" approach. In this feature extraction methodology, we were able to represent the data concisely with new building blocks that retain the characteristic nature of the signal source (i.e., pulsed GMAW) such that we can feed them into a classification algorithm to identify defects and make diagnosis.