Recent projections have the robotic welding industry growing by 60% over the next five years. At the same time AWS estimates there will be a shortage of approximately 290,000 welders in the US by 2020. Robot welders work with high precision, e.g., typical positional repeatability is 0.02 mm, but unlike human welders these robots don’t know a good weld from a bad weld, at least not yet. Also, we’d like them to be able to determine what is good and what is bad while they are making the weld. The key to online weld quality is high quality sensor data and machine learning algorithms. In this paper we present some of our work on this problem and show, at least for the welding conditions used in our experiments, that it is possible to determine the quality of a weld with high confidence.
There are a large number of machine learning techniques. Those that are used for classification (which is what we are doing by identifying good and bad welds) can be roughly separated into supervised (with labeled classification), and unsupervised learning (those without labels), and a third category, semi-supervised learning (SSL), that works to integrate datasets containing both labeled and unlabeled data. In the work done here segments of welds were labeled as either good or bad based upon post-weld evaluation, i.e., x-ray and surface profile results.
The experimental setup for this work is focused on GMAW-P (Pulsed GMAW) due to its ability to adjust to a variety of welding conditions, e.g., out of position, material thickness, etc. The setup included a 6DOF robot, voltage and current sensors, imaging camera, laser line scanner, and a data acquisition computer. Tee-joint fillet welds were made in both the flat [1F] and horizontal position [2F]. Welds were done with clean surfaces and with primer (used to protect weld surfaces from rust) applied at different thicknesses. Voltage and current data were collected, as well as pose data from the robot controller. Video of the weld pool was also collected during welding. All of these data were synchronized using a 1 kHz pulse train that was integrated into each of the data sets. Once the weld was completed, the robot would retrace its trajectory using the line scanner to capture the profile of the fillet weld produced. Photographs of each weld were taken and each of the welds was subsequently sent out for x-ray.
Support Vector Machines is one method of machine learning that maps the collected data into another space, typically of a higher order, in order to separate the data into the labeled classes (e.g., good and bad). Once the model (mapping) has been developed for the training data, it is tested on unseen data to evaluate its performance.
The data collected from the set of welds were analyzed for both surface and sub-surface porosity, and these results were then used to classify the welds as either good welds or bad welds. All of the welds done on clean plates were good as were some of the welds done with primer (typically small thicknesses of primer did not result in porosity). In contrast, weld plates with primer thicker than about 2 mm typically result in both surface and sub-surface porosity.