Predicting welding parameters using Hierarchical Bayesian Models

Abstract

Programming an industrial welding robot is time-consuming work, as every movement has to be hard coded. Therefore, automation mostly takes place in the bulk production, as the robot only has to be programmed once. This is why automation in the single piece production is lagging behind. Since products are different each time, the robot also has to preform a different task each time, and thus has to be reprogrammed for each product. The Instant Robot Programming System aims to reduce this programming time from a couple of days to a couple of minutes. Even though the movement of the robot no longer has to be programmed, there is still a human operator needed to provide some additional welding parameters. The next step in automating this process even further lies in the prediction of these additional parameters. Since there is little training data available such a predictions system should be able to give the correct parameters after seeing only one of a couple of examples. This thesis tries to solve this problem using a Hierarchical Bayesian Model, which is a very powerful tool for one-shot learning.