Neural networks vs. regression in localized scan correction for robotic drilling applications

Due to inherent inaccuracies in equipment and assembly variation, a method of on-the-fly correction is necessary for processes using robots for assembly purposes. This research explored the goodness of using the back-propagation feedforward neural network to perform scan correction for a robot vision system. The main goal of this research was to propose a robust procedure for comparing performance of the neural network system with the current regression analysis. This paper presents the process used to develop neural networks to calculate necessary scan corrections, train the neural networks, test and validate the networks using a threefold cross-validation process, and compare the results obtained. The results of the network development and the four current regression-based models are evaluated using hypothesis tests and prediction error statistics. In this case, the neural network results were not a significant improvement over the current regression-based analysis.