Learning the Automated Setup of Profile Wrapping Lines for New Products from Few Past Setups

Abstract

This study investigates the feasibility of automated setup of profile wrapping processes on new products using machine learning on past setup examples. The task is characterized by high complexity of the considered production system in combination with highly varying products and a very small available database. This database also reveals ambiguous ground truth due to human, unsystematic preferences. A simple geometric-physical motivated preprocessing is proposed. On the resulting data, a Deep Convolutional Neural Network in the form of an autoencoder is shown to be very suitable for predicting wrapping actions for new products. The good but improvable results are discussed extensively with respect to the technological background and possible solutions are proposed.