Machine Learning Support for Repetitive Tasks in Metal Processing SMEs

Abstract

In the metal processing industry, there are time-consuming repetitive tasks, e.g. checking parts if they are producible on a certain machine. In order to relieve the production manager and save time, this paper presents a self-learning system that carries out this task independently. Expert knowledge was collected, a synthetic data generator, a machine learning model based on a neuronal network for part classification as well as feedback modalities for experts were developed together with an Austrian sheet metal profile manufacturer. The solution was well accepted by the target group, but it became clear that it is important to integrate them into the whole development process. Furthermore, they can imagine that they trust the machine learning prediction after several weeks and thus the test of producibility could be automated. Tests on synthetic data showed a data collection period of approx. two years is necessary to provide satisfactory prediction accuracy if the model is trained from scratch. This time can be shortened by using pre-trained models.