Transfer Learning Approaches In The Domain Of Radial-Axial Ring Rolling For Machine Learning Applications

Abstract

Due to increased data accessibility, data-centric approaches, such as machine learning, are getting more represented in the forming industry to improve resource efficiency and to optimise processes. Prior research shows, that a classification of the roundness of shaped rings, using machine learning algorithms, is applicable to radial-axial ring rolling. The accuracy of these predictions nowadays is still limited by the amount and quality of the data. Therefore, this paper will focus on how to make the best use of the limited amount of data, using transfer learning approaches. Since acquiring data for homogenised databases is time, energy and resource consuming, logged data gathered by the industry is often used in research. This paper takes both, industrial data from thyssenkrupp rothe erde Germany GmbH and a smaller dataset of an inhouse research plant, into account. Additionally, a synthetic dataset, created by generative adversarial networks, is considered. To accomplish an improvement of machine learning predictions using accessible data, three transfer learning approaches are investigated in order to extend existing models: (I) transferring from a radial-axial ring rolling mill to a different mill containing less available data with a ratio of 20:1, (II) learning from unlabelled data using an autoencoder and (III) training on synthetic data. The obtained improvements are further evaluated. Based on these results, future possible investigations are elaborated, in particular the consideration of transfer learning from the less complex cold ring rolling process.