Investigation Of Suitable Methods For An Early Classification On Time Series In Radial-Axial Ring Rolling

Abstract

To increase competitiveness in the hot forming sector, there is a constant urge to improve the rolling process and its products. Industry 4.0 and its impact on data acquisition and data availability enable data driven methods for optimization. In order to optimize the quality prediction of rolled rings in Radial-Axial Ring Rolling (RARR) with regard to ovality as early as possible and hence prevent scrap and unnecessary rework, machine learning methods from the early classification on time series subdomain are used and evaluated within this research. Different approaches from the time series classification domain within supervised learning are used and compared. A so-called minimum prediction length of the ring rolling process time series is analysed using real world production data from thyssenkrupp rothe erde Germany GmbH. Building upon results of earlier research regarding the use of time series classification in RARR by FAHLE ET AL. fully automated as well as domain specific minimum prediction lengths will be investigated. The results of both approaches are compared and evaluated with regards to the current maximum prediction accuracy using the whole sequences, which should provide the highest score as it holds all available information of each sample.