Improvement of Machine Learning Models for Time Series Forecasting in Radial-Axial Ring Rolling through Transfer Learning - Abstract

Aug 6, 2026

Upload by Patrick

Machine learning for production technology is increasingly adopted in the Industry 4.0 context, where sufficient and high-quality data is essential. When data quantity or quality is limited, data augmentation, synthetic data, and transfer learning across related datasets can mitigate the gap. This work applies transfer learning to radial-axial ring rolling by predicting the outer diameter over process time between a source mill and a target mill with unequal datasets.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages2

Reading Time5 min

AccessFree to download

Abstract

Transfer learning setup between source and target rolling mills

RNN/LSTM model adaptation for full-process prediction

Evaluation using MSE and standard deviation across hyperparameters

Analysis of prediction behavior and causes of observed anomalies

Improvement of Machine Learning Models for Time Series Forecasting in Radial-Axial Ring Rolling through Transfer Learning SFU 2023 Johannes Seitz 1* , Qinwen Wang2 , Tonias Moser 1 , Alexander Brosius2 , Bernd Kuhlenkötter1 1)Chair of Production Systems, Ruhr-University Bochum, Industriestraße 38c, 44894 Bochum 2) Institute of Manufacturing Technology, Chair of Forming and Machining Processes, Dresden University of Technology , Helmholtzstr. 10, 01062 Dresden *Corresponding author: Johannes Seitz, [seitz@lps.ruhr-uni-bochum.de](seitz@lps.ruhr-uni-bochum.de) Abstract Due to the increasing computing power and corresponding algorithms, the use of machine learning (ML) in production technology has risen sharply in the age of Industry 4.0 [1] . Data availability in particular is fundamental at this point and a prerequisite for the successful implementation of a ML application. If the quantity or quality of data is insufficient for a given problem, techniques such as data augmentation, the use of synthetic data and transfer learning of similar data sets can provide a remedy. In this paper, the concept of transfer learning is applied in the field of radial-axial ring rolling (rarr) and implemented using the example of time series prediction of the outer diameter over the process time. Radial-axial ring rolling is a hot forming process and is used for seamless ring production. In this paper, the transfer between two different rolling mills with unequal data availability is analysed. The rolling mill with a comparatively large database (1300 rolling cycles) is from thyssenkrupp rothe erde GmbH (source space ) and the rolling mill with the small database (60 rolling cycles) is affiliated to the Ruhr-University Bochum (target space ) . This rolling mill is smaller in size and can apply fewer forming forces. An already existing model according to SEITZ ETAL. [2], which represents the first 50% of the rolling process, is taken up and modified in such a way that a prediction of the outer diameter of the entire rolling process is possible. The underlying algorithm is based on recurrent neural networks and is implemented using an LSTM (Long Short-Term Memory) architecture. In order to achieve an extension to the entire rolling process, it is necessary to discard the time scaling used in [2], the time series sampling set to a fixed number of time steps, and to choose a new approach. The selected input dimension now runs each index sequentially step-by-step for all rolling ids and thus takes into account different rolling speeds and lengths . This makes it easier to compare the characteristics of the rolling processes. In addition, the concept of a hybrid database is taken into account, which transfers the training data of the target space into the training data of the source space, so the corresponding rolling strategies of the target are already taken into account in the source model. In addition to the classic source models, model generalization methods are also applied, on the one hand by adding dropout layers and on the other hand by modifying the underlying data split in the hybrid database. For each hyperparameter configuration, five models are trained and the mean value of the mean square deviation (mse) is used as the final evaluation. In addition, the standard deviation of the five models between each other is determined. The base model, which is trained within the data of the target space , shows an mse of 1. 109*10-3 . Transfer learning results in an improvement of 61% of the mse and a reduction of the standard deviation of 45% . Through generalising the source model, an equal improvement is achieved , but consists of a reduced standard deviation of 81% compared to the base model. For this reason, generalisation is recommended. Using the hybrid database does not lead to the best transfer results, but a positive transfer can be observed for all hyperparameter configurations. When considering the source database for

FAQs

Why is transfer learning considered for radial-axial ring rolling time-series prediction?

Data availability and quality can differ substantially between rolling mills, which limits model performance. Transfer learning uses knowledge from a source dataset to improve prediction in the target dataset.

How is the original model adapted to predict the entire rolling process?

The approach discards time scaling and fixed-step sampling, and instead feeds indices sequentially step-by-step for all rolling IDs. This accommodates different rolling speeds and lengths for more comparable process characteristics.

What improvements are reported compared with a base model trained only on the target space?

The base model yields an MSE of 1.109*10^-3. Transfer learning improves MSE by 61% and reduces the standard deviation by 45%, while generalizing the source model achieves the same MSE improvement with an 81% lower standard deviation.

Aug 6, 2026

Upload by Patrick