h1

TY  - THES
AU  - Ziegler, Moritz Andreas
TI  - Entwicklung eines Automatisierungskonzepts für Pfannenabschlackmaschinen; 1. Auflage
VL  - 111
PB  - RWTH Aachen University
VL  - Dissertation
CY  - Aachen
M1  - RWTH-2025-04203
SN  - 978-3-941277-54-0
T2  - Aachener Schriften zur Rohstoff- und Entsorgungstechnik
SP  - 1 Online-Ressource : Illustrationen
PY  - 2025
N1  - Druckausgabe: 2025. - Auch veröffentlicht auf dem Publikationsserver der RWTH Aachen University
N1  - Dissertation, RWTH Aachen University, 2025
AB  - The aim of deslagging in the Linz-Donawitz process for steel production is to remove metallurgical slag from the metallurgical ladle. This deslagging process is currently often mechanised by using a manually controlled slag-raking machine (PAM) to remove the slag from the surface of the ladle at a temperature of around 1500 °C. At present, the PAM is usually controlled from close to the ladle during the deslagging process and, although mechanised, is therefore an uncomfortable and physically demanding workplace. Automation and robotics methods offer far-reaching potential for making the deslagging process safer and more efficient.In this thesis, an automation concept for the PAM is developed and validated. Functional, qualitative and technical requirements for the automated deslagging process are defined as part of a requirements analysis. The automation concept developed as a result closes the chain from the sensory recording of the measurement scene via the algorithmic processing of the sensor information to the motion planning of the PAM and can simultaneously fulfil the defined requirements.A number of optical technologies commonly used in robotics are being evaluated for the sensory environment detection of the automated PAM and tested for the application. During the associated series of tests in the steelworks, all technologies based on active infrared technology (IR) proved to be unsuitable, as no artificial IR signals could be detected due to the intense radiation from the hot ladle. A purely passive stereo camera system, whose exposure time, aperture angle and base length can be customised to the application, is therefore identified as a suitable technology for three-dimensional environmental perception.The slag is detected on the basis of convolutional neural networks (CNN), which are trained using recorded measurement data from the process environment. For this purpose, the camera images from the stereo camera are annotated so that the slag can detected using image processing. The slag detection can thus be projected into three-dimensional space via the relationship between camera pixels of the stereo camera lenses and points of the point cloud generated by stereo vision. The segmentation results of several CNN architectures are compared using the intersection-over-union (IoU) value and the prediction time. The U-Net proves to be the most suitable architecture. With an average IoU value across all classes of 0.813, the U-Net delivers the best segmentation prediction on the validation dataset by a small margin and also achieves the shortest prediction time and thus the highest frame rate by some distance. Under identical hardware conditions, the second-best architecture only achieves 82.4 
LB  - PUB:(DE-HGF)11 ; PUB:(DE-HGF)3
DO  - DOI:10.18154/RWTH-2025-04203
UR  - https://publications.rwth-aachen.de/record/1010585
ER  -