Innovative machine learning based approach for reliable and accurate measurement of guide roll alignment in continuous casting plants

Abstract

Continuous casting (strand casting) is the leading-edge technology in modern steel production. Even slight improvements of the casting process offer a range of benefits, such as improved quality, higher productivity, energy savings and a reduced environmental impact. The precise alignment of a plants guiding rolls is a crucial factor for its performance. Strand condition monitoring systems usually incorporate modules to gauge the alignment of the guiding rolls. These modules measure the angle formed by two successive rolls by using metal rulers fitted with inclination sensors. However, the angle of the ruler only represents a valid measurement where it maintains firm contact with two successive rolls. Most of these systems rely on inductive switches that trigger a measurement at these locations. Yet this approach provides unreliable results due to the trigger’s insufficient accuracy. Furthermore, this method stores only one data point per roll, which is problematic due to significant measurement noise. Apparently, a reliable and accurate alignment measurement remains a significant challenge. As part of an ongoing research project, this paper presents the development of a novel alignment measurement module. The module will provide continuous readings from the inclination sensors without the need for a trigger. In addition, an innovative machine learning based approach to distinguish between valid and invalid measurements is developed. This method enables the averaging of multiple measurements per roll, drastically reducing the impact of measurement noise while providing reliable alignment angles. To overcome the limited availability of real-world data, which is insufficient for model training, the measurement module is simulated to generate artificial training data. Our machine learning model achieves excellent performance on this artificially generated dataset. The paper concludes by outlining how the new approach will be applied to real-world data generated by the new module