Speaker
Description
The RH degassing plant is essential for producing ultra-clean steel, but its vacuum-based nature is hiding the process from close scrutiny. Despite recording the many inputs and outputs of process control and measurements surrounding the plant, the use of produced data is comparatively low, and data handling is a challenge, especially considering correlations of process control data with KPIs such as actual steel homogeneity and cleanliness after casting.
Within the scope of the EC-funded INEVITABLE project, the focus lies on improving sensor data quality and maximizing the utilization of the abundance of data obtained during the RH process. Visual surveillance offers the possibility of image analyses via conventional approaches such as surface flow observation inside the vacuum chamber and flow analysis with PIV or optical flow methods, but also machine learning can be applied, e.g., for blurriness detection or even flow pattern recognition. Time-series data from process control measurements like offgas composition or chamber pressure, even though highly connected to specific domain knowledge, can be tackled with machine learning methods as well, such as recognition of characteristic curve sections or categorization of curves. Ultimately, machine learning could be applied to correlate data on the quality of the cast steel with specific patterns occurring in the data during the cleaning and homogenizing steps of steel production.
In this talk we present examples for the application of machine learning techniques in the context of the RH treatment in steel plants and evaluate their success and relevance for improving the production of ultra-clean steels.