#03 Process
A data-driven approach lies at the heart of digitisation, aimed at employing innovative digital technologies, such as machine learning and artificial intelligence, to extract knowledge from data for process optimisation purposes. We apply data-driven predictive models to several steps along the production process, from through-process optimisation, such as checking the temperature of steel or determining the optimal recipe for ferroalloy additions, to predicting critical events. Human-machine integration is a key aspect of our solutions. This provides process supervision or guidance operators with an efficient decision-making or even decision-automation support.
Advanced process guidance and control solutions for your data-driven intelligent plant
Danieli Intelligent Plant at ABS QWR 4.0
Case
History
#01
Who
A plant that recycles high-quality scrap metal to form SBQ steel bars used primarily by the automotive, heavy truck, agricultural and construction equipment, oil & gas and energy markets. The process starts with melting of the scrap in an electric arc furnace, followed by refining, degassing, casting, rolling and finishing.
Why
The customer needed to reduce the loss of productivity and energy efficiency in the meltshop caused by:
- recirculation of the steel ladle due to arrival at the casting machine at an insufficient temperature (“cold” heats);
- casting at lower speed due to the steel ladle arriving at temperatures that are too hot (“hot” heats).
How
We built a machine learning model that predicts the optimal temperature at ladle furnace exit in order to ensure its arrival at the casting machine at the target temperature, after processing at the vacuum degasser. As input data, we used heat characteristics, such as steel grade and casting sequence position, steel ladle life information, temperature and chemical composition samples, raw material additions, process and transfer times.
The model is based on an initial dataset of approximately 5000 heats and is re-trained on a daily basisdaily to improve steel grade historical information. It is made available to pulpit operators, including in “What-if” mode, enabling simulation of the arrival temperature at the caster for a given temperature at ladle furnace exit and a given processing time at vacuum degassing. It is also made available to process technologists for the monitoring and adjustment of operating practices. The input variables with the greatest impact on the performance of the model include the target duration for vacuum degasser process steps, especially pump-down and degassing. Accurate configuration of operating practices has therefore been highlighted as mandatory for successful implementation.
What
So far, the model has performed with an average forecast error of ± 6.4 °C. The number of “cold” heats has been clearly reduced while the “hot” ones are progressively decreasing as far as crews get confident with the solution, avoiding to request higher than necessary arrival temperatures.
Who
A plant that recycles high-quality scrap metal to form SBQ steel bars used primarily by the automotive, heavy truck, agricultural and construction equipment, oil & gas and energy markets. The process starts with melting of the scrap in an electric arc furnace, followed by refining, degassing, casting, rolling and finishing.
Why
The customer needed to reduce the loss of productivity and energy efficiency in the meltshop caused by:
- recirculation of the steel ladle due to arrival at the casting machine at an insufficient temperature (“cold” heats);
- casting at lower speed due to the steel ladle arriving at temperatures that are too hot (“hot” heats).
How
We built a machine learning model that predicts the optimal temperature at ladle furnace exit in order to ensure its arrival at the casting machine at the target temperature, after processing at the vacuum degasser. As input data, we used heat characteristics, such as steel grade and casting sequence position, steel ladle life information, temperature and chemical composition samples, raw material additions, process and transfer times.
The model is based on an initial dataset of approximately 5000 heats and is re-trained on a daily basisdaily to improve steel grade historical information. It is made available to pulpit operators, including in “What-if” mode, enabling simulation of the arrival temperature at the caster for a given temperature at ladle furnace exit and a given processing time at vacuum degassing. It is also made available to process technologists for the monitoring and adjustment of operating practices. The input variables with the greatest impact on the performance of the model include the target duration for vacuum degasser process steps, especially pump-down and degassing. Accurate configuration of operating practices has therefore been highlighted as mandatory for successful implementation.
What
So far, the model has performed with an average forecast error of ± 6.4 °C. The number of “cold” heats has been clearly reduced while the “hot” ones are progressively decreasing as far as crews get confident with the solution, avoiding to request higher than necessary arrival temperatures.
