STEEL TECH

Self-adaptive data-driven system for process monitoring during ladle refining and continuous casting (H2020 INEVITABLE)

Björn Glaser - Associate Professor/ Head of Unit of Processes - KTH - The Royal Institute of Technology

Real-time process monitoring in relation to steelmaking is challenging due to various physical and chemical interactions of process parameters and stochastic events. The complexity of understanding these events limits the use of conventional monitoring tools. In recent times, increasing global demand for clean steel created a need for discovering novel methods for predictive process monitoring and control. Using applied artificial intelligence, dynamic and data-driven models can be developed to help achieve robust processing by reducing production errors. Considering the advantage of rapid anomaly detection in steelmaking, the use of data-driven models is demanding. One of the applications to use data-driven process models is in the early prediction of the clogging phenomenon associated with the continuous casting of steels. Clogging of the submerged entry nozzle (SEN) during the continuous casting of steel is a major stochastic event that needs to be carefully monitored from castability perspective. Another important application is the online optimization of material additions during ladle refining. As every steelgrade is unique, developing the distinctly trained machine learning (ML) model seems challenging for deployment in existing production servers. Industrial data were collected from steel plants to design and validate the model's set rules. The collected data were examined to optimize the model parameters by continuously monitoring the evaluation of non-metallic inclusions at various steelmaking stages. The differences between predicted and actual castability were analyzed to validate and verify the newly developed models. Special efforts were taken to make the models self-adaptive in case of data drift to avoid degradation in model performance. Such model predictive results coupled with effective thermodynamic calculations can be used as a decision support system for effective process monitoring.

Share