Developing a supervised machine learning model for an optimised aluminium addition based on historical data analytics, for clean steelmaking

Detta är en Master-uppsats från KTH/Materialvetenskap

Sammanfattning: De-oxidation is an important process in clean steelmaking. Al (Aluminium) is mainly used as de-oxidant and controls the final oxygen content and impact the sulphur removal in steel. Adding optimum amount of Al is critical for steel cleanliness and to reduce cost. Unfortunately, recovery of Al is not repeatable due to inherent variation in factors like amount of slag carryover, total oxygen content, tapping weight and so on. To address this challenge, statistical modeling is used to develop a supervised machine learning model to predict Al addition for secondary de-oxidation. Data analytics is used on historical data from production database to gain insights from data on secondary de-oxidation practice, observe patterns, trends and understand correlation among critical process parameters. Simple and multiple linear regression models have been developed with prediction accuracy of 58 and 66% respectively. These models have been trained, tested and cross validated using standard procedures like k-fold cross validation and grid search. To deploy multiple linear regression model into production, a Microsoft Excel based dashboard containing prediction tool, pivot charts, line, and bar graphs for analysing the process is developed. This model when tested in shadow deployment environment perform well on steel grades containing dissolved C (Carbon) up to 0.15% after tapping. In shadow deployment mode the new model can be utilised in parallel to existing tool. For %C greater than 0.15%, prediction accuracy stands at 46%. This is due to nonlinear relationship between oxygen content and added Al. With our model, in process window containing 0 to 0.15 % C after tapping in steel melt, we believe that we can in future achieve better steel quality and repeatability in de-oxidation process, improve productivity in terms of time and resources and facilitates decision making when the model is ready for use in real production environment. Future work in this direction would be to further develop this model for other steel grades.  

  HÄR KAN DU HÄMTA UPPSATSEN I FULLTEXT. (följ länken till nästa sida)