Development of Machine learning methods for extracting data from furnaces and IR camera to enhance tapping process

Abstract

This project was done in collaboration with Eramet Norway and USN with focus on quantifying events during the tapping of molten metal and slag. The main objective was to get data from the furnace tapping process using an IR camera, weight sensors and a level sensor to further the process understanding of the metallurgists. The individual objectives are to describe different machine learning algorithms and apply the models to detect features and quantify them using the available sensor data with the added possibility of real time processing of the image data to provide temperature information to the tapping operators.

The main methods used are image processing and machine learning using different image classification methods and regression models. The classification models could accurately detect coke and classify flaming into three categories, where the best performing models could classify flaming with a 93% accuracy and detect coke 100% of the time. While the performance of the regression models on metal/slag level estimation was unsuccessful, due to poor data quality. Slag temperature can be extracted and uploaded to Eramet database.

In conclusion gathering good quality data a higher sampling rate is vital to further development of level estimation. Image classification and slag temperature analysis can be implemented on the furnace in the current state. The method to find spitting frequency must be improved and better quality data must be collected to further develop this system