Abstract

In the steel making process, vessels and containers need to be protected from the extreme temperatures reached during the production. Thus, these metallurgical vessel are lined with bricks. Over time this lining wears down and needs to be replaced. It is vital to recognise high abrasion to avoid a breakage of the vessel which could be devastating during steel production. Costly laser measurements can be made to determine the state of the remaining bricks. However, for some vessels this is not possible as the environment is hostile and they cannot be opened to correctly conduct these measurements. With the purpose of offering a more cost-effective alternative, this thesis aims to use measurements made during the process to predict the wear of the refractory lining. Since, multiple statistical approaches have been tested, the focus lies specifically on neural networks to improve the predictions. A myriad of neural network architectures were assessed with the aim of finding the best performing model. Multiple data sets with diverse measurements from different manufacturers were evaluated. The findings show that this approach improves the performance on most of the data sets while providing competitive results for all of them which suggest that predicting the lining of metallurgical vessels to be a viable alternative. A 2-dimensional convolutional neural network was used to find patterns in the data. The results improved after applying pre-processing steps like a Gaussian filter to smooth the data and by selecting the most impactful features using SHAP values.