Real-Time Classification of Sinter Flame Layer Using Computer Vision and CNN

Abstract

ABSTRACT

Recently, various automatic control models have been introduced into POSCO’s blast furnace process. However, there are physical limitations in maintaining stable blast furnace conditions using existing control models when the quality of key raw materials like sinter and coke deteriorates. This arises because most of the control models currently applied to the furnace aim solely to maintain favorable furnace con- ditions. Given this, there is an emerging need for a project that employs an Artificial Intelligence (AI)-based automatic control model in the sintering process, a precursor to the blast furnace process, to stabilize the quality of sinter. The sintering process involves a significant amount of unstructured data, creating numerous technical challenges for applying AI. Among them, sinter flame layer data is an essential unstructured variable that represents the ongoing thermal conditions in the sintering process. To address this, our research proposes a classification model for sinter flame layer thickness. During model development, Generative Adversarial Networks (GANs) were employed to solve issues of data set imbalance, and a De- noising Autoencoder (DAE) was used for noise removal. For image classification, a simple CNN model was utilized. Furthermore, for robust on-site application, Majority Voting techniques were employed. Ultimately, the developed classification model for sinter flame layer thickness improved the performance of existing thermal automatic control models by 10%, and is anticipated to be applicable in various other sintering automatic control systems.