Multisource Ensemble Network-Based Learning for Knowledge-Informed FeO Prediction in Sintering

Abstract

Lack of crucial state data is a common problem in processing industries, particularly in iron-making. Real-time measurements are often infeasible in harsh conditions such as high temperatures and heavy dust. In practice, technical experts rely on manual observations to make estimates; however, this knowledge is difficult to formalize and quantify. To date, few studies have effectively addressed these two challenges in a unified manner, hindering progress in data acquisition and process optimization. In this article, a novel knowledge-informed method that integrates a multisource fusion model with an ensemble network (KIMEN) is proposed to predict a key chemical indicator in the industry, the FeO content. First, a sParts-Pair comparing sorting (s-PCS) strategy was introduced for knowledge solidification. Furthermore, a knowledge-informed image processing scheme was proposed. In cases of data scarcity, we proposed a two-layer cascaded structure combining gradient boosting decision tree (GBDT) and gated recurrent unit (GRU), which functions as an ensemble recurrent network. When applied to the Guangxi Liuzhou Iron & Steel (Group) Company, the method demonstrates improved prediction performance and practical effectiveness. Experimental results show that the proposed KIMEN outperforms some conventional methods and state-of-the-art approaches. Ablation studies, small-scale experiments, and transfer learning experiments further validate the advantages of our method.