Deep Learning Based Feature Selection and Ensemble Learning for Sintering State Recognition

Sintering is a commonly used agglomeration process to prepare iron ore fines for blast furnace. The quality of sinter significantly impacts the blast furnace ironmaking process. In the vast majority of sintering plants, the judgment of sintering quality still relies on the intuitive observation of the cross section at sintering machine tail by operators, which is susceptible to the external environment and the experience of operators. In this paper, we propose a new sintering state recognition method using deep learning based feature selection and ensemble learning. First, features from the infrared thermal images of sinter cross section at the tail of the sinterer are extracted based on ResNeXt. Then, to eliminate the irrelevant, redundant and noisy features, an efficient feature selection method based on binary state transition algorithm (BSTA) is proposed to find the truly useful features. Subsequently, an ensemble learning (EL) method based on group decision making (GDM) is proposed to recognize the sintering states. Novel combination strategies considering the varying performance of the base learners are designed to further improve recognition accuracy. Industrial experiments conducted at a steel plant verify the effectiveness and superiority of the proposed method.