A Transformer and Random Forest Hybrid Model for the Prediction of Non-metallic Inclusions in Continuous Casting Slabs

Abstract

Non-metallic inclusions (NMIs) in continuous casting slabs will significantly reduce the performance of final steel products and lead to other defects in steel products. The traditional detection methods of NMIs in continuous casting slabs have the problem of low efficiency, and it is complicated to establish a prediction model of NMIs based on physics and chemistry. Therefore, we tried to use the machine learning method by integrating Transformer and Random Forest and established an RF-1DViT model to predict NMIs in continuous casting slabs. To predict the occurrence and the location of NMIs as accurately as possible, the whole process data of steelmaking, refining and continuous casting were used, and the continuous casting slab was processed in slices. The experimental results show that the proposed RF-1DViT model has an F1 score of 0.8991, surpassing Logical Regression, K-Nearest Neighbor, Support Vector Machine, Random Forest, AdaBoost, GradientBoost, Multi-Layer Perceptron and 1DViT model, and has good interpretability and strong feature extraction ability. By means of the Random Forest and histogram, the process importance can be analyzed and rules of inclusions generation can be given. The t-SNE manifold learning method can further assist researchers to accurately locate the defect.