A deep regression model for silicon content estimation using iron sample images in blast furnace ironmaking

Abstract

Silicon content in hot metal is a crucial indicator of the thermal state and operational stability of the blast furnace (BF) ironmaking process. Real-time and accurate monitoring of silicon content is essential for maintaining furnace efficiency and ensuring high-quality iron production. Current measurement methods, which depend on manual sampling and offline analysis, often lead to significant delays. To address this issue, this paper proposes Images-Based Silicon Content Estimation Network (IBSCE-Net), a deep regression model that leverages iron sample images for real-time silicon content estimation. First, a tailored data augmentation strategy is applied to enhance the limited iron sample images. Second, an improved network structure is designed, where efficient multi-scale attention ) module is integrated into the residual network (ResNet) to enhance its feature extraction capabilities, and a feature calibration module based on feature distribution smoothing is used to overcome the problem of imbalanced label distribution. Furthermore, an adaptive weighted penalty loss function is proposed that considers both the disparity in label distribution and the specific objectives of the silicon estimation task. The experimental results demonstrate that IBSCE-Net achieves a root mean squared error of 0.0583, with 92.71% of the estimations having a margin of error below 0.1%. These results highlight the effectiveness of the proposed approach in accurately estimating the silicon content in BF ironmaking, providing precise decision support for intelligent BF operations.