<p>Scrap steel, as a recyclable secondary raw material with high embedded energy value, plays a critical role in promoting the sustainable development of the steel industry through its efficient utilization. During electric arc furnace (EAF) steelmaking, uncertainty in the charging weight of different scrap categories can easily lead to deviations in burden proportioning, fluctuations in energy consumption, and reduced compositional control accuracy, thereby affecting product quality and production cost. However, existing studies on scrap steel mainly focus on image segmentation and classification, whereas limited attention has been paid to weight prediction of different scrap types under practical EAF charging conditions. Current practice still largely relies on manual estimation, resulting in low accuracy. To address this issue, this study proposes a scrap weight prediction method for different material categories based on semantic segmentation and machine learning. First, a UNet model with VGG as the backbone network was developed to perform semantic segmentation of different scrap images and extract multidimensional features, including area, morphology, texture, and color characteristics. The model achieved high segmentation accuracy on a self-constructed scrap image dataset, with mIoU, mPA, and Acc reaching 92.14%, 95.88%, and 98.26%, respectively. Compared with SegFormer-B0, FCN, and PSPNet, the proposed model improved mIoU by 8.35%, 15.34%, and 14.73%, respectively. Subsequently, the particle swarm optimization (PSO) algorithm was introduced to optimize the hyperparameters of the random forest (RF) regression model, enabling efficient coupling between segmentation features and weight prediction. The results demonstrate that the PSO-RF model achieved superior performance in weight prediction, with MSE, MAE, and RMSE values of 0.197, 0.158, and 0.356, respectively, while the coefficient of determination (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation>) reached 0.932. Compared with MLP, LightGBM, and SVM, the <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation> value increased by 0.064, 0.047, and 0.047, respectively, indicating significantly better performance than other regression models. This method enables accurate estimation of scrap charging weight during the EAF feeding stage and provides effective technical support for improving burdening accuracy, optimizing smelting operations, and predicting smelting endpoints.</p>

Scrap weight prediction for different scrap types based on semantic segmentation and machine learning

期刊论文
发表日期：

2026年6月26日

查看全文

Jihu Yin,
Pengcheng Xiao,
Bixia Zhang,
Liguang Zhu

摘要

Scrap steel, as a recyclable secondary raw material with high embedded energy value, plays a critical role in promoting the sustainable development of the steel industry through its efficient utilization. During electric arc furnace (EAF) steelmaking, uncertainty in the charging weight of different scrap categories can easily lead to deviations in burden proportioning, fluctuations in energy consumption, and reduced compositional control accuracy, thereby affecting product quality and production cost. However, existing studies on scrap steel mainly focus on image segmentation and classification, whereas limited attention has been paid to weight prediction of different scrap types under practical EAF charging conditions. Current practice still largely relies on manual estimation, resulting in low accuracy. To address this issue, this study proposes a scrap weight prediction method for different material categories based on semantic segmentation and machine learning. First, a UNet model with VGG as the backbone network was developed to perform semantic segmentation of different scrap images and extract multidimensional features, including area, morphology, texture, and color characteristics. The model achieved high segmentation accuracy on a self-constructed scrap image dataset, with mIoU, mPA, and Acc reaching 92.14%, 95.88%, and 98.26%, respectively. Compared with SegFormer-B0, FCN, and PSPNet, the proposed model improved mIoU by 8.35%, 15.34%, and 14.73%, respectively. Subsequently, the particle swarm optimization (PSO) algorithm was introduced to optimize the hyperparameters of the random forest (RF) regression model, enabling efficient coupling between segmentation features and weight prediction. The results demonstrate that the PSO-RF model achieved superior performance in weight prediction, with MSE, MAE, and RMSE values of 0.197, 0.158, and 0.356, respectively, while the coefficient of determination ( \(R^2\) ) reached 0.932. Compared with MLP, LightGBM, and SVM, the \(R^2\) value increased by 0.064, 0.047, and 0.047, respectively, indicating significantly better performance than other regression models. This method enables accurate estimation of scrap charging weight during the EAF feeding stage and provides effective technical support for improving burdening accuracy, optimizing smelting operations, and predicting smelting endpoints.