Mask-RCNN-CHFNet: An improved deep learning for 3D reverse modeling of iron tailings (SiO2) real-time melting process

Abstract

The melting process of iron tailings is influenced by thermodynamic and kinetic factors, with particle size directly affecting the melting rate. As iron tailings absorb heat, the slag system's temperature drops and viscosity increases, making particle size and melting rate critical for temperature regulation and heat compensation. In this study, a CCD camera was used to track SiO₂, the main component of iron tailings in a high-temperature molten pool, to monitor its melting behavior. The Mask-RCNN-CHFNet model is used to perform semantic segmentation on images, and an end-to-end convex hull filtering (CHF) framework is constructed to achieve quantitative analysis of the volume change and morphological evolution of high-temperature melts. During neural network training, the loss value is 0.098. On the test set, the model achieves AP50–95 of 45.4, AP50 of 82.0, and AP75 of 40.8. 3D reverse modeling is then performed on the segmented SiO₂ regions. By combining experimental data with intelligent algorithms, the complex high-temperature melting process is translated into a computable mathematical relationship. Compared with the existing water quenching technology, continuous monitoring, tracking and tempering are carried out. This approach establishes a reliable time-sequence law, providing real-time data for iron tailings melting and improving slag cotton quality.