Abstract
When conducting top coal caving, the timely closure of the coal discharge port is crucial for controlling the recovery rate of top coal. The existing semantic segmentation models have issues such as low degree of lightweighting, and can only provide control decisions based on the pixel gangue ratio in the discharge port. This study proposes a real-time prediction framework for end-to-end inference of the mass gangue ratio from coal flow images. Its core is a lightweight model named LEG-U-Net, which integrates knowledge distillation technology and modules such as depth separable convolution, sub-pixel convolution, global shared coordinate attention, and edge guided attention. Experiments show that compared with the teacher model, LEG-U-Net reduces the number of parameters by 87%, increases the FPS by 55%, and achieves an mIoU of 95.27%, achieving a good balance between efficiency and accuracy. Furthermore, through discrete element simulation and mathematical modeling, a quantitative mapping relationship between pixel gangue ratio, volume gangue ratio, and mass gangue ratio was established, enabling end-to-end inference from images to quality gangue ratio. This method provides a new paradigm for real-time intelligent prediction of coal flow gangue ratio, and can effectively support process control and quality optimization decisions in the coal processing process.
Funding Information
  • National Natural Science Foundation of China (U23A20599)

This publication has 31 references indexed in Scilit: