Intelligent photoelectric dry separation for low-rank coal upgrading

Air dense medium fluidized bed (ADMFB) dry separation represents a key technology for achieving the clean and efficient utilization of coal resources. However, it still faces critical technical bottlenecks, including the difficulty of online identification of mismatched materials and the lagging regulation of separation parameters. In this study, an intelligent separation control strategy integrating an improved YOLOv13 model with Model Predictive Control (MPC) is proposed, and a closed-loop control framework of “visual perception–feature quantification–MPC decision–parameter execution” is established. A dual-channel image dataset of dry separation products, termed ADMFB-G3K , was constructed, covering separation conditions under six representative combinations of operating parameters. An improved target detection algorithm, SCU-YOLOv13, was developed by integrating a Lightweight Multi-Head Self-Attention module (LMHSA), a Multi-Scale Convolution–Attention Fusion module (MSCAF), and an Adaptive Intersection-over-Union loss function (AIoU). The LMHSA module reduced computational complexity by 68 % while improving recall by 3.5 %; the MSCAF module increased mAP@0.5 by 5.5 %; and the AIoU loss reduced bounding-box localization error by 40.2 %. Through the synergistic effect of the three modules, the model achieved a precision of 93.4 %, a recall of 92.5 %, and an mAP@0.5 of 97.8 %. Furthermore, an intelligent regulation system based on Model Predictive Control was established, in which the MPC controller was designed with a prediction horizon of Np = 20 and a control horizon of Nc = 10 . Experimental results demonstrate that MPC regulation increased the clean-coal yield by 1.79 %, reduced the ash content of clean coal to 9.97 %, and increased the ash content of tailings to 59.64 %. The probable error E p ​ decreased from 0.080 g/cm 3 under conventional open-loop/manual control to 0.065 g/cm 3 under the MPC closed-loop baseline, and further to 0.055 g/cm 3 after coordinated multi-variable optimization, and the control response time was shortened from 25-35 s to 8–12 s. In addition, the dynamic migration behavior of sulfur during the separation process was revealed. During the steady-state separation stage, the sulfur content in the upper layer decreased to 0.89–0.91 % , whereas the sulfur concentration in the lower layer reached 19.41–19.42 % , representing a 4.4-fold increase relative to the raw coal. After separation, the clean coal exhibited an ash content of 9.97 % and a sulfur content of 0.81 % , with a yield of 75.36 % ; the gangue product showed an ash content of 59.64 % and a sulfur content of 19.43 % , with a yield of 24.64 % , and the probable error E p was 0.055 g/cm 3 . These results demonstrate that high-precision intelligent dry separation of low-rank coal can be effectively achieved.