Abstract: Current shaking table concentration operations rely on manual observation of concentrate zoning positions, followed by manual adjustment of the concentrate cutting plate. This approach suffers from low efficiency and high error rates. To replace manual labor and achieve rapid, automatic recognition of concentrate zoning positions, this paper proposed an intelligent extraction method for concentrate zoning positions on a shaking table based on an improved YOLOv11n algorithm, termed SFEL−YOLO, a lightweight detection model. First, a lightweight StarNet backbone network was adopted to replace the original YOLOv11n backbone, reducing the number of model parameters. Partial convolution (PConv) was introduced in the convolutional layers to replace depthwise separable convolution (DWConv), thereby reconstructing the StarNet architecture into a novel lightweight P−StarNet backbone, which further reduced redundant parameters and improves detection accuracy. Second, the FasterNet Block module was integrated into the C3k2 module of the neck network to enhance detection performance in complex environments. Additionally, an efficient multi−scale attention mechanism (EMA) was incorporated to construct a Faster_EMA module, improving the model's anti−interference capability. Finally, a lightweight shared convolutional detection head (LSCD) was designed, combining shared convolution, group normalization (GN), and the distribution focal loss (DFL) for bounding box regression to reduce model complexity and improve detection efficiency for concentrate zoning positions on the shaking table. A dataset of mineral zoning images was collected from multiple tungsten shaking tables in a tungsten beneficiation plant using industrial cameras. Experimental results show that the improved SFEL−YOLO model achieves an mAP₅₀ of 96.8%, with Parameters, GFLOPs, and Weights of 1.629×10⁶, 3.9, and 3.5 MB, respectively. Compared to the YOLOv11n model, the mAP₅₀ is improved by 1.9%, while Parameters, GFLOPs, and Weights are reduced by 36.9%, 38.1%, and 36.4%, respectively. The deployed model achieves a detection accuracy of 93.5% for concentrate zoning positions on the shaking table. The SFEL−YOLO model ensures high−precision detection of concentrate zoning positions while significantly reducing model complexity and lowering the miss detection rate, meeting the requirements for both accuracy and lightweight deployment in industrial shaking table concentrate zoning position recognition.