To address the problems of complex underground coal mine working environments, low efficiency of manual monitoring, and difficulty in real-time deployment of miner behavior detection models on low-computing-power edge devices, a lightweight miner behavior recognition method based on improved YOLOv11n is proposed. YOLOv11n is used as the baseline model. The original backbone structure is replaced with a lightweight FasterNet backbone to reduce the computational cost in the feature extraction stage. A C3k2 module with re-parameterized branches, namely C3k2-Deep Branch Block (C3k2-DBB), is introduced to enhance the feature representation capability. A lightweight shared convolution detection head (LSCD) is adopted to reduce redundant structures in the detection stage, thereby further decreasing model complexity and improving inference efficiency. The experimental results show that the number of parameters, floating-point operations, and model size of the improved model are reduced by 39.3%, 48.2%, and 39.0%, respectively, while the CPU and GPU inference frame rates are increased by 76.1% and 32.3%, respectively. In terms of detection performance, the precision is slightly improved, while the recall and mAP@0.5 decrease slightly. The proposed method effectively reduces model complexity and improves inference efficiency under controllable changes in detection performance, especially improving real-time inference performance on the CPU side. It can provide technical support for miner behavior monitoring in low-computing-power underground edge scenarios.