Abstract: Aiming at the major hidden dangers such as belt tearing, fire disastes and others caused by foreign objects in downhole belt conveyors, a type of high-precision real-time detection method, YOLO11-MRNet, is proposed, and a three-stage progressive thawing training strategy is designed. On the self-built dataset, the mAP@50-95 of the designed scheme reaches 0.841, with an increase of 15.8% compared to the original YOLO11n, an increase of 14.3% compared to conventional single-stage training, and a decrease of 10.5% in training time consumption. The main innovations include: proposing SOEP pyramid and optimized RFPN, achieving maximum preservation of shallow layer details and efficient direct access to deep layer semantics; A CSP-OmniKernel multi-core reparameterization module is designed to significantly enhance multi-scale receptive fields under zero inference overhead; A three-stage progressive thawing training strategy is proposed to improve the overfitting problem of complex models under small sample industrial scenes. The ablation experiment and visualization results show that this method remains extremely high accuracy and confidence degree under extreme working conditions such as dense small targets, severe occlusion, low lighting at long distances and others, and has excellent robustness and engineering deployment value.