In the emergency response process of coal-fired power enterprises, working environments are often characterized by high dust concentration, insufficient illumination, and frequent occlusion, which severely restrict the performance of unsafe behavior detection models and make it difficult to meet real-time early warning requirements. To address this challenge, this study optimizes the YOLOv8 framework: on the one hand, a local feature filtering mechanism is introduced to effectively enhance the recognition of small and occluded targets; on the other hand, a triple attention module is integrated with the C2f_DWR module to achieve collaborative optimization of feature extraction and representation. Experimental results show that when the triple attention module and the C2f_DWR module are used independently, the detection speed decreases by approximately 6 FPS and 9 FPS, respectively, while both improve accuracy by more than 3%. When the two modules are combined, the detection speed decreases by about 13 FPS, while the mean Average Precision (mAP) significantly increases from 38.5% to 46%, achieving a substantial improvement in detection accuracy while maintaining acceptable detection speed. The findings demonstrate that the proposed method can effectively enhance unsafe behavior detection and early warning in complex industrial environments, providing a practical technical reference for safety monitoring systems in coal-fired power enterprises during emergency response.