This paper constructs a safety belt intelligent monitoring system for power grid construction scenarios, realizing real-time monitoring and early warning of safety belt status through hardware and software co-design. Relying on acousto-optic controller to realize real-time feedback of buckle status, mosaic enhancement and environmental noise simulation are used to improve data diversity. The YOLO network is improved by introducing cavity convolution and depth separable convolution to optimize the feature extraction efficiency, and combining with progressive attention area network to enhance the feature characterization ability of small targets. Experiments show that the accuracy, recall, and AP value of the improved YOLO-DCM-DSCM-PAAN algorithm are improved by 5.65%, 1.07%, and 3.88%, respectively, compared with the YOLO algorithm. After the introduction of the DCM module, the mIOU value and F1 score of YOLO-DCM reached 0.83 and 0.81, respectively. The ablation experiments show that Experiment 4 fuses the two modules, DSCM and PAAN, and improves the mIOU and F1 scores by 6.0% and 7.4%, respectively, on the basis of Experiment 1, which is a more obvious improvement. The detection accuracy of the proposed improved algorithm reaches 96.91%, which is 2.45% higher than the YOLOX algorithm. It proves that the improved algorithm in this paper can meet the demand of real-time monitoring of complex construction scenarios, and the study provides an innovative and intelligent solution for electric power operation safety.
