Abstract: In order to effectively identify the dress code of the field workers in substation and prevent the workers from wearing short sleeves or rolling up long sleeves, the Yolo v5 real-time object detection algorithm is used in this paper, in order to solve the problem of long-sleeve/short-sleeve detection, the classifier is modified and the output layer is modified to 21-dimensional tensor. The 7135 samples collected from the actual working site of substation are annotated, the data set is enlarged by image and adjusted by color, and the initial anchor frame is cluster analyzed. Based on the Yolo v5 pre-training model on COCO, the training targets are adjusted by means such as hard-negative-mining, and finally a standardized detection model for staff dressing in the optimal monitoring scene is obtained. The experimental results show that on the test set of 1000 images, the precision is 97.01%, the recall is 96.60%, and the mAP value is 0.8767. In the experimental environment, the average response speed reached 38 ms. Finally, the model is encapsulated as a micro-service and deployed in a substation auxiliary system. The identification model of dress code has the advantages of high speed and less false alarm, which meets the requirement of safety inspection.