Abstract: With the rapid development of power grid construction, the scale of technical support personnel in operation site is expanding continuously. Operation site belongs to high-risk work site, illegal wearing protective equipment will seriously endanger the workers. In order to improve the inefficiency of traditional manual supervision, this paper uses a real-time deep learning algorithm to detect illegal wearing behavior. The algorithm combines the real-time object detection network YOLOv3 and Self-Attention mechanism, uses the DANet structure for reference, and embeds the Self-Attention module at the high level layers of YOLOv3 network to better mine and learn location relations and channel relations of feature maps. The experimental results show that the mAP and Recall of this algorithm reached 94.58% and 96.67%. Compared with YOLOv3, its mAP and Recall increased by 12.66% and 2.69%. The accuracy of the model is significantly improved, which can meet the detection requirements of the task and improve the intelligence of power grid.

Key words: power operation, illegal wearing, YOLOv3, Self-Attention mechanism, object detection