Research On The Recognition Method Of Abnormal Behavior In Electric Power Operations Based On Deep Learning

The power operation environment is very complex,with many potential safety hazards.The safety awareness of operators is uneven,and there are many situations where they forget to wear safety equipment due to carelessness or make unreasonable posture actions due to lack of concentration during the operation process.These small safety problems can accumulate and erupt,posing a serious threat to the personal safety of operators.Nowadays,the investigation of security issues mostly relies on traditional surveillance videos,making it difficult to detect problems in a timely manner.With the continuous development of intelligent management in the power system,the combination of computer vision and video monitoring is used to achieve automatic supervision in the power operation environment.Intelligent identification of abnormal behavior in the power operation environment effectively solves the problem of low efficiency in human supervision of operators’ behavior norms.At the same time,intelligent identification of abnormal behavior can efficiently and timely report sudden dangerous behaviors of operators,Be able to minimize the impact of safety accidents.With the continuous development of deep learning technology,Convolutional neural network has been widely used in the field of target detection,which makes the recognition of abnormal behavior in the electric power operation scene also have a considerable development.At present,the identification of abnormal behavior in power operation scenarios mostly involves detecting whether the operator is wearing safety items and whether the operator’s trajectory is abnormal.Research on the identification methods of the operator’s own dangerous actions mainly extracts surface features of the human body,which are easily affected by light.This article aims to apply pose estimation algorithms to accurately detect the wearing of safety helmets by operators and track them accurately,and to obtain high-level features such as the coordinates of key points in human bones that are concise,highly distinguishable,and not easily affected by light as the basis for action classification.It aims to build a deep learning detection based abnormal behavior recognition platform for power operation scenarios.This article first studied object detection algorithms based on deep learning,compared and analyzed Yolo series algorithms and R-CNN series algorithms.From the perspectives of real-time performance and accuracy,Yolo v5 was selected as the object detection algorithm in this article,and images from the COCO dataset were selected to train the Yolo model.The excellent performance of Yolo v5 algorithm was verified.On this basis,a multi target tracking algorithm based on object detection was studied,and their respective target association and update mechanisms were analyzed based on the running results of Sort and improved target tracking algorithms in actual scenarios.The quantitative indicators of the improved target tracking algorithm were specifically improved using MOT Challenge as an evaluation tool.Secondly,this article provides a detailed analysis of the Open Pose attitude estimation algorithm,and based on the key point detection results of attitude estimation,designs a behavior recognition algorithm that distinguishes various behavioral features using key point coordinates and bone geometry information.Finally,a more accurate intelligent control and abnormal behavior recognition were established in the power operation scenario.The reliability of the deep learning based anomaly behavior recognition algorithm proposed in this paper has been demonstrated by the excellent performance of improved multi-objective algorithms and behavior recognition algorithms based on object detection in practical scenarios.