Research On Key Technologies Of Dangerous Object Detection In Transmission Corridors Based On Deep Learning

As the transmission corridors environment is becoming more and more complex,and the possibility of damage due to external force is increasing,which seriously threatens the reliability of power supply.At present,the routine transmission line patrol is mainly conducted by manual route patrol,unmanned aerial vehicle(UAV)patrol and real-time monitoring.Manual line patrol is inefficient,time-consuming and laborious.The working distance and endurance of the UAV are limited,which makes it diffilcult to be extended to the power grid system in the large area.Compared with manual line patrol and UAV,the transmission corridor environment can be monitored by monitoring equipment for the long time and the wide range,which alleviates the difficulty of inspection.However,real-time monitoring cannot get rid of manual detection of dangerous foreign objects,and relying on subjective judgment cannot guarantee efficiency and accuracy.With the increase of monitoring coverage,the processing of massive images increases the workload of inspection.Obviously,the use of the monitoring system fails to realize the real intelligent operation and inspection of the transmission corridors,and cannot fundamentally solve the difficulty in handling external force damage events.Deep Learning has made remarkable progress in machine vision,speech recognition,natural language processing,machine translation,data mining,automatic driving and other aspects.It also provides new ideas and methods for detecting dangerous objects in transmission corridors.This thesis studies the image detection method with the aim of alternative artificial identification.This method can help the operation and maintenance staff to timely detect the invasion of dangerous foreign objects in the transmission corridors.The efficiency of the inspection work of the transmission corridors is improved to ensure the safety and stability of the power grid operation environment.Based on Deep Learning convolutional neural network and object detection technology,this thesis proposes the research on key technologies of object detection in transmission corridors Based on Deep Learning.Through image preprocessing and model improvement to improve detection performance,the function of fast and accurate identification and positioning of dangerous objects in transmission corridors is realized.The main work completed in this thesis is as the follows:(1)The thesis introduces the background and significance of the research topic,and analyzes the necessity of the research on the detection of dangerous objects in the transmission corridors.This thesis summarizes the development process of Deep Learning,dangerous object detection in the transmission corridors and the research status of dangerous object detection by Deep Learning.(2)The thesis analyzes the basic principle of deep learning and the structure of convolutional neural network.The feasibility of Deep Learning in the detection of dangerous object in transmission corridors is explained from the object detection principle of convolutional neural network.The YOLOv3 detection model is selected as the research object of this thesis.(3)The monitoring images of transmission corridors have the characteristics of large amount,low proportion of effective data and low quality.Data augmentation is proposed to preprocess the monitoring images.The images are labeled to form the enhancement training sample set.It also reduces the influence of insufficient samples and environmental changes on network model training and identification,and improves the network model robustness and generalization ability.(4)K-means clustering was carried out on the size of labeled box of sample set.The preset bounding box size of YOLOv3 was modified,so as to reduce the difficulty of YOLOv3 detection bounding box adjustment and model learning loss.It also improves model positioning accuracy and detection ability.The effectiveness of the improved method is verified based on the actual monitoring image.(5)Non-maximum suppression algorithm principle and defects are introduced and analyzed.With reference to the Soft-NMS algorithm and GIoU algorithm,Non-maximum suppression algorithm of YOLOv3 bounding boxes is improved,to enhance the target detection precision rate and recall rate of detection model.The effectiveness of the improved method is verified based on the actual monitoring image.