Research On Visual Detection Algorithm For Live Working Robot On Power Transmission Line

Transmission line is exposed to the outside for a long time,and affected by complex weather,charge changing and other factors,which lead to line faults such as bolt loosened,wire broken,accessories broken and so on.In order to ensure the steady transmission of electric energy,the workers must wear shielded clothing to work in high-risk and high-pressure environments.It is an effective way to liberate workers’ pressure that using electric power live working robot.However,the current robots have disadvantages of low automation and low working efficiency.In recent years,machine vision detection technology has been widely used in various fields,among which deep learning as a novel visual detection technology has demonstrated excellent detection performance.In this thesis,aiming to the bolt tightening end device and wire repairing end device of live working robot on power transmission line,visual detection algorithms are studied to improve working efficiency of robot.The related image processing algorithms such as Hough transform and SVM classifier are studied firstly.For the bolt detection of bolt tightening end device,the hexagonal peak voting strategy based on Hough transform is designed to detect hexagon bolt.HOG and SVM techniques are used to classify the objects in pictures taken by robot’s camera.For wire faults detection,Hough transform is used to extract special texture of wire and the least squares algorithm is applied to fit the centerline of wire,finally LBP and SVM technique are used to determine wire is defective or normal.The experiments show that the tow detection algorithms based on Hough transform and SVM classifier can detect target under certain circumstances,while the problems such as illumination,occlusion are still not solved and the positioning accuracy also need to be improved.Aiming at problems of poor robustness and high error detection rate mentioned above,the novel bolt detection and wire faults detection algorithms based on Faster R-CNN are proposed which has good performance among deep learning algorithms.Firstly,the related research on Faster R-CNN is carried out.For difficult samples,online hard example mining network is added to improve the detection accuracy.In order to solve the problem of insufficient samples,a variety of sample expansion methods to increase the number of samples is proposed.Finally,the bolt detection network and wire fault detection network can be obtained through joint training.The experiments demonstrate that the proposed algorithm based on Faster R-CNN perform higher accuracy and location precision for bolt and defective wire detection than Hough and SVM method,has a good applicability.Finally,the trained deep learning network model is put into the robot,and the operation experiments are completed on the 110kV power transmission line.Experimental results show that the proposed detection algorithm based on Faster R-CNN can detect bolt and wire defects greatly,improve the working efficiency,and has broad application prospects.