Algorithms And Software Design For Fault Detection Of Transmission Line Patrol Images

The transmission line is the main artery of the national economy,and the safety of the line is the guarantee for stable economic operation.At present,the main way of patrol inspection is to take pictures of UAV or surveillance camera and then send them back to the background for manual expert review or automatic fault detection by computer.Most fault detection methods are based on image processing and machine learning.However,there are many types of faults and components to be monitored in the line.Traditional methods need to design different feature extractors for different kinds of targets.Few general algorithms can process these targets at the same time,and their processing ability for complex background is weak.The rapid development of deep learning makes it possible for algorithms to process multiple detection targets simultaneously.Based on the background of deep learning technology,a fault detection method for power components based on multi-objective detection is implemented,which can identify and determine multiple faults and locate the faulty components accurately.Firstly,this paper explained in detail the application scenarios and research objectives of the project,as well as the basic principles of object detection method,prepared a dataset containing four types of faults,six types of objects,and analyzed the main problems in fault detection of transmission line based on general object detection algorithms.Then,based on Faster RCNN algorithm,an adaptive brightness and contrast adjustment algorithm for bimodal gray histogram distribution image is proposed to enhance the processing ability of low quality image.According to the appearance characteristics of insulators and shock hammers,a non-maximum suppression algorithm based on area is proposed,which solves the problem of multi-labeling with single object.At the same time,the original image is cropped into several sub-images by the shear detection integration method,and the box fusion algorithm is designed to synthesize the detection results and improve the detection ability of small objects.A Two-Classifier integration method is proposed to make the same object be classified by two classifiers,and a weighted voting algorithm is designed to improve the classification performance of the detector.The optimal detection scheme of the integrated algorithm can reach 95.23% recall rate and 93.86% accuracy rate.Finally,based on Python language and PyQt library,a concise interactive application software is developed.The software can recognize the fault components stored in the local patrol image at the expert intelligent level,and has the functions of generating detection reports and updating detection models.The software achieves detection speed of 4 frames per second under GTX 1070 graphics card,which meets the application requirements of this project.