Research On Electrical Equipment Detection Method Based On Deep Learning

For a long time in the past,the rapid development of China’s economy has accelerated the pace of infrastructure construction,of which China grid system is a typical representative.In the past years,China grid system has achieved remarkable achievements in many aspects,leading the world in technology and reaching the level of enabling extremely remote mountain areas to use electricity.With the increasing scale of power grid system,the urgent challenge of power grid system is the security problem.In order to solve this problem,the traditional method is to take a great many thermal images or visible light images of electrical equipment along transmission lines,distribution lines and substations by using a large number of manpower,and then diagnose and judge whether these images have faults and the types of faults manually.At present,the most common method is to use a large number of thermal images of electrical equipment taken by thermal imager or visible light images of electrical equipment taken by helicopters and unmanned aerial vehicles during patrol inspection,and then analyze images by combining with an electrical equipment intelligent analysis system to judge whether certain types of equipment on the image have certain faults,abandon the tedious labor before,and realize automation of electrical equipment faults discrimination.Electrical equipment detection,whether based on thermal images or visible light images,is very important part of electrical equipment intelligent analysis system.Therefore,the research on electrical equipment detection technology has important theoretical significance and practical value.The concept of smart grid was put forward by China State Grid Corporation in May 2009.Its contents and tasks include technology and management.One of the tasks is to realize automation in technology.The purpose of the research in this dissertation is to realize the automation of equipment fault diagnosis.Due to the rapid expansion of power grid in scale,images of electrical equipment will also increase rapidly,and the traditional manual diagnosis and discrimination methods cannot meet the actual needs.With the rapid development of convolutional neural network,in the field of computer vision,deep learning technology has achieved very excellent results.The model of deep learning can quickly extract features which is more deeper of images and have high flexibility and robustness.Therefore,aiming at the thermal image dataset and visible image dataset of electrical equipment,this dissertation uses the object detection algorithm based on deep learning technology and improvement on the classical algorithm,and proposes the electrical equipment detection method that can be applied in practice.This dissertation proposes a new visual task to solve the problem of automatic fault diagnosis of electrical equipment,which named electrical equipment detection.This dissertation has built the corresponding data benchmark and combined the deep learning technology which is very popular nowadays to carry out the corresponding work and research.The following is a summary of the main work and contributions of this dissertation:(1)Electrical equipment datasets TEED001 and VEED001 of thermal images and visible images are constructed.The dataset TEED001 includes 5558 images,11180 device instances and 21 device categories.The dataset VEED001 contains 3669 images,15182 device instances and 14 device categories.During the labeling process,label and record the location,size and category of all the devices of interest with rectangular boxes.(2)A thermal image electrical equipment detection method based on self-attention mechanics is proposed.At present,advanced object detection algorithms have achieved high detection accuracy,but many methods do not make good use of context information,so the obtained feature representation is not optimal,which leads to the detection performance of many algorithms still have the potential to improve.With improvement of DANet,this dissertation propose CPDANet on the basis of self-attention mechanism.Algorithm use channel attention module and position attention module in CPDANet to adaptively integrate local features and global dependency to obtain more optimized feature representation.Then algorithm use this feature as the input of Faster R-CNN for subsequent processing to obtain better detection performance.(3)A fast multi-scale electrical equipment detection method in visible light images of UAV is proposed.In this dissertation,the preprocessing and post-processing of high-resolution visible light images are combined with YOLOv3 to improve the performance of small target detection.In addition,in this dissertation,the multi-scale feature fusion in YOLOv3 is improved by combining ASFF and SE-Net methods and named as SEFF.Image preprocessing are overlapping clipping of images.A high-resolution image is overlapped and clipped into multiple low-resolution images.Then the detection results of multiple images are combined by NMS,which is called post-processing.These operations greatly improve the detection performance of small targets and targets with severe scale changes.