Research On Power Grid Foreign Objects And Abnormality Detection Based On Image Recognition

The work of power patrol inspection is mainly performed manually,which requires a lot of manpower and material resources,and the real-time performance is poor.It can effectively improve inspection efficiency if image recognition algorithms can be used to analyze the on-site images,and promptly notify staff when hidden dangers that threaten the safe operation of the power grid are detected.There are still some problems with using image recognition algorithms to identify inspection images.Among them,smallsized targets represented by pin defects are difficult to identify using generalpurpose detection algorithms.In addition,taking advantage of deep learning requires a number of effective samples,fewer real samples containing hidden dangers,and sometimes even no suitable samples,which often cannot meet the training requirements of deep learning algorithms.In the face of the current Ubiquitous Power Internet of Things construction needs,it is necessary to introduce edge computing into power inspection image recognition,so that devices on the edge side can complete more computing tasks.This article first set up a power patrol inspection image recognition system to detect foreign objects or abnormal conditions such as broken construction machinery and insulators,and notify staff when problems are found.Aiming at the problem of identifying small targets such as pin defects,a two-level detection algorithm for pin identification is proposed.The first level uses a object detection algorithm to locate the pins,and the second level uses a support vector machine to determine whether pin defects exist.To solve the problem of image shortage,this paper studies two sample generation algorithms for expanding samples.The first one is based on an image fusion algorithm,which fuses targets and background images according to some rules to achieve the purpose of expanding the data set in batches.The second method is based on Generative Adversarial Networks,selecting the appropriate structure,and directly generating the required samples.To meet the need of introducing power image recognition into edge computing,the application of edge computing is studied from the perspective of hardware and software.On the software side,the huge model of deep learning is compressed and accelerated on the premise of ensuring accuracy.Make it more suitable for the weak performance of embedded device hardware.In terms of hardware,we compares the time-consuming operation of image recognition of the general-purpose Arm-based embedded development board and the Da Vinci-based development board.The development boards of the two architectures each have their own advantages and can be targeted at different need to make different types of hardware choices.