Research On The Application Of Machine Learning In Monitoring Images Of Power Transmission System

With people’s dependence on electric power more and more,the safe operation of power system is related to the stability of people’s life,industrial production is carried out in an orderly manner,so it is very important to ensure the safety and stability of power system.Large-scale electricity from production to use to pass through power generation,transmission,distribution and electricity four links,these four links constitute the power system.Among them,transmission link is an important component,transmission system is a large-scale interconnection system consisting of high-voltage transmission lines,substations,etc.distributed in a vast area,and is also the largest artificial energy transmission system.The transmission system is large in scale,complex structure,easy to receive a variety of influence stoush disturbances,such as high-voltage transmission lines in the wild is very easy to be hung foreign bodies,the main alien species include dust-proof nets,plastic bags,kites,etc.,these foreign bodies easily cause high-voltage transmission lines between phase discharge,resulting in large-scale power outages,Sometimes foreign objects ignited by high-voltage transmission lines fall can also cause fire casualties,and in substations,there are many meters and switches need to manually read the value or determine the status,the manual reading practice is very inefficient,when the abnormal value of the meter appears,it is difficult for the staff to timely find timely processing,thus causing power supply accidents.The main research content of this thesis includes two parts,the first part is the use of machine vision for high-voltage transmission line foreign object detection,the second part is the use of machine vision for the automatic identification of substation pointer meters.In the field of foreign object detection of high-voltage transmission lines,a high-voltage transmission line positioning algorithm based on Bresenham is proposed,which is used to accurately locate the location of high-voltage transmission lines and search for high-voltage transmission lines.The abnormal area of the high-voltage transmission line is found by projecting histogram in the normal direction of the vertical transmission line.The abnormal areas identified in this way are not necessarily foreign objects,and may often be transmission line clips,which are then classified.Alien species are often diverse,but the collection of different species samples is very difficult,because the occurrence of foreign body hanging is often accidental,so this thesis presents a GANs-based feature enhancement algorithm FGANs,using GANs to directly learn the characteristics taken by CNN rather than raw data.The accuracy of the classification of foreign objects can be effectively improved even when the sample is very limited.In addition,the deep learning-based target detection algorithm SSD is used for foreign object detection,because SSD is a global search algorithm,which results in targets that are not suspended on power lines similar to foreign objects are detected,such as clouds,high-rises at the edge of the skyline,etc.Therefore,it is necessary to use the high-voltage transmission line positioning algorithm proposed in this thesis to filter and remove false positives.In addition,SSDs need to be classified using a linear regression layer and softmax activation function,so using FGANs-generated features for training SSDs can significantly improve test results and improve model performance.In the automatic recognition part of substation meter,a common automatic meter recognition model based on deep learning is proposed.It is very difficult to automatically read the pointer meter in the image for two main reasons,if you want to get an accurate reading,the input image must be positive view,and not be tilted,but the image obtained by the patrol robot or fixed camera is often unstable,difficult to meet the above requirements.Second,the traditional algorithm can only recognize a specific type of meter,that is,an algorithm corresponds to one meter,but substations often have a variety of meters,for all types of meters to design the algorithm is very laborious,and difficult to ensure its accuracy.In this paper,a pointer meter detection model based on convolutional neural network is proposed,which obtains the feature points on the meter and calculates the single-response matrix to establish a mapping from other views to positive views,and infers the position and angle of the pointer with the semantic information of the image.Therefore,the model can perform both view correction and reading acquisition after obtaining the position and type of the gauge through the target detection algorithm.This model is a generic model that reads all types of pointer meters.