| As an important part of the power grid intelligent inspection system,the detection of foreign objects in power transmission lines is of great significance to the safe and reliable operation of the power system.At present,high-voltage overhead transmission lines are mostly bare wires,insulated by air.Therefore,hanging foreign objects such as kites,balloons and greenhouse films can cause phase-to-earth or phase-to-phase short circuits,resulting in large-scale power outages in the area,leading to outage accidents.In this dissertation,based on the deep learning algorithm,the foreign object detection method in the power inspection image is studied,and the main contributions are as follows:1.At present,there is no internationally disclosed image data set for transmission line inspection.This dissertation analyzes the types and characteristics of common foreign objects in power transmission lines and constructs an inspection image data set.In order to facilitate the training and testing of the algorithm,the data is labeled and divided.The training data is expanded using data enhancement methods such as flipping and brightness adjustment.The image data format is converted for different algorithms to speed up the data reading speed.Finally,the difficult points such as the feature extraction of foreign objects in the inspection images and the limitations of the current related research are summarized.2.This dissertation considers the hazards of missing foreign objects to the power system.Therefore,this dissertation defines the 100% recall rate model of the classification algorithm.Based on this,this dissertation improves the problem that the InceptionV3-retrain algorithm has a higher false positive rate at 100% recall rate model.First,the effectiveness of different data enhancement methods is studied,and it is verified that the data enhancement strategy of horizontal flipping reduces the algorithm’s false positive rate by 6.3%.Second,this dissertation improves the training method,combines the advantages of transfer learning and fine-tuning training,reduces the risk of network overfitting,and reduces the false positive rate by 9.7%.Third,this dissertation improves the network structure.According to the task characteristics,the network structure is optimized.The feature learning ability of the network is improved,and the false positive rate drops by 6%.Finally,the experimental results show that the improved classification algorithm in this dissertation achieves the task requirement of 100% classification of foreign object images,and the false positive rate drops by 22%.3.Based on the classification of foreign object images in the previous article,this dissertation further expands the research on the method of marking the location of foreign objects in the images.First,a "warning-review" foreign object recognition and marking strategy combining image classification and target detection is proposed.Second,this dissertation studies SSD algorithm and Faster R-CNN algorithm.Comparative experiments are designed to verify the effectiveness of the "warning-review" foreign object identification and marking strategy.The results show that adding classification links increases the average AP value of the target detection algorithm by 3.9%.Finally,the joint algorithm of foreign object image classification and auxiliary labeling is studied,and the index of the algorithm is evaluated from the recall rate of foreign object detection.On the Faster R-CNN foreign object auxiliary labeling algorithm based on the residual network,84.1% of foreign object targets are correctly positioned and labelled. |
