Research On Infrared Image Recognition Technology Of Power Equipment Based On Convolutional Neural Network

Power equipment is the cornerstone of the power system.Due to the advantages of infrared detection technology with non-contact detection,power grid companies often use infrared detection technology to conduct regular manual inspections of power equipment.However,at present,the efficiency of manual inspection is low,and it is easily affected by the subjective experience and working time of inspectors,resulting in misjudgment of the status of power equipment,which will be difficult to meet the needs of real-time processing of a large number of power equipment infrared data in the future.In addition,there are still many problems in the process of infrared image detection and recognition of power equipment,such as low infrared image contrast,mutual occlusion of power equipment,and imbalance of positive and negative samples in the image.We carry out the research on infrared image detection of power equipment based on deep convolutional neural network in this thesis.We firstly introduce the common power equipment detection algorithms,analyze the advantages and disadvantages of the classic target detection algorithms in detail,and summarize the detection difficulties of infrared images of power equipment.Secondly,we adopt random cropping,random affine,random erasure and Mosaic data augmentation as online data augmentation of infrared images of power equipment;noise,pseudo-color transformation,and filled polygonal occluders are selected as offline augmentation.Next,super-resolution reconstruction is performed on the augmented dataset using the Fast Super-Resolution Convolutional Neural Network algorithm to achieve image enhancement.In this thesis,a variety of online and offline data enhancement methods are used to augment the infrared images of power equipment,which can build a dataset of infrared images of power equipment,improve the generalization ability and robustness of the network model,and solve the problem of low resolution of infrared images of power equipment,low contrast and small datasets.Finally,improved YOLOv4 model to achieve infrared image detection and recognition of power equipment is proposed.Because the R-CNN series algorithms in the typical two-stage target detection algorithm have low detection speed,and the YOLO series algorithms of the one-stage target detection algorithm have low precision and are difficult to deploy in actual engineering equipment,the lightweight network Mobilenetv3 is first introduced as YOLOv4 The feature extraction structure of the network performs feature extraction on the image of the power equipment at the input end,which greatly reduces the amount of parameters and computation,and speeds up the process of feature extraction;then design an appropriate size Anchor Box according to the K-means++ algorithm;then the traditional NMS algorithm is improved to a double-threshold soft-NMS algorithm,which suppresses redundant frames of power equipment and avoids missed detection of targets,and solves the problem of missed detection caused by mutual occlusion of power equipment;finally,the Focal loss algorithm is introduced into the loss function to reduce the simplicity,the weight of easy-to-classify negative samples in network training,and at the same time increase the weight of difficult samples,which solves the problem of unbalanced positive and negative samples in infrared images of power equipment.In this thesis,a power equipment infrared image detection network based on the YOLOv4 algorithm of the Pytorch deep learning framework is built.Through the verification and effect test of the proposed algorithm model,the results show that the improved YOLOv4 detection method achieves high detection accuracy and real-time detection.This lays the foundation for the automatic and intelligent detection and identification of infrared images of power equipment,and lays the groundwork for the subsequent diagnosis of power equipment status.