| As one of the indispensable equipment in power transmission system,insulator monitoring is crucial for the security and stability of power supply system.Insulator defects will severely impact the normal running of power transmission system.At the same time,it is a challenging task to detect insulator defects in complex background in high-resolution aerial images.Thus,it is of vital importance to carry out high quality continuous monitoring and defect detection of insulators.In this paper,the transmission lines in the power system are considered as the research background and the insulators and their defects in aerial images are regarded as the research object.In view of the current research status and present issues of insulator defect detection algorithms,deep learning algorithms are discussed.The concrete research content and innovation points are stated as follows:Some existing insulator defect detection algorithms focus on the extraction of physical features,which are not suitable for insulator defect detection under complex background,and have the shortcomings of low detection rate,accuracy and average precision.This paper presents a high-precision insulator defect detection algorithm based on Double Multi-level Multi-scale Detector(Double-M2Det),which is suitable for a variety of backgrounds.Firstly,the image data is expanded by data augmentation technology,and this paper uses M2 Det model to locate insulators.Then,the clipping block algorithm is used to process the located insulator image.At the same time,The M2 Det model is used to detect the insulator defects again.The results show that the high-precision insulator defect detection algorithm based on Double-M2 Det proposed in this paper can detect the defects with high precision for different materials insulators and various backgrounds,and greatly improves the precision and robustness of insulator defect detection algorithm.The high-precision insulator defect detection algorithm based on Double-M2 Det is optimized from the aspects of detection accuracy and precision.Although good precision performance is obtained,the improvement of precision performance often comes at a price.The high precision insulator defect detection algorithm based on Double-M2 Det has the problem of cumbersome detection process,resulting in low detection speed.To solve this problem,this paper optimized the detection process and detection speed.A fast insulator defect detection algorithm based on Improved Pyramid Scene Parsing Network(IPSPNet)and Single Shot Multibox Detector(SSD)was proposed.Firstly,the loss function and training stage of PSPNet network are optimized to generate a IPSPNet model suitable for insulator segmentation so that the insulator background in aerial image is removed.Then,SSD is used to detect insulator defects directly on the insulator image after removing the background.The results show that the fast insulator defect detection algorithm based on IPSPNet-SSD can greatly cut out the detection process and promote the detection speed while ensuring the precision.The insulator defect detection algorithm based on Double-M2 Det and based on IPSPNet-SSD are optimized from the two aspects of detection precision and detection speed,respectively,and the characteristics of high-precision and fast defect detection are realized.However,the high computational resources and model complexity of the proposed algorithm exceed the capabilities of mobile and embedded devices.To address this problem,this paper optimizes the algorithm from the perspective of lightweight,and proposes an insulator defect detection algorithm based on lightweight model.Firstly,the lightweight model Mobile Networks(Mobile Net)is used to improve the IPSPNet,and the focal loss is used to optimize the IPSPNet loss function to generate a lightweight insulator segmentation model for insulator segmentation.Then,the lightweight model Mobile Netv2 and depthwise separable convolution are used to lightweight target detection algorithm based on regression to generate a lightweight insulator defect detection model.Finally,the two lightweight models are combined to detect insulator defects.The results show that the insulator defect detection algorithm based on the lightweight model significantly reduces the model parameters under the dual conditions of ensuring precision and speed,so as to reduce the complexity of the model.Insulator defect detection algorithm based on Double-M2 Det network,based on IPSPNet-SSD network and based on lightweight model are proposed respectively in order to optimize the detection precision,detection speed and model complexity.The test results show that the three algorithms proposed in this paper improve the precision of defect detection,accelerate the training and detection speed,reduce the complexity of the model,and realize the lightweight of the defect detection algorithm. |
