Research On Visual Inspection Technology Of Transmission Line Insulator Defects Based On Deep Learning

Insulators are a significant part of transmission lines.Perceiving the insulators through inspections is of great significance to the operation and maintenance of power systems.In recent years,intelligent power inspections that combine deep learning and UAV application technology have become a new trend.However,the images taken by UAV have the problem of diverse insulator string sizes and difficulty to detection,and the lack of insulator defect samples will also reduce the training effect of the neural network and affect the identification ability of the model.Therefore,designing and training suitable networks to complete the detection and defect-recognition of the insulator has important research significance.The specific research content of this thesis is as follows:(1)The basic conditions of porcelain,tempered glass,and synthetic silicone rubber insulators used in power transmission lines are introduced,and their usage and common faults are analyzed.According to the characteristics of traditional inspection,a two-stage defect detection method is proposed,which detects the insulator string and the single defective insulator caps in the insulator string step by step.(2)Aiming at the problem of insufficient detection accuracy due to the large aspect ratio and various tilt angles of the insulator string in the inspection image taken by the UAV,an Inclined Insulator String Detector(IISD)based on the Oriented Bounding Box(OBB)is proposed.The Feature Pyramid Network(FPN)is used to extract features of the image,and the direction angle parameter is introduced in the regression process of the bounding box,and the network training is completed by designing a reasonable loss function.The comparison results with the one-stage object detection network using the horizontal bounding box confirm that the IISD proposed in this thesis can more accurately locate the inclined insulator string in the inspection image,and the effect of filtering the background is great.(3)Given the insufficient image samples of defective insulators,the model is prone to overfitting.Based on the study of Feature Reweighting Network(FRN)in Few-Shot Learning(FSL),a Multi-scale Feature Reweighting Network(MFRN)was designed.And the model can extract more subtle defect features from multiple scales and reweight the feature maps of the query.By adopting the 5-Way 5-Shot pre-training and fine-tuning training strategy,the recognition of five common defects of insulators such as the falling of the porcelain insulator caps or the self-explosion and falling of the glass insulator caps has been realized.The effect of the MFRN is verified by the experiment.And the two-stage defect detection method is verified by ablation experiments.In summary,the thesis takes the insulator in the UAV inspection images as the research object and uses CNN,object detection,FSL,and other methods to study the insulators defect detection problem under the condition of complex background and scale,as well as the insufficient samples.Experimental results show that the method and detection frameworks proposed in this thesis have famous effects.