Research On Hydrophobicity Live Detection Of Composite Insulators Based On Deep Learning

Composite insulators are widely used in electric power systems because of their good performance,however,the organic material of their umbrella skirts made of silicone rubber will age with the growth of their service life,which poses a potential threat to the safe operation of power grids.The existing traditional manual detection method for hydrophobicity of composite insulators based on the water spray classification method has problems such as high safety risk and low efficiency,and the discrimination results are greatly influenced by human subjective factors.The digital image processing methods for composite insulators’ hydrophobicity classification have been researched but have poor robustness and other problems.Therefore,this paper focuses on a hydrophobicity detection device based on deep learning technology and UAV platform,which can greatly improve the efficiency of hydrophobicity detection of composite insulators and reduce the labor cost.The process of composite insulator hydrophobicity detection proposed in this paper is that the UAV carries a water spraying device to spray water on the composite insulator,and the camera device transmits the video stream to the edge c omputing platform for inference calculation in real time.The umbrella skirt positioning model locates the umbrella skirt area,and then crops and extracts the water trace area on the umbrella skirt according to the location information of the positioning frame and the cropping coefficient,passing the water trace area image to the image classification model for hydrophobicity level discrimination diagnosis to realize the edge calculation of composite insulator hydrophobicity charged detection,The operati on and inspection personnel can read the results of composite insulator hydrophobicity to be tested in real time.Firstly,to implement models for deployment at the edge with limited computational resources,two lightweight versions of the YOLO target dete ction algorithms were selected,and the YOLOv3-tiny and YOLOv4-tiny models were trained based on the constructed umbrella skirt dataset,and the results were compared on the test set in order to compare the recognition and localization effects of the models,showing that the YOLOv4-tiny model has better localization recognition effects,and has greater precision,recall,and average Io U of the predicted umbrella skirt region to the real region.Then,the Efficient Net algorithm,which combines speed and accu racy,was selected for the hydrophobicity classification network.A data set containing seven hydrophobicity classes was constructed by photographing and collecting hydrophobicity images of composite insulators and data enhancement methods.Based on the constructed dataset,six algorithmic networks of different depths and widths of Efficient Net-b0～b5 were trained with the transfer learning method,and the algorithmic network with the highest accuracy,Efficient Net-b3,was selected after comparison on the test set.And the balance of speed and accuracy of the model was further demonstrated by comparison with other classical convolutional neural networks.Finally,an unmanned airborne water spray device was designed.The selection and assembly of the UAV,water spray gun,water pump and water tank were realized,and the YOLOv4-tiny-based umbrella skirt localization algorithm model and the Efficient Net-b3-based hydrophobicity classification diagnosis model were deployed on the edge computing platform Jetson Xavier NX,with the average inference speed of the models being 17.30 FPS and 7.74 FPS,respectively.The test results on the unmanned water spray device and the combined deep learning algorithm model at the edge prove the feasibility and effectiveness of the deep learning-based hydrophobicity detection method for composite insulators designed in this paper.