Identification And Defect Detection Of Aerial Insulators Using Lightweight Network

Insulation plays a crucial role in the safety of power system,and intact insulator strings on high-voltage transmission lines are the key to ensure insulation performance.In order to judge whether the insulator strings are intact,it used to be common to manually climb the high-voltage piezoelectric tower for inspection,but now the development of intelligence makes the inspection way more convenient.The hardware equipment of UAV is combined with the automatic detection and recognition function of convolutional neural network in deep learning for inspection.Due to the Angle and complex environment of UAV aerial photography of insulator images,the common convolutional neural network is not suitable for insulator identification and defect detection,no matter in terms of detection accuracy or detection speed.Therefore,this paper preprocesses the original data set according to the morphological characteristics of insulators and takes SSD(Single Shot Multi Box Detector)target detection network as the basic model to improve the research from three aspects.To make its detection effect on insulators better,the specific improvement research is as follows:(1)In view of the characteristics of backbone network of SSD model,which has high accuracy in public data set but low accuracy in insulator identification,as well as large number of parameters and calculation amount,this paper designs a lightweight backbone network suitable for insulator identification according to the principle of lightweight network proposed in recent years,which greatly reduces the number of parameters and calculation amount of backbone network.And the experiment proves that the detection accuracy and speed are improved.(2)Aiming at the feature that the model loses part of the global information due to the depth divisible volume in lightweight networks,this paper increases the attention mechanism to improve the model detection accuracy.Based on the characteristics of the attention mechanism,we designed the Enhanced Channel Attention Module(ECAM)and the Enhanced Spatial Attention Module(ESAM),and constructed the Teeth Attention Module(TAM)according to the form of channel-spatial-channel.The effectiveness of the Teeth Attention Module to improve the accuracy of the model was proved by experiments.(3)In view of the morphological characteristics of complete insulator strings and defective insulator strings in the data set,this paper uses K-MEANS clustering to perform cluster analysis on insulators in the data set.The clustering results are compared with the default box of the original SSD model.After determining the morphological characteristics of insulators in the data set,the default box of the original SSD model is improved.The generated anchor frame fits the shape characteristics of insulators better,so as to improve the detection accuracy of insulators in the image.(4)Based on the above three directions of improvement research and fusion with the original SSD model,and constructs the lightweight network model of the fusion SSD.Through the comprehensive comparison with the original SSD model and various single improved SSD models,it proves the high efficiency of the fusion SSD lightweight network model in terms of detection accuracy,speed,memory consumption and other indicators.