Research On Insulator Self-explosion Defect Detection Based On Deep Learning

Insulator is an important part of transmission line,which plays the role of support and electrical insulation.Due to its long-term exposure in the complex natural environment,it is easily to cause self-explosion and other faults,leading to the interruption of power supply in the whole area,so it is necessary to detect its status in time.With the development of technology,UAV aerial inspection has replaced manual inspection as the main way of transmission line.The efficiency of inspection image acquisition is greatly enhanced and the traditional image recognition method to detect insulator self-explosion fault has been difficult to meet the demand.Therefore,the target detection method based on deep learning is introduced into the insulator self-explosion fault detection.The specific tasks is as follows.Aiming at the problem of small target size of self-explosive defect blocks and insufficient defect samples,a cascade network of "detector + classifier" is first proposed to detect the self-explosion fault of insulators in aerial images.The detector and classifier are implemented based on CNN.Among them,the detector uses Faster R-CNN with Alexnet as the skeleton network to identify the whole insulator string,on this basis,a classifier is used to distinguish the self-explosive insulator insulator from the normal insulator.Comparing the performance of the three classifiers of Alexnet,VGG-16,and Resnet-101,it is found that Alexnet has the weakest performance and the performance of VGG-16 and Resnet-101 classifiers is similar,but the training time of VGG-16 is shorter.Due to the small number of self-explosive insulator samples,the classification experiment needs to be expanded to enhance the data set.Comparing the enhancement effects of various traditional enhancement methods such as flipping,rotating,sharpening,and adding noise,it is found that the use of flipping,rotating,sharpening and other methods can enhance the data set to a certain extent,but adding noise will weaken the effect.From an overall point of view,the enhancement effect of traditional enhancement methods still cannot meet the demand.In response to this problem,the Generative Adversarial Network(GAN)is introduced to enhance the data set.The loss function of the classical deep convolution generative adversarial network(DCGAN)is improved,and EM distance is used to replace JS divergence to make it more stable.The experimental results show that the model can generate high-quality defect samples and extend them to the original data set,which can improve the performance of CNN more significantly.On the basis of the enhanced data set,the feature fusion strategy is introduced to enhance the original Faster R-CNN detector.The shallow features and deep features are fused to retain more details and enhance the detection ability of the detector for small size targets.The improved detector is used to directly locate the insulator self-explosive defect block from the map,and realize single-stage detection to avoid the tedious work of the cascade network.Experimental results show that the enhanced detector can directly use the insulator self-explosive defect block as the detection target,and the missed alarm rate and false alarm rate can be reduced to a low level.In addition to detecting self-explosive defect blocks of insulators,this strategy also has certain reference significance for detecting other small-sized targets.