Insulator Defect Detection Based On Deep Learning

In modern society,all production and life are closely related to electricity.Under the huge demand for electricity,the scale of power system is increasing,and EHV transmission lines came into being.For transmission lines,real-time monitoring and maintenance of equipment is the key to ensure power security.However,insulators are prone to defect when they work under harsh natural conditions for a long time.More than half of the power grid outage accidents are caused by insulator defect every year.Therefore,regular fault detection of insulators is of great significance for the normal operation of transmission lines.Firstly,the image is transformed from RGB color space to gray space,which reduces the computer workload without changing its characteristics;Then,histogram equalization processing is carried out on the image to make the distribution of each gray level in the image more uniform,enhance the distinction between the target and the background,and facilitate feature extraction;Finally,the image is filtered.The purpose is to eliminate the isolated noise points on the basis of retaining the target edge information,select the nonlinear filter for processing,then use the median filter and bilateral filter with different template sizes to process the image respectively,evaluate the filtering effect through its peak signal-to-noise ratio(PSNR),process and compare the data of multiple photos,Finally,the convolution kernel with the best effect is 5 X 5 median filter template.Then it introduces and analyzes the network architecture commonly used in the field of target recognition,and mainly compares faster.The detection speed and accuracy of r-cnn family and Yolo are compared,and their error types are analyzed.The detection accuracy of the former is 6.1%higher than that of the latter,but the detection speed of the latter is more than 20 times higher.At the same time,considering the huge amount of patrol data and the development needs of real-time detection,Yolo algorithm is selected as the basic framework of this paper.In the next research,Combined with the characteristics of the detection target,it is improved for its low accuracy.Then,in the aspect of insulator target recognition,firstly,according to the characteristics of different insulator shapes and sizes,the generation algorithm of the initial frame is replaced by the kmeans++algorithm with lower sensitivity to the initial point and more stable results,which better fits the distribution of the target frame of the training set,and improves the map value of the original network from 0.937 to 0.951.Then,before training process,the pre training model parameters on the ImageNet data set are invoked to help the training of the insulator data set.The improved network convergence rate is faster than that of the original network,and the iteration cycle needed to achieve convergence is less than that of the original network.At the same time,for the same number of iterations,the mAP after migration is higher than the original network.It shows that it is easier to make the network converge and the convergence effect is better.The detection rate of the improved network is slightly higher than that of the original network,and the recognition confidence rate is better.In order to improve the generalization performance of the model and simulate the situation that the target is blocked by obstacles,this paper also carries out random erasure processing on the insulator data set,and combines other augmentation means,so that the finally trained model also has good recognition effect on the blocked insulator picture.Finally,several important parameters in the training process are introduced.According to the first training results,the super parameters are fine tuned and the optimal results are obtained.On the basis of insulator target recognition,the defect recognition network is constructed.Firstly,the correctness of five common network models in insulator defect recognition is compared and tested.Using the best densenet feature extraction network algorithm for reference,each convolution layer is tightly connected to improve the feature reuse rate and reduce the amount of calculation.In the aspect of target detection,learning from the idea of FPN algorithm,the fusion of top-level feature map and bottom-level feature map can not only better grasp the global information,but also get more detailed information and improve the accuracy of small target detection;And each fused feature layer is predicted independently.Finally,the indexes of the model are analyzed.The accuracy and accuracy of the model constructed in this paper are better than the original model.Figure[47]table[9]reference[82]...