Research On Intelligent Cognition Method Of Self-Exploding State Of Glass Insulator Based On Deep Migration Learning

Insulator is a kind of special insulation control,which has the function of line support and electrical insulation.It is an important component in overhead transmission lines.At the same time,it is also a component with many faults during the circuit transmission.Problems such as damage,contamination,and self-explosive will affect the safe operation of the transmission line and cause incalculable losses.Therefore,with the increasing scale of high-voltage transmission lines,the tasks of manual line inspection and visual inspection of insulator failures are becoming increasingly burdensome,and it is necessary to perform real-time detection on the operating status of insulators.Therefore,in this paper,in view of the deficiency of the generalization ability of the open-loop model and the defects of the deep neural network structure in the existing insulator self-detonation state detection method,this paper imitates the human cognitive model,draws on the transfer learning and feedback mechanism,and explores a deep transfer learning-based intelligent detection method for self-detonation state of glass insulator.The main work of this paper is as follows:(1)In view of the large amount of field environment information in the aerial image of insulators,and the situations where drones or helicopters encounter turbulence and jitter during aerial photography and are at different shooting angles,input data preprocessing and subsequent accurate positioning self-explosion are provided for the insulator state recognition model,based on the YOLO-v3 algorithm,the insulators in the complex aerial insulator images are located,and the rotation transformation and normalization preprocessing are performed to provide an insulator image database for subsequent state recognition.(2)In order to solve the problem of large amount of computation caused by the extraction of redundant feature maps in convolutional neural networks,this paper introduces interleaved group convolutional networks to reduce the channel redundancy of convolution operations and build deeper on the basis of higher-order abstract feature space The hierarchical feature space data structure realizes the conversion of input image channel spatial information to multi-level high-order abstract feature information.For the insulator feature images extracted by the convolution module group,there is a feature map space to determine the mapping relationship,and a distinguishability measure index is constructed to evaluate the distinguishability of each feature map for multi-category insulator images,and a strong distinguishable and simple differentiation is established.Feature map space.(3)Aiming at the defect of insufficient generalization ability of softmax layer in the deep neural network,the random configuration network classifier is used to adaptively set the parameter range and generate a new random basis function under the constraints of random input weights and offsets according to the output residuals ensure the ability of the learning model to approximate all situations.In the learning process of deep neural networks,based on the alternating optimization strategy of forward propagation and backward update,the parameters of the deep learning network are iteratively updated globally to adapt to the diverse data sets obtained at the input and generate a generalized feature space data structure and its classification criteria.(4)Aiming at the traditional method of a posteriori statistics that can not evaluate the uncertain cognitive results of the insulator image in real time,it imitates the human information interactive cognitive model of repeated comparison and comparison from the whole to the part.Based on the generalized error and entropy theory,a target optimization function evaluation index based on semantic error information entropy is established,real-time evaluation of the reliability of the uncertain cognitive results of the insulator state,dynamic update of the number of convolution layers,based on the transfer learning mechanism to achieve the multi-level differentiated feature space of the insulator image and its classification estimation criteria optimize and reconstruct,re-cognize the image state of insulators with low reliability.Experimental results show that compared with other open-loop and closed-loop algorithms,this method enhances the generalization ability of the model and improves the cognitive accuracy of the model.