Self-blast state detection of glass insulators based on stochastic configuration networks and a feedback transfer learning mechanism

Abstract

The self-blast state of a glass insulator directly affects the safety and reliability of transmission lines. To address the insufficient generalization ability of existing detection methods for insulator self-blast states and the drawbacks of deep neural network structures, the theories of transfer learning and closed-loop control are drawn upon to provide an intelligent detection method for the self-blast states of glass insulators. The method proposed in this paper is based on stochastic configuration networks and a feedback transfer learning mechanism. First, to reduce the redundancy of convolutional kernels in the channel extent, the interleaved group convolution strategy is employed to reconstruct the convolutional layers of the Inception network. Second, in view of the different feature applicabilities of different glass insulator images and based on the adaptive convolution module groups, the data structure of the dynamic feature space of insulator images is built with a certain mapping relationship from global to local. Then, the discriminative measure index is used to evaluate the discriminative information of the feature space to enhance the interpretability of the compact feature spance. Third, the fully connected feature vector of the compact feature space is sent to stochastic configuration networks (SCNs), which have universal approximation property to establish the classification criteria of the self-blast states of insulator images with generalization ability. Finally, an imitation of human thinking patterns is employed that exhibits repeated deliberation and comparison. Consequently, based on generalized error and entropy theories, the evaluation index of the objective function is established to evaluate the uncertain detection results of the self-blast states of glass insulator images in real time. Then, the dynamic transfer learning mechanism is constructed based on the constraint of the measurement index of uncertain detection results to realize self-optimizing regulation of the feature space that exhibits multihierarchy and discrimination and reconstructed classification criteria. The experimental results show that compared with other algorithms, the proposed method enhances the generalization ability and detection accuracy of the model.