Recognition method of hydrophobicity grade of composite insulator based on RS-RepVGG-B0

Abstract

Hydrophobicity class (HC) of composite insulator is the evaluation index of its pollution flashover resistance. In order to identify the hydrophobicity level of composite insulators accurately, a method based on RS-RepVGG is proposed in this paper. The algorithm is based on RepVGG-B0. Firstly, during training, a residual branch across the nonlinear layer is added to RepBlocks to solve the depth constraint problem of RepVGG model without affecting the calculation efficiency of the model. Secondly, the reserved and merged (RM) operations are used to remove the added residual connections equivalently and the re parameterization method is used to decouple the network branches with VGG reasoning structure. Finally, based on this model, spectral nonlocal block (SNL) is introduced to replace the se attention mechanism module at stage2 in the original network, so that the network can capture long-range correlation more flexibly and robustly. The test was carried out in the water spray image data set of composite insulator of Christos-Christodoulos A. Kokalis team. Compared with repvgg-b0, rs-repvgg-b0 increased the model volume by 12.6%, but the floating-point calculation ability of the model was improved by 7.7%, the average recognition accuracy was improved by 2.72%, reaching 97.86%, and the uncertainty of misjudged sample results was not more than ± 1 HC level, which was in line with IEC TS 62073 standard. The experimental results show that, compared with many other existing methods, the RS-RepVGG based composite insulator hydrophobicity grade recognition method proposed in this paper has the advantages of high recognition accuracy, small size and strong floating-point computing ability, and meets the requirements of composite insulator hydrophobicity grade determination.