Silicone Composite Insulator Localization and Condition Assessment Using Weakly Supervised Learning in the Power IIOT

The assessment of the hydrophobicity of silicone composite insulators, a key component of power equipment, is an important task in the maintenance of power systems. Due to the complex working conditions and convolutional network characteristics, the problems of image distortion and oversized model structure occur in the detection process. To solve these problems, this paper proposes a weakly supervised learningbased insulator detection model. The method consists of three parts: image enhancement module, feature extraction backbone, and hydrophobicity grade evaluation. Among them, the image enhancement module solves the problem of a small number of aerial images by Gray Rotating Mosaic algorithm. The feature extraction backbone uses a lightweight residual network oriented to the a priori information of the image, combined with a convolutional block attention mechanism module, which improves the computationally intensive problem of traditional convolutional networks. At the same time, the feature information at the edge of the water trace is enhanced by the Gray-scale Balanced Segmented linear transformation module, and the accurate assessment of the hydrophobicity class is realized based on the Comparative Intersection over Union loss function calculation. The experimental results show that the method can effectively realize the long-distance nondestructive detection of insulators through the combination of edge computing devices and servers. In addition, by taking advantage of weakly supervised learning, the model can also provide the necessary location information for insulator string detection.