The reliability of electric power systems is largely determined by the insulation condition of electrical equipment. Insulation damage can lead to power losses, reduced service life of lines and devices, and emergency shutdowns, so insulation diagnostics is critical to prevent technological disruptions. However, traditional approaches to insulation monitoring are often labor-intensive and subjective. In this regard, the role of computer vision and deep learning methods, capable of automatically detecting insulation defects and thereby increasing the efficiency and objectivity of monitoring, is increasing. This study considers the application of modern architectures of deep convolutional neural networks for the problem of recognizing insulating elements of electrical equipment. Particular attention is paid to the comparative analysis of several state-of-the-art models. The considered architectures show effective results and provide deep multi-scale analysis of scene features based on convolutional networks. In this paper, the models are used in conjunction with image augmentation algorithms. Data augmentation allows you to artificially expand limited sets of training images through various transformations, which is especially important for a small dataset. The application of these methods is aimed at improving the quality of training data and reducing the risk of overfitting models, as well as overcoming the imbalance of classes in the sample by generating additional fault samples. The proposed approach includes conducting a sequential comparative experiment on a small and limited set of image data from power facilities. A comparison was made of the accuracy and completeness metrics of various neural network architectures when combining various augmentation strategies in order to identify a combination of models and data augmentation methods that provide the highest recognition accuracy. The results of the study will help determine the most effective augmentation models and techniques suitable for real-life operating conditions at power facilities, taking into account complex backgrounds, variable lighting, and different angles of equipment shooting. Identifying such optimal solutions based on deep learning is intended to improve the reliability and efficiency of automated insulation monitoring in the power industry.

Keywords: computer vision, convolutional neural networks, isolation, defect, data augmentation, machine learning, energy, automation of image analysis