Abstract
Power transmission lines use high-voltage insulators for both electrical insulation and mechanical support. In normal operating conditions, insulators are often exposed to outdoor contamination and stress from different weather conditions. These factors can lead to insulation malfunctions and power losses. Consequently, electricity grid operating companies must conduct regular check-ups and evaluations to minimize possible power outages. This paper introduces a new method that uses deep convolutional neural networks (CNNs) for high-voltage insulator surface image classification. Given the effective use of CNNs in computer vision, numerous pre-trained models are available for various classification tasks. Typically, researchers fine-tune these models on a specific dataset, and a single model is selected and used for the final prediction. In contrast with this “winner-take-all” approach, in this study, an ensemble of pre-trained CNNs is proposed to classify the type of pollution accumulated on the insulator surface. This focus is warranted because each pre-trained convolutional base operates as a feature extractor, and thus a combination of them as proposed in this study functions as a composite feature extractor. Our numerical results clearly show the superiority of the proposed ensemble technique over the common “winner-take-all” approach for classifying insulator surface contamination.