DEFECTS DETECTION AND INSULATOR CLASSIFICATION FOR POWER DISTRIBUTION NETWORKS USING DEEP LEARNING
Visual evaluation, deep learning, electrical network insulators.
Overhead Power Distribution Lines (OPDL) correspond to a large percentage of the medium-voltage electrical systems available worldwide. In these networks, Visual Inspection activities (VI) are usually performed without resorting to automated systems, requiring a significant investment of time and human resources. VI are also vulnerable to subjective evaluation and human error, which can lead to incorrect and/or inaccurate inspection results. In this perspective, the present PhD study proposes to introduce a set of intelligent techniques aimed at vision-based automatic inspection of medium voltage OPDL, aiming at the components identification as well as the diagnosis of defects visible to the naked eye. For this purpose, different computational models were developed based on Digital Image Processing (DIP) and Artificial Intelligence (AI) techniques. As an additional part of this research, it was created an image database of OPDL components collected in a photographic studio and from a realistic OPDL created outdoors. To optimize the computational models performance, different types of intelligent algorithms and deep learning techniques were evaluated, with emphasis on Convolutional Neural Networks (CNNs), hybrid configurations and frameworks of Generative Adversarial Networks (GANs). In addition, other approaches were implemented, such as: data augmentation; transfer learning and Multi-Task Learning (MTL); background invariance and Image Generators (IG). The preliminary results indicate satisfactory performance of the proposed techniques in class identification and defect detection of distribution insulators, providing a series of innovations in relation to other available solutions.