Scholarlink

<p>Deep learning approaches are the state-of-the-art in image classification. However, they require representative data for training, and acquiring such data is not always feasible, due to rarity, high costs, environmental hazards, or other restrictions. This paper addresses one such scenario: recognizing the state of high-voltage disconnect switches in transmission substations in a noninvasive manner, based on images captured by regular surveillance cameras. The impossibility of maneuvering switches without affecting substation services, along with strict regulations that limit when and how such maneuvers can occur, makes any real image dataset for that problem severely imbalanced. To address this, we build three-dimensional models of substations, from which we render a large number of synthetic images of open and closed switches, from different camera viewpoints, while also randomizing lighting and other simulation properties. In our experiments performed at two real transmission substations, models trained on synthetic data along with a small number of real samples were able to correctly identify the switch state in over 98% of the test cases, consisting of real photos showing switches of four different types from various viewpoints. That way, we show how synthetic data can be explored to effectively deal with a real-world scenario that poses obstacles for usual data acquisition practices.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Exploring Synthetic Data for Training Deep Learning Models for High-Voltage Disconnector State Identification

  • 期刊论文

  • 发表日期：

    2025年8月26日

查看全文
  • Victor Barpp Gomes,
  • Bruno Marchesi,
  • Yuri Arnold Gruber,
  • Lourival Lippmann Jr.,
  • Marcos Scremin,
  • Bogdan Tomoyuki Nassu

摘要

Deep learning approaches are the state-of-the-art in image classification. However, they require representative data for training, and acquiring such data is not always feasible, due to rarity, high costs, environmental hazards, or other restrictions. This paper addresses one such scenario: recognizing the state of high-voltage disconnect switches in transmission substations in a noninvasive manner, based on images captured by regular surveillance cameras. The impossibility of maneuvering switches without affecting substation services, along with strict regulations that limit when and how such maneuvers can occur, makes any real image dataset for that problem severely imbalanced. To address this, we build three-dimensional models of substations, from which we render a large number of synthetic images of open and closed switches, from different camera viewpoints, while also randomizing lighting and other simulation properties. In our experiments performed at two real transmission substations, models trained on synthetic data along with a small number of real samples were able to correctly identify the switch state in over 98% of the test cases, consisting of real photos showing switches of four different types from various viewpoints. That way, we show how synthetic data can be explored to effectively deal with a real-world scenario that poses obstacles for usual data acquisition practices.