Scholarlink

<p>Reliable operation of electrical transmission lines is vital to power-system stability, but fault diagnosis is hindered by noise, environmental variation, equipment failures, and the distance between generation and load centers. To improve fault identification under these conditions, this study constructs and proposes a multiclass version of the feature space partition (FSP) framework specifically designed to operate on and classify multiclass datasets. The proposed FSP is an appropriate pattern-identification methodology based on a two-phase local–global strategy. First, it performs unsupervised segmentation of the feature space to identify structurally similar regions. Next, supervised multiclass models are trained within each region, allowing decisions to adapt to local patterns. Region quality is assessed using Cauchy–Schwarz divergence, enabling adaptive partitioning and improved separation of subtle fault signatures, even under noisy or incomplete data conditions. Unlike one-vs-rest or one-vs-one schemes, FSP learns all classes jointly, producing more consistent decision boundaries, shared feature use across classes, lower computational cost, and better behavior under class imbalance. The pipeline also includes robust preprocessing and noise-aware training, with dropout and simulated fault scenarios. On a benchmark transmission-line fault dataset, FSP achieved 83.03% multiclass accuracy, outperforming k-nearest neighbors, decision trees, support vector machines, and standard neural networks. These results indicate that FSP methodology applied in this work is a robust and scalable multiclass solution for safer and more resilient grid operation.</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

A Multiclass Feature Space Partition Learning Framework for Transmission-Line Fault Typing

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月15日

查看全文
  • Bruno Alves do Carmo,
  • Marcello Goulart Teixeira,
  • Elizabeth Fialho Wanner,
  • Carolina Gil Marcelino

摘要

Reliable operation of electrical transmission lines is vital to power-system stability, but fault diagnosis is hindered by noise, environmental variation, equipment failures, and the distance between generation and load centers. To improve fault identification under these conditions, this study constructs and proposes a multiclass version of the feature space partition (FSP) framework specifically designed to operate on and classify multiclass datasets. The proposed FSP is an appropriate pattern-identification methodology based on a two-phase local–global strategy. First, it performs unsupervised segmentation of the feature space to identify structurally similar regions. Next, supervised multiclass models are trained within each region, allowing decisions to adapt to local patterns. Region quality is assessed using Cauchy–Schwarz divergence, enabling adaptive partitioning and improved separation of subtle fault signatures, even under noisy or incomplete data conditions. Unlike one-vs-rest or one-vs-one schemes, FSP learns all classes jointly, producing more consistent decision boundaries, shared feature use across classes, lower computational cost, and better behavior under class imbalance. The pipeline also includes robust preprocessing and noise-aware training, with dropout and simulated fault scenarios. On a benchmark transmission-line fault dataset, FSP achieved 83.03% multiclass accuracy, outperforming k-nearest neighbors, decision trees, support vector machines, and standard neural networks. These results indicate that FSP methodology applied in this work is a robust and scalable multiclass solution for safer and more resilient grid operation.