Scholarlink

<p>Dissolved Gas Analysis (DGA) is essential for transformer fault diagnosis but is constrained by limited data volume and severe class imbalance, leading to diagnostic inaccuracies. Based on the DGA raw data following a multivariate Gaussian distribution after logarithmic transformation, this paper proposes a novel framework combining Gaussian Copula data augmentation with deep transfer learning to overcome these limitations. Key steps include: (1) Logarithmic transformation of DGA raw data to approximate Gaussian distributions; (2) Synthesis of balanced, statistically representative DGA logarithmic data using Gaussian Copula modeling; (3) Fault classification via fine-tuning a pre-trained deep neural network on the augmented dataset. Comparative evaluation against various machine learning, deep learning, augmentation, and transfer learning methods demonstrates the proposed approach achieves <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\( &gt; 97\)</EquationSource> </InlineEquation>% diagnostic accuracy, with significant performance gains particularly for minority fault classes. The method effectively resolves data scarcity and imbalance challenges in transformer DGA diagnosis.</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

DGA Fault Diagnosis Method Based on Gaussian Copula Data Augmentation and Transfer Learning

  • 期刊论文

  • 发表日期：

    2025年8月20日

查看全文
  • Hao Tang,
  • Chang Liu,
  • Xiang Zhang,
  • Xiaoqing Li,
  • Xiaofei Zhang

摘要

Dissolved Gas Analysis (DGA) is essential for transformer fault diagnosis but is constrained by limited data volume and severe class imbalance, leading to diagnostic inaccuracies. Based on the DGA raw data following a multivariate Gaussian distribution after logarithmic transformation, this paper proposes a novel framework combining Gaussian Copula data augmentation with deep transfer learning to overcome these limitations. Key steps include: (1) Logarithmic transformation of DGA raw data to approximate Gaussian distributions; (2) Synthesis of balanced, statistically representative DGA logarithmic data using Gaussian Copula modeling; (3) Fault classification via fine-tuning a pre-trained deep neural network on the augmented dataset. Comparative evaluation against various machine learning, deep learning, augmentation, and transfer learning methods demonstrates the proposed approach achieves \( > 97\) % diagnostic accuracy, with significant performance gains particularly for minority fault classes. The method effectively resolves data scarcity and imbalance challenges in transformer DGA diagnosis.