Scholarlink

<p>This study demonstrated an improvement in partial discharge (PD) recognition for underground cable joints by saturating the images. The recognition rate increased as the phase-resolved PD (PRPD) images transitioned from sparse to saturated through the accumulation of more PD signals. Difference hashing (dHash) was used to calculate similarity scores among sequentially accumulated images. Furthermore, an image was identified as that corresponding to a saturated PRPD when its similarity score exceeded a set threshold. For validation, a PD dataset containing three different types of defects in cable joints was used to train a convolutional neural network model. Additionally, four thresholds—0.9, 0.95, 0.9925, and 0.997—were tested. The threshold of 0.95 is recommended as it effectively balances accuracy with power cycle efficiency, yielding accuracy rates of 93.12%, 95.14%, and 90.84% for defects A, B, and C, respectively. The proposed method effectively improves pattern recognition accuracy, particularly in scenarios with unclear or sparse measured data.</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

Enhancing PRPD patterns to improve defect recognition in real-time measurements

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月28日

查看全文
  • Chien-Kuo Chang,
  • Supawit Mahajak

摘要

This study demonstrated an improvement in partial discharge (PD) recognition for underground cable joints by saturating the images. The recognition rate increased as the phase-resolved PD (PRPD) images transitioned from sparse to saturated through the accumulation of more PD signals. Difference hashing (dHash) was used to calculate similarity scores among sequentially accumulated images. Furthermore, an image was identified as that corresponding to a saturated PRPD when its similarity score exceeded a set threshold. For validation, a PD dataset containing three different types of defects in cable joints was used to train a convolutional neural network model. Additionally, four thresholds—0.9, 0.95, 0.9925, and 0.997—were tested. The threshold of 0.95 is recommended as it effectively balances accuracy with power cycle efficiency, yielding accuracy rates of 93.12%, 95.14%, and 90.84% for defects A, B, and C, respectively. The proposed method effectively improves pattern recognition accuracy, particularly in scenarios with unclear or sparse measured data.