Scholarlink

<p>Power Grid Line Fault (PGLF) detection is critical for uninterruptible power supply and grid resiliency. However, existing object detection algorithms achieve unsatisfactory results in PGLF detection, since (a) subtle line anomalies (e.g., broken and loose strands) require more sensitive IoU loss for these fine details, and (b) the targets are easily susceptible to background objects, such as hanging kites, cloth, and sticks, making it necessary to capture comprehensive global context. To solve these issues, we propose a modified RT-DETR framework with Null Intersection over Union (NuIoU) loss and TranCFormer module (NT-RTDT) for real-time PGLF identification. First, we propose NuIoU loss to focus on shape and positional offsets between boxes, enhancing sensitivity to small anomalies. Second, TranCFormer with lightweight attention mechanism is introduced to enrich global context across the entire image during channel compression. To verify the effectiveness of proposed method, we collect and present a new Power Grid Drone Inspection (PGDI) dataset. The experimental results show that the proposed method achieves 38.97 AP<sup>5095</sup>, outperforming other mainstream real-time detectors. Source code and pre-trained models are available at <a href="https://github.com/he13689/NTRTDT">https://github.com/he13689/NTRTDT</a>.</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

Real-time grid line fault detection with null IOU loss and TranCFormer

  • 期刊论文

  • 发表日期：

    2026年2月28日

查看全文
  • Yiqing He,
  • Tiantian Li,
  • Zhuowei Wang,
  • Yunyun Zhang,
  • Chuxiu Guo,
  • Lianglun Cheng

摘要

Power Grid Line Fault (PGLF) detection is critical for uninterruptible power supply and grid resiliency. However, existing object detection algorithms achieve unsatisfactory results in PGLF detection, since (a) subtle line anomalies (e.g., broken and loose strands) require more sensitive IoU loss for these fine details, and (b) the targets are easily susceptible to background objects, such as hanging kites, cloth, and sticks, making it necessary to capture comprehensive global context. To solve these issues, we propose a modified RT-DETR framework with Null Intersection over Union (NuIoU) loss and TranCFormer module (NT-RTDT) for real-time PGLF identification. First, we propose NuIoU loss to focus on shape and positional offsets between boxes, enhancing sensitivity to small anomalies. Second, TranCFormer with lightweight attention mechanism is introduced to enrich global context across the entire image during channel compression. To verify the effectiveness of proposed method, we collect and present a new Power Grid Drone Inspection (PGDI) dataset. The experimental results show that the proposed method achieves 38.97 AP5095, outperforming other mainstream real-time detectors. Source code and pre-trained models are available at https://github.com/he13689/NTRTDT.