Scholarlink

<p>Rail fastener defects threaten track integrity and operational safety, making reliable automated inspection essential. This study develops an energy-efficient real-time railway fastener detection framework for UAV-based monitoring by integrating Spiking Neural Networks (SNN) into a modified Spiking-YOLOv8 architecture. Unlike conventional CNNs, the event-driven SNN processes only informative visual changes, reducing redundant computation, and achieving a better accuracy–energy balance for on-board inspection. A dataset of 173 images of ballasted railway tracks captured by UAVs was used for data augmentation, resulting in a total of 2,061 training images. The dataset is used to compare SNN and CNN performance under identical conditions, measuring accuracy, latency, and energy per inference. Results show that SNN achieves near CNN accuracy (mAP@0.5 = 0.975 vs. 0.995) while reducing energy consumption by about 80% (0.50&#xa0;J vs. 2.50&#xa0;J) and maintaining real-time inference speed (~ 34ms per frame). The approach demonstrates a 4.9 times improvement in accuracy-per-joule, supporting longer UAV endurance for inspection and more autonomous railway inspections.</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

Energy-efficient, real-time detection of railway fastening systems from drone-based imagery using spiking neural networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月21日

查看全文
  • Sakdirat Kaewunruen,
  • Zibo Chen,
  • Aryadhatu Dhaniswara,
  • Zahid Hamarat

摘要

Rail fastener defects threaten track integrity and operational safety, making reliable automated inspection essential. This study develops an energy-efficient real-time railway fastener detection framework for UAV-based monitoring by integrating Spiking Neural Networks (SNN) into a modified Spiking-YOLOv8 architecture. Unlike conventional CNNs, the event-driven SNN processes only informative visual changes, reducing redundant computation, and achieving a better accuracy–energy balance for on-board inspection. A dataset of 173 images of ballasted railway tracks captured by UAVs was used for data augmentation, resulting in a total of 2,061 training images. The dataset is used to compare SNN and CNN performance under identical conditions, measuring accuracy, latency, and energy per inference. Results show that SNN achieves near CNN accuracy (mAP@0.5 = 0.975 vs. 0.995) while reducing energy consumption by about 80% (0.50 J vs. 2.50 J) and maintaining real-time inference speed (~ 34ms per frame). The approach demonstrates a 4.9 times improvement in accuracy-per-joule, supporting longer UAV endurance for inspection and more autonomous railway inspections.