Scholarlink

With power systems advancing to high-voltage and wide-area setups, traditional UAV inspection methods face hurdles in dynamic adaptability and multi-agent coordination. This paper presents an embodied intelligence-driven UAV inspection framework. It features three key innovations: Spatiotemporal multimodal fusion using hardware-synchronized sensors raises submillimeter defect detection accuracy to 94.7% (32.4% improvement).Dynamic Federated Contract Network Protocol (CNP), integrating gradient-isolated federated learning and auction-enhanced task allocation, reduces multi-agent conflicts from 15% to 5% and keeps 85% collaboration efficiency. 4D trajectory planning, combining spatiotemporal corridors (0.5 m3 + 0.1s resolution) and model predictive control, attains 98.5% obstacle avoidance under 50V/m electromagnetic interference. The three-layer architecture enables a 200ms edge-response closed-loop for perception-decision-execution. Field tests in 500kV substations demonstrate a 60% time cut for 10km line inspections, with 280ms cloud latency meeting industry norms. The framework can be extended to Industrial IoT applications like pipeline and renewable energy facility monitoring.

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

Embodied Intelligence-Driven Framework for UAV Power Inspection Systems: Federated Reinforcement Learning for Multi-agent Collaborative Optimization

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Chuanlei Zhang,
  • Chen Zhang,
  • Hongli Cui,
  • Pengfei Li,
  • Yan Wan,
  • Di Sun,
  • Haifeng Fan,
  • Hui Ma

摘要

With power systems advancing to high-voltage and wide-area setups, traditional UAV inspection methods face hurdles in dynamic adaptability and multi-agent coordination. This paper presents an embodied intelligence-driven UAV inspection framework. It features three key innovations: Spatiotemporal multimodal fusion using hardware-synchronized sensors raises submillimeter defect detection accuracy to 94.7% (32.4% improvement).Dynamic Federated Contract Network Protocol (CNP), integrating gradient-isolated federated learning and auction-enhanced task allocation, reduces multi-agent conflicts from 15% to 5% and keeps 85% collaboration efficiency. 4D trajectory planning, combining spatiotemporal corridors (0.5 m3 + 0.1s resolution) and model predictive control, attains 98.5% obstacle avoidance under 50V/m electromagnetic interference. The three-layer architecture enables a 200ms edge-response closed-loop for perception-decision-execution. Field tests in 500kV substations demonstrate a 60% time cut for 10km line inspections, with 280ms cloud latency meeting industry norms. The framework can be extended to Industrial IoT applications like pipeline and renewable energy facility monitoring.