Scholarlink

<p>To address the limitations of conventional path planning methods in dynamic wind fields, particularly their inadequate adaptability and constrained spatiotemporal modeling capabilities, this study proposes a deep reinforcement learning network architecture termed SMA-DDQN, which integrates the Mamba architecture with a dual-channel attention mechanism. This network synergizes the Mamba architecture’s long-sequence modeling strengths with the attention mechanism’s key features—including dynamic focusing and adaptive weighting to construct a deep reinforcement learning network with robust temporal modeling capacity and spatial feature perception. Furthermore, by constructing phased guidance signals using flight path nodes to alleviate sparse rewards, combined with a multi-phase decaying <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\varepsilon \)</EquationSource> </InlineEquation>-greedy exploration strategy and a multi-objective composite reward function balancing energy consumption, safety, and efficiency, a decision-making system harmonizing local and global optimization is established. Simulation experiments and real-world scenario simulations have shown that the algorithm significantly enhances flight stability and path planning efficiency, exhibiting notable advantages in energy consumption control, path accuracy, and system robustness.</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

Uav path planning and flight control based on Mamba and dual-channel attention mechanism under wind field disturbance

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Jianxin Feng,
  • Song Jiang,
  • Jiadong Sun,
  • Zumin Wang,
  • Zhiguo Liu,
  • Yuanming Ding

摘要

To address the limitations of conventional path planning methods in dynamic wind fields, particularly their inadequate adaptability and constrained spatiotemporal modeling capabilities, this study proposes a deep reinforcement learning network architecture termed SMA-DDQN, which integrates the Mamba architecture with a dual-channel attention mechanism. This network synergizes the Mamba architecture’s long-sequence modeling strengths with the attention mechanism’s key features—including dynamic focusing and adaptive weighting to construct a deep reinforcement learning network with robust temporal modeling capacity and spatial feature perception. Furthermore, by constructing phased guidance signals using flight path nodes to alleviate sparse rewards, combined with a multi-phase decaying \(\varepsilon \) -greedy exploration strategy and a multi-objective composite reward function balancing energy consumption, safety, and efficiency, a decision-making system harmonizing local and global optimization is established. Simulation experiments and real-world scenario simulations have shown that the algorithm significantly enhances flight stability and path planning efficiency, exhibiting notable advantages in energy consumption control, path accuracy, and system robustness.