Scholarlink

<p>Missile systems are inherently subject to strong non-linearities, model uncertainties, and flight-condition-dependent external disturbances, which pose significant challenges for autopilot design. To address these issues, this paper proposes a novel angle-of-attack-free, data-driven robust control framework for missile longitudinal autopilot design, formulated as a two-player zero-sum differential game. Unlike conventional three-loop autopilots, the proposed approach does not require angle-of-attack measurements, while still achieving high-performance acceleration tracking. In addition, the method avoids the need for explicit or highly accurate system models by leveraging reinforcement learning to learn optimal control policies directly from operational data, thereby enhancing adaptability to uncertain environments. First, the acceleration tracking problem is formulated as a <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(H_{\infty }\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>H</mi> <mi>∞</mi> </msub> </math></EquationSource> </InlineEquation> control problem based on the missile’s dynamic characteristics, which is subsequently transformed into a zero-sum game formulation. Second, the Nash equilibrium is learned online using Q-learning based on data collected during operation. Moreover, a Robust Integral of the Sign of the Error (RISE) feedback term is incorporated to attenuate matched disturbances while requiring only knowledge of the input matrix. Simulation results validate the effectiveness of the proposed framework, demonstrating its significant improvements in robustness and performance under different complex and critical flight conditions.</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

Data-Driven Q-Learning Based Robust Longitudinal Missile Autopilot Design Without Angle-of-Attack Measurement

  • 期刊论文

  • 发表日期：

    2026年5月4日

查看全文
  • Thang Nguyen-Tat,
  • Nhat-Minh Le-Phan,
  • Nga Thi-Thuy Vu

摘要

Missile systems are inherently subject to strong non-linearities, model uncertainties, and flight-condition-dependent external disturbances, which pose significant challenges for autopilot design. To address these issues, this paper proposes a novel angle-of-attack-free, data-driven robust control framework for missile longitudinal autopilot design, formulated as a two-player zero-sum differential game. Unlike conventional three-loop autopilots, the proposed approach does not require angle-of-attack measurements, while still achieving high-performance acceleration tracking. In addition, the method avoids the need for explicit or highly accurate system models by leveraging reinforcement learning to learn optimal control policies directly from operational data, thereby enhancing adaptability to uncertain environments. First, the acceleration tracking problem is formulated as a \(H_{\infty }\) H ∞ control problem based on the missile’s dynamic characteristics, which is subsequently transformed into a zero-sum game formulation. Second, the Nash equilibrium is learned online using Q-learning based on data collected during operation. Moreover, a Robust Integral of the Sign of the Error (RISE) feedback term is incorporated to attenuate matched disturbances while requiring only knowledge of the input matrix. Simulation results validate the effectiveness of the proposed framework, demonstrating its significant improvements in robustness and performance under different complex and critical flight conditions.