基于深度强化学习的三维变形机翼反设计方法

An inverse design method for three-dimensional morphing wings based on deep reinforcement learning

  • 摘要: 本文提出了一种基于强化学习的三维变形机翼反设计(RLID)框架,并将其应用于可变工况的自适应变形飞行任务中。选取类别/形状变换函数(CST)设计三维变形机翼,并采用拉丁超立方抽样(LHS)方法对变形设计空间进行抽样,从而获取样本点;通过计算流体力学(CFD)求解得到对应的气动参数,并通过深度置信网络(DBN)代理模型构建从变形设计参数到气动参数的输入-输出模型。针对可变工况环境,基于无监督学习的深度Q网络(DQN)强化学习智能体可为机翼实时提供变形策略,以达到预期的气动性能要求。此外,本文将DQN智能体与基于贪心的条件生成对抗网络(G-CGAN)智能体进行了对比,结果表明,本文所提出的RLID框架在多变工况条件下能够提供可靠的变形策略,且相较于G-CGAN,DQN智能体更注重整体任务的收益。

     

    Abstract: In this study, an RLID (reinforcement learning inverse design) framework is proposed and applied to the reverse design of three-dimensional morphing wings for adaptive morphing flight missions under variable operating conditions. The CST parameterization method is chosen to define three-dimensional morphing wings, and the Latin hypercube sampling method is used to sample in the design space and generate sample points. Computational fluid dynamics simulations are performed to obtain corresponding aerodynamic parameters, and the deep belief network surrogate model is constructed to map the input-output relationship between morphing design parameters and aerodynamic parameters. To address the variable operating conditions, a DQN (deep Q-network) reinforcement learning agent, leveraging unsupervised learning, is used to provide real-time morphing strategies to achieve expected aerodynamic performance. Furthermore, the design results via the DQN agent are compared with those via the G-CGAN (greedy-based conditional generative adversarial network) agents. The results indicate that the proposed RLID framework efficiently obtains a satisfactory strategy of morphing wings under variable operating conditions and that the DQN agent focuses more on overall task rewards than the G-CGAN agent.

     

/

返回文章
返回