Scholarlink

<p>To safely control helicopters, it is critical to accurately model their dynamics and to capture the highly nonlinear effects generated by aerodynamic forces, propeller-fuselage interactions, vibrations and other phenomena. However, such nonlinear effects can hardly be measured or modeled by the physics-based model. Although deep learning methods (e.g., CNNs and RNNs) were used to learn these complex dynamics, they exhibit limitations in fully harnessing the multidimensional information of flight data, particularly the state-control space and the temporal dimension. This paper proposes a residual hybrid model based on a temporal-spatial Transformer (TS-Trans) network, efficiently utilizing the information hidden in flight data to learn the residual dynamics that is ignored by the physics-based model. Compared with several baseline models, the proposed model has the best performance on a real-world helicopter flight dataset. Furthermore, in order to tackle the problem of formidable sizes and computational costs induced by the TS-Trans model, an autonomous rank-search algorithm is proposed to integrate the low-rank approximation into the model compression of the TS-Trans model, which can significantly reduce the computational costs by 69% with a slight performance degradation of 2.70%.</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

Modeling residual dynamics of helicopters based on temporal-spatial Transformer and low-rank compression

  • 期刊论文

  • 发表日期：

    2025年3月4日

查看全文
  • Hailang Zhang,
  • Jing Liu,
  • Yu Hu,
  • Xiaoqing Tian

摘要

To safely control helicopters, it is critical to accurately model their dynamics and to capture the highly nonlinear effects generated by aerodynamic forces, propeller-fuselage interactions, vibrations and other phenomena. However, such nonlinear effects can hardly be measured or modeled by the physics-based model. Although deep learning methods (e.g., CNNs and RNNs) were used to learn these complex dynamics, they exhibit limitations in fully harnessing the multidimensional information of flight data, particularly the state-control space and the temporal dimension. This paper proposes a residual hybrid model based on a temporal-spatial Transformer (TS-Trans) network, efficiently utilizing the information hidden in flight data to learn the residual dynamics that is ignored by the physics-based model. Compared with several baseline models, the proposed model has the best performance on a real-world helicopter flight dataset. Furthermore, in order to tackle the problem of formidable sizes and computational costs induced by the TS-Trans model, an autonomous rank-search algorithm is proposed to integrate the low-rank approximation into the model compression of the TS-Trans model, which can significantly reduce the computational costs by 69% with a slight performance degradation of 2.70%.