基于PINN的空间电磁捕获刚体动力学建模与预测

PINN-Based Modeling and Prediction of Rigid-Body Dynamics in Space Electromagnetic Capture

摘要: 在空间电磁吸附任务中，铁质目标的动力学预测对于轨迹规划与控制具有重要意义。传统有限元仿真虽然精确，但计算开销大，难以满足快速预测需求。近年来，深度学习方法如多层感知机(MLP)、长短期记忆网络(LSTM)等被用于轨迹预测，但其泛化能力有限，在未见过的工况下往往出现偏差。为此，本文提出了一种基于物理信息神经网络(Physics-Informed Neural Network, PINN)的预测方法，将刚体动力学方程作为物理约束引入神经网络训练。通过Ansys Motion与Maxwell联合仿真构建的1200组多变量轨迹数据集进行验证，结果表明，PINN在位置、速度和角速度等关键指标上的预测精度显著高于传统数据驱动模型。在跨质量、速度、电流和偏移量的组合外推实验中，PINN依然保持较低误差和良好稳定性，体现了物理约束对数据预测的有效增强。研究结果表明，PINN可在保证物理一致性的同时，实现高效、鲁棒的动力学预测，为空间电磁捕获任务提供支持。

Abstract: Accurate prediction of rigid-body dynamics under electromagnetic capture is crucial for trajectory planning and control in space missions. While finite element-based simulations provide high fidelity, they are computationally expensive and unsuitable for rapid prediction. Recently, deep learning methods such as multilayer perceptrons (MLP) and long short-term memory networks (LSTM) have been applied to trajectory prediction, but their generalization ability remains limited, often leading to significant deviations under unseen conditions. To address this, we propose a Physics-Informed Neural Network (PINN)-based prediction framework that incorporates rigid-body dynamics equations as physical constraints during training. Using a dataset of 1200 trajectories generated from co-simulations of Ansys Motion and Maxwell under varying mass, velocity, current, and offset conditions, we evaluate the performance of PINN against traditional data-driven models. Results show that the PINN achieves significantly higher accuracy in predicting position, velocity, and angular velocity, and maintains low error and strong stability in extrapolative scenarios across unseen parameter combinations. This demonstrates that embedding physical knowledge into neural networks effectively enhances data prediction, providing a robust and efficient tool for space electromagnetic capture applications.