融合CNN和BiLSTM的挠性太阳帆时变惯性张量辨识

Identification of time-varying inertia tensor of flexible solar sail based on CNN and BiLSTM

摘要:

为准确辨识挠性太阳帆航天器在轨展开过程中的惯性张量参数,并提升复杂干扰环境下参数辨识模型的精度和鲁棒性,结合卷积神经网络（CNN）的局部特征提取能力和双向长短期记忆网络（BiLSTM）的时间序列建模优势,提出了一种基于CNN-BiLSTM混合网络的数据驱动型参数辨识方法。首先,建立挠性太阳帆航天器的姿态－振动耦合动力学模型,并利用领域随机法生成大量训练数据。其次,构建CNN-BiLSTM网络参数辨识模型,并采用预先设计的训练策略对其进行训练。最后,使用训练完成的CNN-BiLSTM模型在不同的测量噪声条件下辨识太阳帆的惯性张量,并比较分析了该模型和单一模型的辨识结果。仿真结果表明,所构建的网络模型在无干扰的条件下,辨识出的系统惯性张量参数的平均相对误差低于1%；而在具有干扰力矩和较低信噪比的复杂测量噪声条件下,辨识结果的平均相对误差仍不超过1.5%,性能显著优于CNN和BiLSTM模型。验证了本文方法能有效克服单一网络模型的不足,不仅能准确辨识出挠性太阳帆的时变惯性张量,还能增强在复杂干扰环境中的鲁棒性。

Abstract:

To accurately identify the inertia tensor parameters of flexible solar sail spacecraft during on-orbit deployment and improve the accuracy and robustness of the parameter identification model in complex disturbance environments, this paper proposed a data-driven parameter identification method based on a convolutional neural network (CNN) and bidirectional long short-term memory (BiLSTM) hybrid network. This method combined the local feature extraction capability of CNN with the time series data modeling advantage of BiLSTM. First, the paper established the attitude-vibration coupling dynamics model of the flexible solar sail spacecraft and generated massive training data using domain randomization. Second, a CNN-BiLSTM network parameter identification model was constructed and trained using a pre-designed training strategy. Finally, the trained model was applied to identify the inertia tensor of the solar sail under different measurement noise conditions, and the identification results of this model and single models were compared and analyzed. Simulation results demonstrate that the constructed network model has a mean relative error of less than 1% in identifying the system inertia tensor parameters under disturbance-free conditions. In addition, under complex measurement noise conditions with disturbance torques and lower signal-to-noise ratios, the mean relative error of the identification results still does not exceed 1.5%, which is significantly better than that of the CNN and BiLSTM models. The proposed method effectively addresses the shortcomings of single network models, not only accurately identifying the time-varying inertia tensor of the flexible solar sail but also enhancing robustness in complex disturbance environments.