基于TSPU-Net的航天器测试数据分析方法及应用研究

Research on the Analysis Method and Application of Spacecraft Test Data Based on TSPU-Net

摘要: 针对航天器在轨温度预测中传统热网格法对瞬态温度预测效果不佳的问题，本文提出了一种基于历史遥测数据驱动的多元时序数据预测模型——TSPU-Net。该模型通过设计时序处理单元(TSPU)，有效融合了一维卷积(Conv1D)、多尺度归一化、Dropout、GELU激活函数以及门控循环单元(GRU)等多种深度学习技术，旨在全面提高航天器遥测数据的特征提取和表示能力，并捕捉多尺度时间依赖关系。数值实验结果表明，TSPU-Net模型在参数量较少的情况下，对航天器温度遥测数据具有较高的预测精度和普适性。与TCN和2D-CNN等主流模型相比，TSPU-Net在均方误差(MSE)、平均绝对误差(MAE)等评价指标上均表现出显著优势，尤其在多步预测任务中展现出更强的鲁棒性。本研究为航天器故障预警、健康管理以及大规模星座数字化模型构建提供了新的解决方案，具有重要的工程应用价值。

Abstract: Addressing the limitations of traditional thermal grid methods in predicting transient temperatures for on-orbit spacecraft, this paper proposes a multivariate time series prediction model driven by historical telemetry data, named TSPU-Net. By designing a Time Series Processing Unit (TSPU), this model effectively integrates various deep learning techniques, including one-dimensional convolutional neural networks (Conv1D), multi-scale normalization, Dropout, GELU activation function, and Gated Recurrent Unit (GRU), aiming to comprehensively enhance the feature extraction and representation capabilities of spacecraft telemetry data and capture multi-scale temporal dependencies. Numerical experimental results demonstrate that the TSPU-Net model achieves high prediction accuracy and universality for spacecraft temperature telemetry data with fewer parameters. Compared to mainstream models such as TCN and 2D-CNN, TSPU-Net exhibits significant advantages in evaluation metrics such as Mean Squared Error (MSE) and Mean Absolute Error (MAE), especially showing stronger robustness in multi-step prediction tasks. This research provides a novel solution for spacecraft fault early warning, health management, and the construction of large-scale constellation digital models, possessing significant engineering application value.