基于ASTRA-Net的航天器遥测数据异常检测方法研究

Research on Anomaly Detection Method of Spacecraft Telemetry Data Based on ASTRA-Net

摘要: 随着航天技术的飞速发展，航天器在轨运行的复杂性日益增加，其健康状态的实时监控与异常诊断成为确保任务成功的关键。遥测数据作为航天器运行状态的直接反映，蕴含着丰富的系统信息。然而，海量的多源异构遥测数据也给传统异常检测方法带来了巨大挑战。本文针对航天器遥测数据中存在的多元性、时序性、非线性及潜在异常模式复杂等特点，提出了一种基于深度学习的自适应时空循环注意力网络(ASTRA-Net)模型，用于实现航天器遥测数据的精准异常检测。ASTRA-Net模型创造性地融合了卷积神经网络(CNN)在局部特征提取方面的优势、双向门控循环单元(BiGRU)在时序依赖建模方面的能力，以及注意力机制(Attention Mechanism)在关键信息加权方面的效能。通过CNN层，模型能够有效捕捉多元遥测数据内部的局部相关性和空间特征；BiGRU层则进一步学习数据在时间维度上的双向依赖关系，增强对时序模式的理解；而引入注意力机制，使得模型能够自适应地关注对异常检测更具判别力的特征和时间步，从而提升了模型的预测精度和异常检测的灵敏度。在此基础上，结合改进的广义自回归条件异方差(GARCH)模型构建动态阈值，实现了对遥测数据异常的精确识别与定位。通过模拟航天器遥测数据进行案例验证，实验结果表明，ASTRA-Net模型在预测精度和异常检测性能上均表现出色，显著优于传统方法，为航天器在轨健康管理与故障诊断提供了新的技术途径。

Abstract: With the rapid advancement of space technology, the complexity of spacecraft in-orbit operations has significantly increased, making real-time health monitoring and anomaly diagnosis crucial for mission success. Telemetry data, as a direct reflection of spacecraft operational status, contains rich system information. However, the massive volume of multi-source heterogeneous telemetry data poses significant challenges to traditional anomaly detection methods. This paper addresses the characteristics of spacecraft telemetry data, including its multivariate, temporal, nonlinear, and complex potential anomaly patterns, by proposing a deep learning-based Adaptive Spatio-Temporal Recurrent Attention Network (ASTRA-Net) model for precise anomaly detection. The ASTRA-Net model innovatively integrates the strengths of Convolutional Neural Networks (CNN) in local feature extraction, Bidirectional Gated Recurrent Units (BiGRU) in temporal dependency modeling, and Attention Mechanisms in weighting critical information. The CNN layer effectively captures local correlations and spatial features within multivariate telemetry data; the BiGRU layer further learns bidirectional dependencies in the temporal dimension, enhancing the understanding of temporal patterns; and the introduction of the Attention Mechanism enables the model to adaptively focus on features and time steps that are more discriminative for anomaly detection, thereby improving prediction accuracy and detection sensitivity. Building upon this, a dynamic threshold is constructed by combining the predicted mean sequence with an improved Generalized Autoregressive Conditional Heteroskedasticity (GARCH) model, achieving accurate identification and localization of telemetry data anomalies. Through case studies using simulated spacecraft telemetry data, experimental results demonstrate that the ASTRA-Net model exhibits excellent performance in both prediction accuracy and anomaly detection, significantly outperforming traditional methods, thus providing a new technical approach for in-orbit health management and fault diagnosis of spacecraft.