Robust Anomaly Detection of Spacecraft Telemetry Data Based on Hybrid Time-Frequency Decomposition and Deep Sequence Learning

Real-time monitoring and anomaly detection of spacecraft health status in orbit are essential for ensuring mission safety and reliability. However, the strongly non-stationary nature of telemetry data under extreme space environments poses significant challenges to conventional methods that rely on the assumption of data stationarity, making it difficult to distinguish genuine anomalies from environmental fluctuations. Moreover, the inherent scarcity and non-repeatability of anomaly events lead to a severe shortage of labeled data for model training. To overcome these limitations, this paper proposes a robust unsupervised detection framework that integrates physically inspired time–frequency decomposition with deep sequential learning. The core of the framework lies in a hybrid decomposition module employing a threefold mechanism: bandwidth-adaptive local regression for nonlinear trend extraction, phase-aware periodicity decoupling via analytic signal transformation and normalized instantaneous phase computation, and residual isolation for fine-grained fluctuation modeling. The decomposed components are then processed by a multi-component attention Transformer, in which cross-attention independently enhances component-specific representations while maintaining inter-component dependencies. Finally, unsupervised anomaly detection is accomplished by combining reconstruction-based errors with correlation-aware anomaly scores. Extensive experiments conducted on three real-world spacecraft telemetry datasets and one industrial benchmark demonstrate that the proposed method achieves superior detection accuracy, stability, and robustness compared to existing state-of-the-art algorithms.