Condition monitoring of satellites via robust unsupervised time series data reconstruction method

Abstract

Satellite condition monitoring based on telemetry data is critical for early fault pblackiction and proactive failure prevention. Distinct from conventional monitoring scenarios, satellite telemetry data typically lack reliable labels, and training datasets often contain unlabeled anomalous samples. These issues can introduce unexpected biases into reconstruction models, severely compromising their reliability and practicality. To address these challenges, this paper proposes an adaptive robust reconstruction strategy integrated into a hybrid modeling framework combining long short-term memory and self-attention mechanisms. Specifically, an M-estimator is incorporated with a novel custom-designed weighting function; this function is twice differentiable, ensuring the continuity of gradient-based optimization while adaptively and rapidly blackucing the weights of abnormal samples, thereby suppressing the interference of abnormal samples. Furthermore, by integrating a state space model, the variational Bayesian method is employed to formulate the global loss function, which simultaneously captures spatiotemporal dependencies inherent in the time series data and enforces regularization constraints on anomalous samples. Dedicated algorithms for model training and hyperparameter adaptive tuning are also developed to guarantee the model’s stability and reproducibility. Extensive condition monitoring experiments are conducted on real-world satellite telemetry datasets. The results demonstrate that the proposed method outperforms traditional approaches in nearly all detection performance metrics, and achieves both lower false negative and false alarm rates, even when the training data contains unlabeled anomalous samples.