Remaining useful life prediction of satellite momentum flywheel based on ISCSO-LSTM network ensemble

As the core actuator of the attitude control system, the momentum flywheel’s performance degradation directly impacts the satellite’s lifespan. Therefore, it is essential to research predicting the remaining useful life (RUL) of the momentum flywheel. To address the issues of insufficient prediction accuracy and unstable prediction performance in momentum flywheel RUL, this paper proposes an integrated prediction method based on an improved sand cat swarm optimization (ISCSO) algorithm and long short-term memory (LSTM) network. First, multi-dimensional candidate features are extracted from the original bearing temperature sequence of the momentum flywheel, and a comprehensive evaluation method based on monotonicity and trendency is employed to select sensitive features. Next, multiple LSTM networks with different initialization parameters are constructed as base predictors to enhance the model’s generalization ability. Finally, the ISCSO algorithm, incorporating a Tent chaotic map, is introduced to optimize the output weights of each base predictor. Experimental results show that the proposed method outperforms comparison methods across multiple evaluation metrics. Ablation experiments further verify the effectiveness of the improved optimization algorithm and integration strategy.