Active Learning-based Reduced Order Model for Thermal Behavior Prediction of a Lunar Orbiter

초록

In this study, an active learning-based reduced order model was developed for efficient thermal behavior prediction of lunar orbiters. Since lunar orbiters experience extreme temperature variations during missions, accurate temperature prediction is essential for mission success. Although conventional numerical analysis-based full-order models have been widely used, they clearly have limitations in terms of computational cost. To address this issue, we propose a model that combines principal component analysis for dimensionality reduction with ensemble neural network-based active learning. The proposed methodology enables high prediction accuracy with a small number of training data samples by selectively choosing additional training data from regions with high prediction uncertainty in the ensemble neural network. Model validation showed that the active learning-based model outperformed conventional deep neural networks and simple reduced order models in all error metrics. Particularly after three active learning cycles (total of 80 training data points), prediction results were almost indistinguishable from those of the full-order model. This demonstrates the efficiency of the uncertainty-based data selection strategy and is expected to be utilized for real-time temperature prediction and mission operation optimization in computationally expensive lunar orbiter thermal analysis problems.

키워드