Neural network-based thermal model for virtual metrology of lunar orbiter temperatures via active and transfer learning

초록

South Korea's Danuri lunar orbiter is currently conducting an observation mission at an altitude of 100 km above the Moon. Accurate temperature prediction is crucial for ensuring the mission's success in the harsh thermal environment of space. However, existing physical models based on finite element or difference methods are computationally intensive and challenging to fully validate due to limited ground testing opportunities. In this study, we developed a reduced order model using active learning and a deep neural network to replace traditional physical models. We then introduced a neural network model that incorporates transfer learning with real orbital temperature data to improve upon the active learning-implemented model to better predict real-world temperatures. Notably, our results demonstrate that the model fine-tuned with the flight data is the most accurate for predicting the lunar orbiter's temperature from the ground. We are confident that the proposed active and transfer learning-implemented model can also be employed to create a virtual measurement system that serves as a digital twin of the lunar orbiter and can be adapted for thermal design, analysis, and operation of new planetary spacecrafts by adjusting the input parameters.

키워드