Artificial neural network-based temperature prediction of a lunar orbiter in thermal vacuum test: Data-driven reduced-order models

Abstract

This study presents data-driven reduced-order models (ROMs) of a lunar orbiter based on principal component analysis (PCA) and artificial neural networks (ANNs) for a ground thermal vacuum test to simulate space thermal environments. We employed a radial basis function network (RBFN) and deep neural network (DNN) from among the various types of ANNs. PCA extracts features from high-dimensional data, such as thermal analysis data. It is utilized in machine learning algorithms as a preprocessing step before inputting the data into neural networks. This process improves the convergence speed and training performances compared to using neural networks alone. The coefficients of the extracted principal component modes were regressed using the RBFN and DNN. Twenty thermal design parameters comprising infrared emissivity, effective thermal conductivity, thermal contact conductance coefficients, and thermal conductance were used to train the ROMs. We conducted training and test of the proposed models during the cold and hot balance phases of the ground test. Consequently, the temperature map can be estimated in seconds for the new design parameters, and the model results are consistent with thermal analysis and measurement data.