Use of Cokriging for Thermal Analysis in Small Satellites

In space, where human intervention is not possible, faults must be autonomously detected and rectified. Inaccuracies, delays, or disturbances can cause component failures that can lead to catastrophic failure. In light of this, a dynamic system simulation can greatly enhance the operational phase of a satellite. This work investigates a multi-fidelity framework for the simulation of small satellites and compares it to traditional regression methods, in particular Gaussian processes and Gated Recurrent Units. The framework combines a computationally cheap, low fidelity surrogate model with an accurate high-fidelity model. In the case of the former, recurrent neural networks, particularly a Gated Recurrent Unit is considered. For the latter, a finite element model is used to produce sparse high-fidelity data describing the satellite's state. High fidelity simulations are expensive. However, abundant low fidelity data can be taken advantage of to speed up the process. Therefore, by means of cokriging, low fidelity data are corrected by high-fidelity data through a comprehensive correction, where the parameters are given by the use of Gaussian processes to provide uncertainty quantification. When some new data arrives, the model can be refitted for a minimal computation cost. The framework is demonstrated through a set of simulations, using thermal analysis data from the NSPO-1 satellite. NSPO-1 is a Taiwanese Space Organization's (NSPO) 6U cube satellite, co-developed by the Intelligent Space Systems Laboratory (ISSL) of the University of Tokyo, intended to orbit in LEO and is intended to provide a convenient validation platform to test optical sensors developed by the NSPO.