Intelligent sensitivity analysis framework based on machine learning for spacecraft thermal design

Abstract

The variance-based method of global sensitivity analysis (GSA) has been widely applied in spacecraft thermal design, which is typically calculated using Monte Carlo estimations. However, such estimations require a large number of samples to ensure sufficient accuracy, which makes GSA expensive to perform when modeling is difficult. Moreover, multimodal or highly skewed output distributions may result in the use of variance as an uncertain agent that generates contradictory results. Therefore, an intelligent density-based GSA framework based on machine learning and multi-fidelity metamodels called IDGSA-3M is proposed. An intelligent batch processing system based on a real-time data interaction between MATLAB and NX/TMG was designed that uses many cheap low-fidelity sample points to reduce the cost of model evaluation while using a small number of expensive high-fidelity sample points to maintain high accuracy, thus achieving trade-offs between high accuracy and low computational cost. A radial basis function (RBF) neural network based on an improved mind evolutionary algorithm was applied to approximate the multi-fidelity metamodel of a spacecraft thermophysical model calculated using a batch processing system, which had a computational speed that was 1000+ times faster than that of the traditional thermophysical model and a high computational accuracy of 99%+. The output distributions of the RBF were then characterized by its cumulative distribution functions to obtain density-based sensitivity indices. Both the theoretical and experimental results of GSA for the thermal design parameters of the extreme ultraviolet radiation detector on the space-based Lyman-Alpha Solar Telescope, developed in China, demonstrated that the convergence rate of IDGSA-3M can be improved up to 10-fold for a fixed convergence level in comparison with two other GSA methods, thereby verifying its superiority.