Global Sensitivity Analysis Based on BP Neural Network for Thermal Design Parameters

Abstract

In order to obtain the thermal design parameters that have a great influence on the temperature of the spectrometer frame, the sensitivity of the thermal design parameters of a balloon-borne spectrometer system was analyzed and calculated by the global sensitivity analysis (GSA) method based on the backpropagation neural network (BPNN) surrogate model. Firstly, the BPNN with 12 selected thermal design parameters as input and temperature as output was well trained. Then, two kinds of variance-based GSA methods, the Sobol’ method and the extended Fourier amplitude sensitivity test (EFAST), were used to calculate the values and ranking results of sensitivity indices of 12 parameters based on the established BPNN. Moreover, the GSA results were verified based on the finite element model of the balloon-borne spectrometer system built by I-DEAS/TMG (software developed by Structural Dynamics Research Corporation for space thermal analysis), which indicates that the BPNN surrogate-model-based GSA is reliable. Finally, the sensitivity calculation accuracy and speed of two methods, the Spearman rank correlation coefficient formula and the GSA method based on BPNN, were compared, and the EFAST method based on the BPNN surrogate model has been proved to have obvious advantages in the reliability and speed of calculation results. Also, the GSA method based on a surrogate model like BPNN is of great significance in the thermal analysis of an optical remote sensor.