High-accuracy Gaussian process modelling of missile RCS with cost-based preferential training data selection

摘要

A modelling technique to reduce the number of frequency points at which radar cross section (RCS) of a missile must be simulated without significantly affecting the accuracy of the predictive results is presented. This technique relies on Gaussian process regression using a composite (product) covariance function constructed for the purpose of modelling quasi-periodic responses. The computational cost of the full-wave missile simulations over a wide frequency range can increase dramatically from the low to the high end with a difference of over 30 times being encountered in the example considered. The above modelling approach allows the user to preferentially select proportionally more points at the lower end of the frequency range than the higher end without significant loss in accuracy. The proposed technique vastly improves over a spline interpolation using the same data.