Comparison between a polynomial chaos surrogate model and Markov Chain Monte Carlo for inverse uncertainty quantification based on an electric drive test bench

Abstract

The development of uncertainty quantification schemes has been pushed forward due to the increasing demands for complex physical and computational simulation models. In industrial applications, distributions on model parameters play a crucial role and quantifing them is a big challenge. In this work, a test bench hardware is presented, which is designed to measure the motor characteristic of an electric drive. The special aspect of this setup is that parameter distributions, which in general are unknown, can be defined a priori. The obtained measurements serve as a reference to analyse the convergence of Polynomial Chaos (PC) and Markov Chain Monte Carlo (MCMC) in context of Bayesian inference. Our focus is on analysing the feasibility of the PC approach as a surrogate model to replace the forward model in the Bayesian inference. In comparison to the classical approach, which directly uses the simulation model, we investigate the number of simulations needed to obtain a good estimation of the parameter distribution. In addition we use different orders for the PC expansion to fit the surrogate model. In our benchmark, we show that the PC expansion is able to significantly reduce the computational cost compared to a pure MCMC approach.