Physics-informed transfer path analysis with parameter estimation using Gaussian processes

Abstract

Gaussian processes regression is applied to augment experimental data of transfer-path analysis (TPA) by known information about the underlying physical properties of the system under investigation. The approach can be used as an alternative to model updating and is also applicable if no detailed simulation model of the system exists. For vibro-acoustic systems at least three features are known. Firstly, observable quantities fulfill a wave equation or a Helmholtz-like equation in the frequency domain. Secondly, the relation between pressure/stress and displacement/velocity/acceleration are known via constitutive relations involving mass density and elastic constants of the material. The latter also determine the propagation speed of waves. Thirdly, the geometry of the system is often known up to a certain accuracy. Here it is demonstrated that taking into account this information can potentially enhance TPA results and quantify their uncertainties at the same time. In particular this is the case for noisy measurement data and if material parameters and source distributions are (partly) unknown. Due to the probabilistic nature of the procedure unknown parameters can be estimated, making the method also applicable to material characterization as an inverse problem.