Modelling long-term vibration monitoring data with Gaussian Process time-series models

Abstract

Gaussian Process (GP) time-series models are a special type of models for Linear Parameter Varying (LPV) systems in which the parameters are represented as stochastic variables following a Gaussian Process regression of the scheduling variables. GP time-series models are ideal for the representation of LPV systems where some of the scheduling variables are uncertain or immeasurable, as is the case in most real-life Structural Health Monitoring (SHM) applications. In this work, a fully parametric version of GP is adopted, most suitable for identification based on large datasets typically originated in SHM campaigns. Here, the model identification problem is addressed via global and local approaches, while is demonstrated that the latter case corresponds to a sub-optimal version of the global optimization. Finally, the GP time-series modelling methodology is demonstrated on the identification of the simulated vibration response of a wind turbine blade, where temperature and wind speed act as scheduling parameters.