Gaussian process models for mitigation of operational variability in the structural health monitoring of wind turbines

The analysis presented in this work relates to the quantification of the effect of a selected set of measured Environmental and Operational Parameters (EOPs) on the dynamic properties of low and high frequency vibration, in the context of a vibration monitoring campaign implemented on the blade of an operating wind turbine. To this end, a Gaussian Process (GP) time-series modelling approach is adopted, in which the coefficients of a time-series model are driven by a Gaussian Process Regression on the selected EOPs. The properties of the data acquisition system allow to evaluate low and high frequency dynamics, the former associated with the structural dynamics of the blade, and the latter with the wave transmission properties of the material, assessed with the help of an electromechanical actuator installed on the blade. In this form, a multi-temporal-scale approach is adopted here, where a GP Linear Parameter Varying Auto-Regressive model is selected to represent low frequency (structural) dynamics, while in parallel a GP Continuous Wavelet Transform model is used to represent high frequency dynamics (associated with wave transmission properties in the material). In both cases the blade is considered in its healthy state as well as in various operational regimes, including idle, and rotating at two different set points. As a result, it is demonstrated that GP time-series modelling succeeds in evaluating and isolating the influence of different EOPs in the features of the vibration response of the wind turbine blade, and at the same time, normalize their effects to enhance the detectability of damage. (C) 2020 Elsevier Ltd. All rights reserved.