Gaussian Processes for Structural Health Monitoring of Wind Turbine Blades

Abstract

Nonstationary environmental and operational variables (EOVs) acting on wind turbines present challenges for the successful application of structural health monitoring systems. A contributing factor to these challenges is the fact that EOVs are often not monitored sufficiently, leading to uncertainties being introduced into monitored components. The method proposed in this paper takes advantage of the fact that all blades on a wind turbine possess nominally identical properties and encounter the same EOVs. Gaussian processes (GPs) are used to learn the relationships in the properties between pairs of blades when they are in a healthy state. The GPs then predict the properties of one blade, given that of another, and deviations between the actual and predicted properties (i.e. the residual errors) are used to indicate the occurrence of damage. To validate this method, it is applied to data from a real wind turbine, where some form of blade damage has been known to have taken place. X-bar control chart analysis shows that this method identifies damage as early as six months before the damage led to problems.