In many SHM applications, a statistical framework for the distinction between damaged and normal states is often desirable. Usually, conventional methods which employ such statistical tests aim at the identification of abnormal, or extreme, data points when compared to a model of normality. This comparison is mainly done piecewise on individual points in a univariate case, or on datasets, if multivariate features are available. The latter case commonly involves the fusion of the multivariate feature into a scalar variable which is then individually assessed, again in a statistical framework. Depending on the actual application, this approach may be limited, or prone to false identifications. A new theory, previously proposed by researchers at Oxford University and applied for novelty detection, attempts to identify functions instead of just datasets, which are then assessed in terms of their novelty. The extreme function theory develops a normal model of the structure based on functions which can be acquired from different sources of data, and thus account for a more robust damage detection. This paper presents the extreme function theory as an approach to SHM. A case study of monitoring wind turbines with the help of power curves i.e. models which relate the wind speed to the power produced, is presented. The data used in the study originate from a real wind farm.
doi: 10.12783/SHM2015/342