A machine learning approach to Structural Health Monitoring with a view towards wind turbines

The work of this thesis is centred around Structural Health Monitoring (SHM) and is divided into three main parts. The thesis starts by exploring di�erent architectures of auto-association. These are evaluated in order to demonstrate the ability of nonlinear auto-association of neural networks with one nonlinear hidden layer as it is of great interest in terms of reduced computational complexity. It is shown that linear PCA lacks performance for novelty detection. The novel key study which is revealed ampli�es that single hidden layer auto-associators are not performing in a similar fashion to PCA. The second part of this study concerns formulating pattern recognition algorithms for SHM purposes which could be used in the wind energy sector as SHM regarding this research �eld is still in an embryonic level compared to civil and aerospace engineering. The purpose of this part is to investigate the e�ectiveness and performance of such methods in structural damage detection. Experimental measurements such as high frequency responses functions (FRFs) were extracted from a 9m WT blade throughout a full-scale continuous fatigue test. A preliminary analysis of a model regression of virtual SCADA data from an o�shore wind farm is also proposed using Gaussian processes and neural network regression techniques. The third part of this work introduces robust multivariate statistical methods into SHM by inclusively revealing how the in uence of environmental and operational variation a�ects features that are sensitive to damage. The algorithms that are described are the Minimum Covariance Determinant Estimator (MCD) and the Minimum Volume Enclosing Ellipsoid (MVEE). These robust outlier methods are inclusive and in turn there is no need to pre-determine an undamaged condition data set, o�ering an important advantage over other multivariate methodologies. Two real life experimental applications to the Z24 bridge and to an aircraft wing are analysed. Furthermore, with the usage of the robust measures, the data variable correlation reveals linear or nonlinear connections.