Increasing growth of wind turbine systems suggests a more systematic research around their design, operation and maintenance is needed. These systems operate under challenging enviromental conditions and failure of some of their parts, for the time being, is frequent, although undesirable. Wind turbine gearboxes, more particularly, seem to be so problematic that some wind turbine designs avoid including them. Structural health monitoring and condition monitoring of wind turbines appear to be necessary in order to determine the condition and lifespan of the wind turbine components and the drivetrain respectively. In this way reparative actions could be taken whenever needed resulting in reduction of maintenance costs.
This thesis focuses on the condition monitoring of wind turbine gearboxes, taking into account the varying loads that they endure. Currently, the vibration-based damage detection methods used in real life wind turbine condition monitoring systems are based on conventional methods that generally fail to detect damage at its early stage under the operational conditions observed in wind turbines. Load and speed variations of the drivetrain that are observed commonly in wind turbines influence the vibration signals and can possibly affect potential damage features. This shows a demand for effective methods for early damage detection. Developments in the area of advanced signal processing should be examined and applied in damage detection of wind turbine gearboxes. Methods from time-frequency analysis, time-scale analysis, pattern recognition, multivariate statistics and econometrics are examined in this study in a condition monitoring context.
One important part of the work presented is the development of a simple gearbox model interfaced with realistic wind loading, a model feature that appears to be novel. Other interesting aspects of this thesis are related to the use of the empirical mode decomposition method for time-frequency analysis. The use of Teager-Kaiser energy operator as an alternative technique to Hilbert transform for the estimation of the instantaneous characteristics of the decomposed signals is one of these aspects. The study showed that for some cases and under certain conditions this operator could help to improve the time-frequency analysis. Another aspect is the observation of the change of the number of the intrinsic mode functions produced, for the different load and damage cases, during the decomposition process. This observation was connected theoretically with what is known as the mode mixing problem of the empirical mode decomposition method. For the feature discrimination part of this work, the simplest novelty detection method, outlier analysis, was used in a slightly different manner than in previous studies and the results obtained were compared with a novel adaptive thresholding technique, the 3D phase-space thresholding method. The previously described approaches were applied on the simulated gearbox data but also on real wind turbine gearbox data. Finally, cointegration analysis was proposed as a potential method for removing the effects of the gearbox load variations. This is a novel concept for the condition monitoring of wind turbine gearboxes. An approach which makes it possible to use data from just a single sensor in order to perform cointegration analysis was developed and the process for applying multiscale cointegration using either wavelets or the empirical mode decomposition method was discussed. This final part of the work is an initial step towards applying cointegration to condition monitoring data.
