The cost of energy generated from wind power plants (particular if located offshore) is challenging societies in terms of desiring cheaper and more environmentally friendly generated electrical energy. The high cost reduction targets can be aided by broad application of condition monitoring systems, which bear the potential to support plant owners reducing turbine downtime and lowering costs. In this research a global condition monitoring system is proposed, which provides a platform to take advantage of the different information sources available to operators. One of the most common sources for information about the component condition is Supervisory Control And Data Acquisition (SCADA) data, e.g. temperature, current orvoltage measurements from different components. Using newly developed Adaptive Neuro-Fuzzy Interference System (ANFIS) models, a normal behavior model based approach is taken to extract information from these data, which otherwise are covered by the high signal variance. A novelty here is the application of ANFIS models in this context, which is traditionally covered by neural networks or regression based approaches. Advantages in training speed and back traceability of results are the motivators to apply these promising model types. Methods proposed in literature usually take an autoregressive approach, i.e. a delayed version of the output signal is used as input alongside other highly correlated signals. In this research, a full signal reconstruction method is proposed in which the output signal is entirely reconstructed by using other correlated signals. Benefits in fault visibility and lead-time to failure estimatesare observed. A very important signal to monitor contained in the SCADA data is the wind turbine power output. The power output has a direct influence on the revenue of the operators and thus the cost of energy. Normal behavior models from literature neglect the influence of the wind direction and the ambient temperature, which can, theoretically speaking, influence the expected power output by up to 20%. A new set of model inputs is proposed and validated in this research. Higher prediction accuracy and thus earlier anomaly detection in case of turbine performance deviations is achieved. Another information sourcepotentially available to operators is vibration measurements from the wind turbine drive train. With respect to the condition of gearbox bearings, a recently proposed method to separate discrete (e.g. originating from gears) from random (e.g. originating from bearings) signal components is applied and validated in this research. This state of the art method named“signal pre-whitening” enhances the fault pattern visibility in the envelope spectra in a very efficient manner. The validation is based on real measured high-speed and intermediate speed bearing faults. Additionally, a semi-automated method for bearing diagnosis suggested in literature is updated with the most promising signal pre-processing techniques to enhance fault visibility and further developed leading to fully automated fault diagnosis. For this purpose a frequency content identifier is developed extracting the frequency content from the envelope spectrum building the basis for automated diagnosis. A modified parameter, namely the Kurtosis of the Amplitude Envelope Spectrum (KEAS), is proposed and its performance is compared to the original parameter, the Spectral Kurtosis (SK), which is used to indicate the fault severity. Moreover, state of the art methods to monitor wind turbine gears have been briefly compared and the most promising parameters are identified using a gear pair with false brinelling. Finally, a global fuzzy expert system is developed giving the possibility of linking all available information in terms of fuzzy logic rules. It is important to highlight that this research is based on real measured data coming from two wind power plants with turbines of a different type and brand. In total, operational data from 39 turbines were available for this research project.