Wind energy plays a pivotal role in the energy transition, however, a major obstacle for offshore wind farms remains the high Operations and Maintenance (O&M) costs, which can be up to 35% of the levelized cost of energy (LCOE). Traditional maintenance strategies often result in excessive downtime, increased costs, or premature component replacements. To counter these problems, condition-based maintenance (CBM) uses real-time data to estimate machine health and guide maintenance decisions. Although most CBM studies address either health assessment or maintenance optimisation in isolation, this thesis proposes an end-to-end CBM framework that derives health indicators (HIs) and integrates them into a multiple-threshold CBM framework.
The framework is validated on rotor bearing failures from the CARE data set, which supplies real-world Supervisory Control and Data Acquisition (SCADA) measurements. Rotor bearings, despite their relatively low failure frequencies, cause high downtime and exhibit a prolonged degradation pattern. HIs are extracted with a long- and short-term memory (LSTM) autoencoder (AE) trained solely on healthy turbines to learn normal behaviour. On avarage, the resulting HIs identify anomalies 195 days before failure.
These HIs are fed into a CBM strategy that prioritises timely, minor interventions over costly lastminute replacements. Compared to a purely corrective replacement policy, the strategy reduces annual maintenance expenditure on rotor bearings by an average of 62.5%, mainly by prolonging the useful life of the components and reducing downtime. This thesis therefore demonstrates the value of rotor-bearing HIs, derived from widely available SCADA data, for wind turbine maintenance.

Accurate fault detection in wind turbines is a key factor for improving their reliability and reducing maintenance cost. This report investigates probabilistic target variable modeling using a conditional Generative Adversarial Network (cGAN) in a Normal Behavior Modeling framework for fault detection on wind turbine drivetrains. The dataset used for the investigation is the open-source Supervisory Control and Data Acquisition (SCADA) dataset provided by Energias de Portugal (EDP). A cGAN and a traditional Recurrent Neural Network (the base model) are developed and trained to model the generator bearing temperature. The performance of the cGAN is compared to that of the base model, to highlight differences associated with the modeling approach. The cGAN’s output is a conditional distribution of the target temperature, given the operational state of the turbine. From this distribution, two methods of deriving a Health Indicator are investigated: (1) By reducing the distribution to a point estimate, and (2) by taking into account the whole distribution. From the base model, a third Health Indicator is derived.
For all three Health Indicators, anomalies are detected by performing a non-parametric hypothesis test designed to detect deviations from the empirical healthy data distribution. The findings indicate an enhanced temperature modeling performance of the cGAN on train, validation and test data. Especially the robustness appears to be significantly improved. While interpreting the outcome of applying the Health Indicators for fault detection remains challenging, the results suggest that the cGAN’s probabilistic modeling improves fault detection by reducing the number of false responses.