Creation of a tool for predictive maintenance in wind turbines

Over the years wind turbines have become even more complex pieces of engineering. Given the huge planned growth for the upcoming years, reducing maintenance and downtime costs is key to maximize efficiency. Detecting failures ahead of time and being able to schedule maintenance accordingly is a live challenge in this industry. Performance monitoring is usually based on Supervisory Control and Data Acquisition (SCADA) information and the expertise of the analysts. With the new technology improvements and the new Artificial Intelligence (AI) techniques, a new framework of possibilities is opened in this topic to help decision makers understand hidden patterns in the data. This thesis goes over the process of creating a Python tool based on AI to predict the Time To Failure (TTF) of different components of several wind turbines building the necessary models and algorithms based on works, events and SCADA measures. After analysing the given data the modeled components are the blades and yaw systems, the generator, the gearbox and the generator bearings. The predictions are based on the best of the following four algorithms in each particular case: Linear regression (LR), Gradient Boost- ing Regressor (GBR), Support Vector Regressor (SVR) and Decision Tree Regressor (DTR). To determine the optimal model R2 score has been used to test the correlation. After testing all the previous models and successfully building up to 152 predictors the thesis reasures what had previously been found in the literature review; that ensemble algorithms work best for this kind of regression cases. Moreover, an analysis of the most relevant features is performed to successfully double check the physic resemblance and the reliability of the models. Finally, a cost estimation of the potential benefits of implementing the predictive maintenance algorithm is assessed together with its environmental impact