Wind farm operators can profit substantially from shifting towards predictive maintenance strategies rather than employing corrective tasks using SCADA-based condition monitoring systems (CMS). Nevertheless, the black-box character found in the market inhibits the generalised use of these systems.

This project will combine all the available data from a wind turbine, alarms, low and high- frequency SCADA and maintenance reports, with explainable machine learning and probabilistic techniques to develop a prognosis tool for the preventive maintenance of the turbine.

Unsupervised learning using classification methods will first allow the identification of important features in the datasets. Importance Ranking and Principal Component Analysis will help to reduce the feature space dimensionality providing a selected set of features.

Finally, supervised classification and regression techniques, will be applied to the selected set of features to obtain prognosis models of failures including uncertainties.

The models will be grouped in a practical tool /app capable of anticipating component failures in the turbine (with an explainable method), including the uncertainty, which will help indecision making for when to repair a component.

The fellow will be trained on the use of the different covered machine learning techniques to obtain and develop research and industrial results. Research results will be obtained at three different levels:

• first, unsupervised ML will provide a set of features of the Wind Turbine SCADA data to be considered;
• second, the features dataset will be reduced by selecting the most appropriate methods;
• third, supervised ML methods will generate models for the prediction of failures. As the industrial result, the prognosis tool (concept) will be developed from the obtained models and validated with real data and benchmarking tools.

Two secondment periods of 3 months each are foreseen to support collaboration on data driven ML predictive maintenance with TU-DELFT (Prof. Simon J. Watson, M24-M26) and to test the final tool with ANNEA (Dr. Maik Reder, M35-M37).