LAUSR

The installed wind power capacity is growing worldwide. Remote condition monitoring of wind turbines is employed to achieve higher up-times and lower maintenance costs. Machine learning approaches can be used for detecting developing faults in wind turbines in their earlier occurrence. However, training fault detection models may require large amounts of past and present data. These data are often not available or not representative of the current operation behaviour. These data can be acquired with supervisory control and data acquisition systems. Note also that newly commissioned wind farms lack data from previous operation, whilst older installations may also lack representative working condition data as a result of control software updates or component replacements. After such events, a turbine’s operation behaviour can change significantly so its data are no longer representative of its current behaviour. Therefore, this paper shows that cross–turbine transfer learning can improve the accuracy of fault detection models in turbines with scarce data from supervisory control and data acquisition systems. In particular, it highlights that combining the knowledge from turbines with scarce data and turbines with plentiful data enables earlier detection of faults than prior art methods. In this way, the reuse and the knowledge transfer across wind turbines allows us to overcome this lack of data, thus enabling accurate fault detection in wind turbines.