Random forests for wind turbine radiator diagnostic measure

Offshore wind turbines are high-capital investments that need to maximise their operational capability to best provide return on their investment. These machines are equipped with remote sensing capability (SCADA systems) that can be used to provide diagnostic health measurements. This research seeks to address the problem of wind turbine cooler failures through the use of decision trees that are provided with additional domain expert crafted features that are based on the SCADA measurements. The decision trees are assessed using area under the ROC (receiver operating characteristic) curve, and the most promising are investigated in depth. The decision trees are constrained to be short to enable human interpretation, and thereby provide the potential for domain expert review. The best performing rules are shown to be highly effective at identifying clogged cooler radiators with precision rates above 80%, and that post-maintenance the rules indicate successful and effective maintenance action has been taken. This rule generation process demonstrates the potential of effective and transparent diagnostic warning that can lead to efficient maintenance activity scheduling.