Supervised Feature Selection Methods for Fault Diagnostics at Different Speed Stages of a Wind Turbine Gearbox

Individual condition monitoring (CM) strategies are capable to diagnose 30–40% of the defects, when they are performed individually. However, combining two or more individual CM strategies can provide more reliable information which will enhance the ability of fault detection. In this investigation, two intrusive CM strategies (vibration and lubrication oil analysis) and one non-intrusive CM strategy (acoustic signal analysis) are combined to form an integrated CM scheme. Experiments are performed on a miniature wind turbine gearbox bench top and the raw data is acquired and the defect sensitive features are extracted using discrete wavelet transform. Feature level fusion is accomplished to achieve integrated feature data set and the selection of optimal subset of significant features is done by various supervised featured selection methods. Finally, the obtained optimal feature subset is classified using SVM algorithm in order to diagnose the local defects of bearings as well as gears present in different stages of the wind turbine gearbox.