Wind turbine gearbox planet bearing failure prediction using vibration data

This paper presents a methodology for predicting planet bearing failures utilising vibration data acquired through accelerometers installed on the gearbox surface. The proposed methodology applies certain signal pre-processing techniques in order to remove the speed variations of the turbine and separate the stochastic bearing components from the deterministic gear ones. Then, spectral kurtosis is used to enhance the impulsiveness of the bearing fault signatures and envelope analysis is used to demodulate the signal. Features are extracted from the envelope spectrum and are used as an input to a classification model. The classification labelling is performed based on the time before failure. The methodology is tested on real offshore wind turbine vibration data collected at various times before failure. The performance of the classifier is assessed using k-fold cross validation. The results are compared with methods of classic envelope analysis that uses a constant demodulation band.