A comparison of feature ranking techniques for fault diagnosis of ball bearing

In rotating machinery one of the prominent causes of malfunction is faults generated in ball bearings, therefore, diagnosis and interpretation of these faults is essential before they become severe. Feature extraction methodology has been presented in this paper based on application of lifting wavelet transform. Minimum permutation entropy is considered as decision making for selecting level of lifting wavelet transform. Sixteen features are calculated from measured vibration signals for various bearing conditions like defect in inner race, outer race, ball defect, combined defect and no defect condition. To achieve better fault identification accuracy selection of features carrying useful information is needed. To select highly distinguished features various ranking methodologies such as Fisher score, ReliefF, Wilcoxon rank, Gain ratio and Memetic feature selection are used. The ranked feature sets that are fed to machine learning algorithms support vector machine, learning vector quantization and artificial neural network for identification of bearing conditions. Tenfold cross-validation results show that selected features give enhanced accuracy for detecting faults. Features selected through Fisher score-support vector machine and ReliefF-artificial neural network gives 100 % cross-validation accuracy. Result shows that proposed methodology is feasible and effective for fault diagnosis of bearing with reduced feature set.