Radial basis function neural network for effective condition monitoring of rolling element bearing

The performance of the Radial Basis Function Neural Network (RBFNN) in the defect classification of a Rolling Element Bearing (REB) has been investigated in this work. The features (input) required for training the RBFNN have been extracted from the non-overlapping segments of the raw and denoised bearing vibration signals. A kurtosis based wavelet denoising method has been used to reduce the noise components in the vibration signals. The Fisher's Criterion (FC) has been used to select a few sensitive features and form a reduced feature set. The centers of the RBF units have been optimized using a modified Fuzzy C-Means (FCM) algorithm, viz., Cluster Dependent Weighted FCM (CDWFCM). The performance of the RBFNN has been compared for four training strategies: two types of feature sets (all features and FC selected features) and two types of RBF centers selection method (centers selected randomly and centers selected using CDWFCM). These strategies have also been tested for the bearing vibration signals provided by the Case Western Reserve University database.