Abstract
Abstract
Vibration signals provided by rotating machinery are informative signals about their operating states. By nature, the vibration signal behavior is non-stationary. To this end, the extraction of discriminating and fault-sensitive parameters is a major challenge in the field of monitoring rotating machines. Conventional fault diagnosis methods based on signal processing use statistical feature parameters in time domain, frequency domain and time-frequency representation. In this article, a new method is proposed for the detection and classification of bearing defects based on spectral subband using frequency membership functions. Statistical parameters including subband energy, Center frequency, root variance frequency and Shannon entropy are considered. Compared to the common features, the extracted parameters can provide discriminating information. These feature parameters are finally fed into a generalized RBF neural network system trained with the Resilient Backpropagation (Rprop) algorithm to classify seven pre-established fault types in ball bearings operating under multiple shaft speeds and load conditions. The results suggest that the proposed system can significantly improve the diagnostic performance in terms of accuracy and estimation of the bearing fault level.