Criterion fusion for spectral segmentation and its application to optimal demodulation of bearing vibration signals

The defective bearing signatures can be detected by resonance demodulation of the vibration signals. The decision of the bearing fault detection largely depends on the quality of the identified resonant frequency band. Two key issues in locating the resonance frequency band are the proper segmentation of the frequency spectrum of interest and the criterion used to guide the search for the resonance band. To deal with the two issues, this paper proposes a criterion fusion approach to guide the spectral segmentation process. With the proposed approach, the frequency spectrum of the bearing signal is first divided into initial fine segments which are then adaptively merged into different subsets using an enhanced bottom-up segmentation technique. To guide the spectral segmentation and merging process, three commonly used criteria, i.e., kurtosis, smoothness index and crest factor are fused into a synthesized cost function using an entropy-based method. The final frequency band delivered by this approach has a good coverage of the resonant band and is then used to demodulate bearing signals. Both simulated and experimental signals have been employed to evaluate the proposed approach, which has also been compared to single-criterion methods. The comparison indicates that the fused criterion yields better results than those from the single-criterion. (C) 2015 Elsevier Ltd. All rights reserved.