Abstract: Due to the existence of noises, a statistical criterion of the vibration signal can be used to guide the identification of the resonance frequency band, which is then demodulated for extracting the faulty signature of the bearing. The commonly used resonance demodulation methods, such as the Kurtogram, coarsely portion the vibration band by empirically are used a top-down strategy. Employing only one criterion for the resonance band identification, moreover, is often interfered by the noises and hence being lack of robustness. To improve the accuracy of the bearing fault diagnosis, a fuzzy fusion technique using multiple criteria is proposed for the optimal band demodulation. In the present method, a fine approximation is first generated for the vibration spectrum. The fine spectrum is then merged bidirectionally using a bottom-up strategy. In this way, the frequency band can be segmented precisely to achieve the minimum cost function. In the present method, the cost function is constructed by a fuzzy neartude-based data fusion of kurtosis, smoothness index and crest factor, and therefore leading to the robust resonance band. Both simulated and actual signals are collected for testing the proposed technique. The results show that, comparing to the existing mono-criterion methods, the proposed technique is capable of correctly diagnosing the condition of the rolling element bearings.

Key words: fault diagnosis, fuzzy data fusion, rolling element bearing