Abstract: It is a common practice of rolling bearing fault diagnosis by choosing rolling bearing resonance band based on kurtosis index and demodulating fault features. Kurtosis can only represent the strength of transients' features, but it cannot differentiate impulse noise and transients, which is cyclical generated in rolling bearing vibration signals. As a result, it leads to the inaccurate identification results of the rolling bearing's resonance band and the unsatisfactory demodulation results of the fault features. In order to overcome these shortcomings, this manuscript redefined the correlated kurtosis and put forward the constructor method of correlated kurtosis in squared envelope spectrum. Combining Morlet wavelet filtering and particle swarm optimization, the proposed method can adaptively choose the optimal resonance band of rolling bearings and demodulate rolling bearing's fault features. The analysis results of simulation data and test data showed that compared with index, such as kurtosis and envelope spectral kurtosis, the proposed correlated kurtosis in squared envelope spectrum can overcome the lack of kurtosis-based index. Moreover, combining optimal resonance frequency demodulation, it can select the exact optimal resonance band's central frequency and bandwidth of the rolling bearing. Improved results of the rolling bearing fault diagnosis verified the validity and advantage of this method.

Key words: fault diagnosis, resonant band demodulation, rolling bearing, squared envelope spectral correlated kurtosis