Cyclogram: an effective method for selecting frequency bands for fault diagnosis of rolling element bearings

Frequency band selection for repetitive transient extraction using the kurtogram and its variants plays a vital role in fault diagnosis of rolling element bearings. However, cyclostationarity, one of the most typical symptoms of faulty bearings, is always neglected in these methods, leading to failure of the extraction of the weak fault features. To address this shortcoming, a novel method for selecting frequency bands, called Cyclogram, is here proposed based on kurtosis and cyclostationarity. In the proposed method, a signal is decomposed into several signals in different frequency bands by a wavelet packet transform, and squared envelopes (SEs) are calculated for these decomposed signals. Then, a robust indicator of SEs for evaluating repetitive transients is constructed based on cyclic spectral coherence and kurtosis, which helps to select useful frequency bands. Afterwards, the envelope spectrum of these selected frequency bands are averaged rather than only selecting one frequency band to enhance fault features. Compared with traditional fault-diagnosis methods for rolling element bearings, the proposed method is able to identify faults from signals corrupted seriously with Gaussian and non-Gaussian noise. The effectiveness of Cyclogram is validated based on simulation and three real-world vibration signals from faulty bearings.

Automated summary

Frequency band selection is crucial for fault diagnosis of rolling element bearings using the kurtogram and its variants. Existing methods often neglect the cyclostationarity, a typical symptom of faulty bearings, resulting in the inability to extract weak fault features. To address this issue, a novel method called Cyclogram is proposed, which combines kurtosis and cyclostationarity to select frequency bands effectively. The proposed method decomposes a signal into different frequency bands using a wavelet packet transform, calculates squared envelopes for these bands, and constructs a robust indicator based on cyclic spectral coherence and kurtosis. This method outperforms traditional fault-diagnosis methods and can identify faults in signals corrupted with Gaussian and non-Gaussian noise.