Morphological Analysis Based Adaptive Blind Deconvolution Approach for Bearing Fault Feature Extraction

How to accurately extract the fault related periodical impulses is the key to bearing fault diagnosis. The blind deconvolution (BD) method has been positively affirmed its ability in this field. However, the experience dependent parameter-setting and vulnerable to interference under complex working condition are two main problems that seriously limit its application. To address these issues, an improved BD method, named adaptive morphological BD, is proposed in this article. A new indicator, the morphological frequency negentropy, is first constructed through morphological analysis and adopted as the objective function for deconvolution. With its robustness to random impact and noise being verified, the optimal Morlet wavelet filter is selected with morphological frequency negentropy (MFN) and used as the initial filter. The sampling matrix is enhanced with varying morphological filtering and its size is adaptively determined by power spectral density. Through adaptive setting of the length of the filter, the dependence of prior knowledge for parameter setting is therefore reduced. Finally, the diagonal slice spectrum is applied on the filtered signal to remove in-band and residual noise. The effectiveness of the proposed method is validated by simulation signal and real datasets. Comparison analysis with other typical filter methods further shows its superiority.

Automated summary

This article proposes an improved adaptive morphological BD method for accurately extracting fault-related periodic impulses in bearing fault diagnosis. By constructing a new indicator called morphological frequency negentropy and selecting the optimal Morlet wavelet filter as the initial filter, the method's robustness is enhanced. By adaptively setting the filter length and enhancing the sampling matrix, the dependence on prior knowledge for parameter setting is reduced. Finally, the diagonal slice spectrum is used to remove residual noise. The effectiveness of the method is validated through simulation signals and real datasets, and comparison analysis with other filter methods demonstrates its superiority.