Weak Fault Feature Extraction of Rolling Bearings Using Local Mean Decomposition-Based Multilayer Hybrid Denoising

Abstract

Extraction of the weak fault features under strong background noise is crucial to early fault diagnosis in bearings. A new method called local mean decomposition (LMD)-based multilayer hybrid denoising (LMD-MHD) is proposed for signal denoising in this paper. LMD is a novel self-adaptive time-frequency analysis method. It can decompose the signal into a set of product functions (PFs), and is thus particularly suitable for processing of multicomponent amplitude-modulated and frequency-modulated signals. The first filtering layer of LMD-MHD is to use a multiple criteria decision to select the effective PF components. The second filtering layer is to use the wavelet threshold denoising (WTD) as the prefilter of singular value decomposition (SVD) implementation, which enables the preserved singular values to consist of the most important information of the PFs. The final denoising layer of LMD-MHD is to use SVD to extract the most important principal features from Hankel matrix of the PFs. The order of effective ranks of the Hankel matrix is determined by the number of main frequencies in the fast Fourier transformation (FFT) result of the signal. The filtering is finished based on the reconstructed signal from the decomposition result of SVD. The results of experimental analysis on the simulation signals and vibration signal collected from rolling bearings indicate that LMD-MHD is effective for extracting the weak fault features and performs well for bearing fault diagnosis.