Periodic impulse feature separation by combination of bi-frequency map and non-negative matrix factorization

Abstract

For the fault detection of rolling element bearings, generally, the periodic impulsive fault features would be mixed with other cyclic components in the spectral coherence map of the raw vibration signal. Authors aim to extract faint fault features from the bi-frequency map, especially in the incipient fault cases. To achieve this goal, this paper presents a novel method termed Double Spectrum (DS), where Non-negative Matrix Factorization (NMF) is used to analyze bi-frequency maps in two steps: firstly, the basis matrix helps identify the cyclic frequencies where the high-amplitude cyclic components locate, and secondly, the activation matrix is considered as a set of filter bands which are adopted for further demodulation and envelope analysis. We combine the information contained in the basis matrix and the activation matrix by multiplying the enveloped spectrum and the corresponding column of basis matrix to enhance the fault features. Tests on the experimental data and comparisons to the state-of-the-art diagnosis techniques have verified the performance of the proposed method.