Impulse Response Extraction for Rolling Bearing Faults Based on Semi-NMF

For the localized faults of bearing components, generally, the resulted impulse response of vibration signal would be contaminated by the noise or impacts generated by other components. This paper combines the concepts of time-frequency distribution and semi-nonnegative matrix factorization (SNMF), and proposes a time-frequency matrix factorization method to enhance identification of the impulse response. With the short time Fourier transform (STFT), the time-frequency coefficient matrix of a vibration signal is built to describe the localized impulse components. Then, the sparse semi-NMF mapping is adopted to decompose and extract the resonance frequency band of the impulse response. With the advantages of the SNMF in the frequency band decomposition and the inverse STFT, the localized fault features of interest can be extracted as well. To evaluate the performance of the proposed method, the experiment on vibration signals of rolling bearing is conducted. Compared to the standard NMF, the results reveal that the proposed method is more effective in feature extraction.