A new fault feature for rolling bearing fault diagnosis under varying speed conditions

Introduction

Rolling bearings are widely used in rotating machines, and bearing failure is one of the most common causes of machine breakdown and accidents.¹ Thus, it is important to study the defect diagnosis technology of rolling bearings. Vibration signals usually carry rich information about mechanical conditions and are often used to diagnose the faults of bearing’s localized defects.^2–4

During the past decades, many intelligent fault diagnosis approaches have been proposed to analyze vibration signals. Fast Fourier transform is the most common method proposed in Corinthios.⁵ Theoretically, periodic impulses will be generated when the rolling elements pass over the defect in time domain and a distinct profile emerges on the frequency spectrum. However, in practice, it is much more complex, almost no machine can work under stationary condition.^4,6 Under varying operation conditions, the vibration signals collected from rolling bearing systems usually carry heavy background noise and the fault characteristic frequency is not only modulated as a series of harmonics but also is smeared on the frequency spectrum.^7,8 Therefore, existing techniques based on the assumption of working in stationary or approximate stationary condition such as FFT cannot work well in extracting the overwhelmed remarkable information for fault diagnosis. In order to solve this problem, researchers have proposed numerous methods based on time-frequency analysis (TFAs) and have made some achievements. However, each of these methods has its own limitation.

Short-time Fourier transform (STFT) was presented in 1947,⁹ with a fixed length window function sliding along the time axis to intercept the signal into several segments. Corresponding frequency and amplitude variation along with time are obtained, but for the restriction of Heisenberg uncertainty principle,^10–12 the frequency and time resolutions may not be ideal at the same time. Wavelet transform (WT) was proposed in 1984 by Grossmann and Morlet,¹³ which probably posed a considerable ideal resolution, for that the basic wavelet function can be modified according to the specific needs of specific applications.¹⁰ According to different types of data analyzed, WT is divided into two classes, that is, continuous wavelet transform (CWT) and discrete wavelet transform (DWT), both of which have been developed successfully in the fault diagnosis field.^14–17 Whereas, selecting a suitable wavelet basis, which is the key...