Dictionary Learning Method for Cyclostationarity Maximization and Its Application to Bearing Fault Feature Extraction

Abstract

It has been demonstrated that fast convolutional sparse dictionary learning (FCSDL) is a useful instrument for diagnosing rolling bearing faults and can recover rolling bearing fault shocks unaffected by random slippage. However, although FCSDL is not impacted by random fluctuations and can rapidly reconstruct fault shock without truncating the signal, its performance for repetitive fault shock reconstruction is not optimal when dealing with strong noise vibration signals. Therefore, this article proposes cyclostationary convolutional sparse dictionary learning (CCSDL), which is guided by fault features (cyclostationarity) to achieve the greatest signal reconstruction performance. First, the proposed method is based on the rotation frequency, and various frequency-band-covering components in the vibration signal are reconstructed successively. In the meanwhile, the harmonic significance index (HSI), which can indicate the cyclostationarity of the fault shock, evaluates the fault characteristics of each reconstruction result and finally obtains the most significant reconstruction result. Compared with FCSDL and variational mode decomposition (VMD), the proposed method performs far superior in signal reconstruction when processing low SNR vibration data.