A rolling bearing fault diagnosis method based on vibration sensor signals using adaptive VMD and correlation kurtosis feature enhancement

Abstract

In response to the serious noise interference in the fault signals obtained by the vibration sensors and the difficulty in effectively extracting the fault characteristics, a rolling bearing fault diagnosis method based on adaptive modal decomposition and correlation kurtosis feature enhancement is proposed. This method is based on the time-domain variation model of the vibration sensor output signal to obtain fault characterization features with enhanced impact and noise suppression capabilities. It uses feature reconstruction and classification networks as the core to achieve rolling bearing fault identification. The Zebra optimization algorithm (ZOA) is employed to adaptively optimize the penalty factor and the number of modal decompositions in the Variational mode decomposition (VMD), using the regularized envelope entropy as the fitness function to achieve adaptive decomposition and modal feature extraction of the vibration signal. The weighted correlation kurtosis (WCK) is used to evaluate and select the effective modes for signal reconstruction, thereby effectively suppressing modal aliasing and noise interference. On this basis, to address the issue of weak impact features being not obvious in the reconstructed vibration signal, the ZOA-optimized maximum correlation kurtosis deconvolution (MCKD) is introduced. By adaptively adjusting the filtering length and shift order, the periodic impact feature expression ability is further enhanced, and the fault characterization effect of the vibration sensor signal is improved. The experimental results show that the proposed method can effectively suppress noise interference and enhance the ability to express fault features. This method has achieved excellent diagnostic results on the self-built bearing dataset, the publicly available CWRU dataset, and the publicly available PU dataset. The fault identification accuracy rates reached 96.5%, 97.0%, and 95.9% respectively. Compared with the traditional VMD-MCKD method, the method proposed in this paper has achieved significant performance improvements on different datasets, verifying that the proposed method has better fault feature extraction ability and generalization performance under different operating conditions.