A New Feature Extraction for Rolling Bearing Using Sparse Representation Based on Improved K-singular Value Decomposition and VMD

Abstract

Bearings are an essential part of the rotating machinery, which are easy to fail prematurely because of the complex working environment. When rolling bearings are damaged, the multi-component coupling is constantly vibrated, which makes fault signals appear nonlinear and nonstationary. The periodic shock components associated with the fault information are mixed with a great many noise. Unluckily, existing k-singular value decomposition (KSVD) algorithm is susceptible to interfering by noise in the process of dictionary learning. In addition, KSVD algorithm usually needs prior knowledge for determining the key parameters, which cannot effectively extract fault features. To address above problems, this paper proposes a sparse representation framework based on improved KSVD dictionary learning and variational mode decomposition (VMD) algorithm (IKSVD-VMD). The hidden fault characteristics are decomposed by VMD. The optimal intrinstic mode functions (IMF) component is selected by quantitative indexes. Then, the spectral negentropy constraint index is introduced to change the constraint condition of KSVD algorithm. Finally, fault types can be identified effectively with envelope analysis of reconstructed sparse signals. The experimental signal and measured signal prove that IKSVD-VMD effectively extract bearing fault features, which outperform traditional dictionary learning algorithm.