An Improved Shift-Invariant Sparse Bearing Fault Diagnosis Method Based on Feature Learning Base Atoms

Bearings, as crucial components in the power systems of high-end equipment, are susceptible to noise and other redundant elements when faults occur. In this paper, an improved shift-invariant sparse (ISiS) feature extraction method based on the feature learning base atom (FLBA) is proposed to extract the weak features in the signal. Firstly, a low-dimensional learning dictionary consisting of base atoms is acquired by shift-invariant K-Singular Value Decomposition (K-SVD) learning. The FLBA selection method based on the kurtosis criterion is proposed to select the dictionary base atom with the largest kurtosis value for matching the feature information in the signal. Secondly, the ISiS model is proposed, which aims to minimize the iteration residuals, thus reducing redundant iterations at the same index position and improving the parsing ability of features. Finally, the effectiveness of the algorithm is verified through experimental analysis of simulated signals containing periodic impulses and bearing fault signals. Meanwhile, the advantages of the proposed model are confirmed by the comparative analysis experiments of the generalized minimax concave sparse enhancement diagnostic method and the fast spectral kurtosis method.