A similarity-guided block-structured dictionary learning method for fault feature extraction of rolling bearings

Abstract

Periodical impulses generated by locally damaged bearings are frequently drowned out by harmonics and noise interference, making it challenging to accurately detect rolling bearing faults. Dictionary learning-based sparse representation provides an effective way to accurately recover fault impulses from noisy signals. However, current dictionary learning algorithms still have some deficiencies, such as low computational efficiency and insufficient accuracy of signal reconstruction. How to obtain an overcomplete dictionary whose atoms match the fault features and have low coherence between atoms is still an open problem. Therefore, a sparse representation method via similarity-guided block-structured dictionary learning is investigated to extract periodic impulse features for fault detection of rolling bearings. First, an initial signal matrix is constructed by performing period segmentation and cyclic shift operations on the original signal, and then the harmonics are eliminated column by column by least square trend estimation algorithm. Secondly, a block-structured dictionary learning method based on K-SVD is designed by using the similarity between atoms, which makes the atoms in a block as similar as possible and the atoms between blocks as different as possible, which in turn significantly improves the efficiency and accuracy of signal recovery. Finally, the block-structured dictionary optimized by soft threshold shrinkage is applied to recover the periodic impulse signal from the initial signal matrix to identify rolling bearing faults. Numerical simulation and analysis results of rolling bearing fault signals reveal that the method suggested can recover periodic impulse signals more accurately, and its denoising performance is better than some existing comparison methods.