Shift-Invariant Sparse Filtering for Bearing Weak Fault Signal Denoising

Weak fault signal caused by defects in the raceway and roller is significant for bearing fault diagnosis, and the noise makes those weak fault signal hard to recognize. Sparse learning, an adaptive learning method, has great potential in weak fault signal detection under strong background noise. However, mode construction and sparse coefficient solution are the two principle problems for sparse learning. Focusing on these issues, this study proposes a new shift-invariant sparse filtering (SISF) method for weak signal denoising. It extracts modes directly from the sparse mapping process rather than sparse results and locates the impulses in the fault signal by a convolution sparse method, so as to achieve the purpose of strengthening the weak fault feature. Primarily, the latent modes can be adaptively mined by sparse filtering (SF) from short-sequence signal segments in a self-learning way. Futhermore, phase space reconstruction (PSR) combined with singular value decomposition (SVD) is employed to enhance the latent modes. With the input of the entire signal and the mined modes provided by optimal mode searching, a convolution sparse way is applied to find the position of the modes in the signal. Finally, notice that the values at different locations represent the magnitude of the defect impulse, so removing the smaller values will achieve the effect of noise reduction. Through experiments and comparative results, it verifies that the SISF can better denoise and enhance the weak fault characteristics and is helpful for the accurate diagnosis of bearing faults.