Application of MMI-SVP theory in spindle fault feature extraction

Abstract

Singular value decomposition (SVD) is an effective method for estimating and separating noise in a signal. However, its performance depends on several factors, including the construction of the Hankel matrix, the number of retained reconstructed components, the length of data analysis, and even the fault feature frequency of the monitored object. This paper aims to discuss these issues and proposes an adaptive selection method of singular value pairs (SVP) based on minimum mutual information (MMI). Firstly, SVD is used to obtain singular values (SV), and the anti-angle averaging method is employed to generate a set of sub-signals for reconstruction. SV is utilized to characterize the energy of the sub-signals and determine the SVP. Subsequently, mutual information is introduced to quantitatively evaluate mutation signals and adaptively select SVP to avoid excessive or insufficient noise reduction. Additionally, the optimal dimension for the Hankel matrix decomposition is automatically determined by considering the singular value ratio and MMI index. Finally, based on prior knowledge of bearing size and fault characteristics, the minimum sample length of the Hankel matrix is determined. The effectiveness of MMI-SVP is verified through simulation analysis and real bearing fault cases.