Local and global SR for bearing sensor-based vibration signal classification

Spectral regression (SR) is a method of feature extraction that realizes dimension reduction by the least squares method and can avoid eigen-decomposition of dense matrices. However, it only considers the affinity graph and misses the global information. In this paper, a novel feature extraction algorithm, called local and global spectral regression (LGSR), is proposed and applied to extract fault features from frequency-domain and time-domain features of vibration signals of bearing sensors. LGSR, which is the development of SR, is able to discover both local and global information of data manifold. Compared with other similar approaches (such as NPE, PCA, and SR), experiments of bearing defect classification validate that LGSR shows better ability to extract identity information for machine defect classification.