Local and Nonlocal Preserving Projection for Bearing Defect Classification and Performance Assessment

The sensitivity of various physical features that are characteristics of bearing performance may vary significantly under different working conditions. Thus, it is critical to extract the most effective information from the original physical features generated from vibration signals for bearing defect classification and performance degradation assessment. This paper proposes a local and nonlocal preserving projection (LNPP)-based feature extraction algorithm, which is different from principal component analysis that aims to discover the global structure of Euclidean space and regular manifold learning algorithms that attempt to preserve local structure in data. LNPP is capable of discovering both local and nonlocal structures of data manifold. This may enable LNPP to find the meaningful low-dimensional information hidden in the high-dimensional feature set and then to serve as a preprocessor for defect classification. Furthermore, an LNPP-based quantification index is proposed for the assessment of bearing performance degradation. An LNPP-based contribution plot for feature selection is developed to improve the degradation detection sensitiveness and to reduce false alarms. A dynamic LNPP for bearing performance assessment is further developed to consider inherent autocorrelation existing in vibration data. Detailed results are very promising and are reported in this paper. This paper will provide guidance for the applications of manifold learning algorithms (e.g., LNPP) in machine fault diagnosis and performance prognostics.