A Bearing Health Assessment Technique Based on UMAP and K-Medoids Clustering

To effectively assess bearing health, detect early failures, and construct health index (HI) curves that quantitatively reflect bearing degradation, we propose a method combining the Unified Manifold Approximation and Projection (UMAP) algorithm and K-medoids clustering for bearing health assessment. Bearing signals are characterized by 18 features in both time and frequency domains. The UMAP algorithm is applied to reduce the feature set, balancing global and local structural information using Euclidean distance-based computational indices. This results in streamlined feature sets for bearings in early failure-free and late failure stages. K-medoids clustering is then used to identify the centers of early failure-free and late failure feature sets at different dimensions, further determining the optimal dimensionality reduction. This approach retains key health-related information while minimizing dimensionality. The centers of the feature sets are combined with a membership function to calculate a health index, enabling quantitative assessment of bearing health. Experimental results demonstrate that the proposed method effectively integrates bearing health information, detects bearing health changes at various stages, and identifies early failures that are difficult to detect.