Fault Diagnosis Method Based on a New Supervised Locally Linear Embedding Algorithm for Rolling Bearing

In view of the complexity and nonlinearity of rolling bearings, this paper presents a new supervised locally linear embedding method (R-NSLLE) for feature extraction. In general, traditional LLE can capture the local structure of a rolling bearing. However it may lead to limited effectiveness if data is sparse or non-uniformly distributed. Moreover, like other manifold learning algorithms, the results of LLE and SLLE depend on the choice of the nearest neighbors. In order to weaken the influence of the random selection of the nearest neighbors, RNSLLE, a supervised learning method, is used to find the best neighborhood parameter by analyzing residual. In addition, a nonlinear measurement based on SLLE is proposed as new criterion. In this paper, the original feature set is obtained through singular value decomposition in the phase space reconstructed by the C-C method. RNSLLE is used for nonlinear dimensionality reduction, which can further extract fault features. Following this, R-NSLLE is compared with other nonlinear methods of dimensionality reduction, such as SLLE, LLE, LTSA and KPCA. The effectiveness and robustness of R-NSLLE have been verified in the experiment, and the accuracy and silhouette coefficient of the proposed method have been further discussed. These show that this feature extraction method, which is based on R-NSLLE, is more effective and can identify the intrinsic structural of rolling bearing even when there is a little fault. Key-Words: Phase Space Reconstruction, Manifold Learning, SLLE, LLE, Rolling Bearing, Fault Diagnosis, Feature Extraction