Bearing feature extraction using multi-structure locally linear embedding

摘要

The locally linear embedding aims to extract the significant features by only digging the individual geometric structure of original data set, for which the intrinsic features can not be completely expressed. In this study, two LLE-based multi-structure fusion methods are proposed. In the proposed methods, the least squares and sparse structures are first estimated in original space, and then the coefficient and function fusion approaches are introduced to integrate the least squares and sparse structures. Furthermore, the relationship between the two fusion methods are analyzed and we demonstrate that the solution of coefficient fusion is a subset of the one of the function fusion. Extensive experiments are carried out on benchmark fault data set and the bearing data set collected from our own laboratory, and experimental results indicate that the proposed multi-structure methods outperform the existing state-of-art related methods. (C) 2020 Elsevier B.V. All rights reserved.