Abstract: The original data sets are disjunctively distributed in the feature space frequently because of the diversity of faults; the completeness of the geometric structure cannot be keep up by the classical neighborhood structure with k-neatest neighbor. A new approach nonnegative linear least squares locally linear embedding (NLLS-LLE) is presented to solve this problem. The boundary points are searched with the nonnegative linear least squares constraints, the neighborhoods of boundary points are determined by the first principal straight line. The neighborhood graphs of boundaries are reconstructed, and the intrinsic distribution and geometry structure of data set is discovered by the classical locally linear embedding, and the fault modes are recognized accurately by KNN classifiers where the low dimensional spaces projection used as inputs. The simulation data analysis and experiment show that the proposed algorithm highly keeps geometry and topology of original data sets, enhances the validity of low dimensional spaces, and improves the classification correctness of fault.

Key words: Fault diagnosis, Locally linear embedding, Multi-manifold, Nonnegative linear least squares