A fault diagnosis method using Hyper-Ellipsoidal learning based Locally Linear Embedding

Fault data samples do not necessarily obey the normal distribution and often contain noise points. The reconstruction error obtained by Locally Linear Embedding (LLE) based on the k-Nearest Neighbor will be large. The completeness of geometric structure cannot be maintained by the classical neighborhood structure constructed based on the k-Nearest Neighbor. A Hyper-Ellipsoidal based LLE method (HE-LLE) is presented to solve this problem. The neighborhood graph is constructed by the neighborhood defined by hyper-ellipsoid. Then the intrinsic distribution and geometry structure of data is discovered by classical Locally Linear Embedding. The analysis result of simulation data and experiment data shows that the presented method has the ability to suppress noise. The analysis of engineering fault shows that the new algorithm can maintain geometry and topology structure of original data sets well. The recognition correctness of fault is improved.