Neighborhood preserving discrimination for rotor fault feature data set dimensionally reduction

Peng Wu,Kunju Shi,Mingshuai Liu,Yuanjun Dai

doi:10.21595/vp.2020.21691

Peng Wu, Kunju Shi + Show 2 more

Open Access

https://doi.org/10.21595/vp.2020.21691

Copy DOI

Abstract

NPP (Neighborhood Preserving Projections) is an incremental subspace learning methods which has a nature of maintaining the data local neighborhood geometry constant. To improve the discriminatory power of NPP, NPD (Neighborhood Preserving Discrimination) algorithm was proposed to be used for the rotor system fault data set feature dimensionality reduction. Floyd algorithm based on graph theory and MMC (Maximum Margin Criterion) were introduced in the NPP which makes NPD avoid the short-circuit problem that occurs in the high curvature high dimensional space data sets, while enhancing data discrimination information during the dimensionality reduction. In addition, NPD can maintain the manifold of data set unchanged. At last, the rotor-bearing experiment has been made to verify the effectiveness of the NPD method.

Highlights

There are significant differences between the data dimensionality reduction techniques used to describe data characteristics and those used for discriminative classification
In order to improve the discrimination ability of Neighborhood Preserving Projection (NPP), this chapter proposes the NPD (Neighborhood Preserving Discrimination) algorithm, which is used to reduce the dimension of the fault feature data set of the rotor system
The training and test sample set of the four running state sample points of the bearing, the class distance calculated by the NPD algorithm is much smaller than the result calculated by the KPCA algorithm

Summary

Introduction

There are significant differences between the data dimensionality reduction techniques used to describe data characteristics and those used for discriminative classification. LDA and its improved version-Maximum Edge Criterion (MMC) belong to this type of method [5] It tries to find a projection direction that can achieve the dual purposes of reducing the number of variables and extracting discriminant information, but its projection direction may not fully express the sample data. Neighborhood Preserving Projection (NPP) is an approximate incremental expression of the LLE algorithm It is an incremental subspace learning meth-od to describe the characteristics of data and reduce the dimension. In order to improve the discrimination ability of NPP, this chapter proposes the NPD (Neighborhood Preserving Discrimination) algorithm, which is used to reduce the dimension of the fault feature data set of the rotor system. The method was validated with the data of a two-span rotor test bench

Method of NPP

Derivation of NPD data dimension reduction formula

Design of dimension reduction method for LLD fault feature dataset

Rotor system’s original fault characteristic data set construction

NPD dimension reduction of rotor system fault data set

Conclusions