The Integration Of Global And Local Feature Information Of Rotating Machine Fault Data Set Dimension Reduction Method

Fault diagnosis technology is one of the effective means that ensuring the safe of the machine equipment operation.In order to obtain more accurate and effective diagnostic results,it is effective to extract the valuable information of equipment.So,how to obtain effectively the real reflecting fault information data which has become a hot research topic in nowadays’ fault diagnosis.It is well known that machinery fault diagnosis is mainly divided into three steps: signal acquisition,feature extraction and pattern recognition.And dimension reduction is the key step of feature expression,which can reduce the pressure of the subsequent fault pattern recognition even extract the essence of fault information.Research shows that the global and local information of data for dimensionality reduction and classification are beneficial.From the perspective of information extraction,the traditional dimension reduction algorithms,such as principal component analysis(PCA)and linear discriminant analysis(LDA)and local keep projection(LPP),which cannot obtain the good effect of dimensionality reduction and classification.According to the above problem,the dimension reduction extraction method based on take into account the global and local which are widely studied.In order to make fault information to retain the more perfect,and diagnosis effect better,this paper is based on global and local dimensionality reduction to extract the fault feature set,our work mainly includes the following contents:1)From the perspective of dimension reduction to keeping data structure,under the basis of the contrast analysis of global and local dimension reduction methods,this paper expounds the importance of dimension reduction for global and local information.2)When training samples are insufficient,local information is more important than global information.Therefore,on the basis of the local RKMMP dimension reduction method,a novel fault diagnosis method of the rolling bearing based on dimension reduction with Regularized Kernel Maximum Margin Projection(Regularized Kernel Maximum Margin Projection,RKMMP)is proposed.Firstly,in the method,using RKMMP to reduce the dimension of mixed fault data set.Then,after the dimension reduction,sensitive feature subset of low-dimensional is input into Kernel Extreme Learning Machine(Kernel Extreme Learning Machine,KLEM)classifier for training and fault identification.The experiments of rolling bearing fault simulation show that the method,to a certain extent,can improve the generalization ability of fault diagnosis and recognition accuracy.3)Aiming at the problem of traditional dimension reduction methods cannot take into account extraction of the global feature information and local discriminant information.A kind of method of dimension reduction of the rotor fault dataset based on Kernel Principal Component Analysis(Kernel Principal Component Analysis,KPCA)and Orthogonal Locality Sensitive Discriminant Analysis(Orthogonal Locality Sensitive Discriminant Analysis,OLSDA)is proposed.Firstly,KPCA algorithm in the method can retain the maximized original data information of global nonlinear.Then,OLSDA algorithm is used to excavate fully local manifold structure information of data,which implements the propose that extracting the low-dimension essential feature with high discrimination.Rotor experiment showed that this method can comprehensively extract the global and local discriminant information,which made the classification of faults more clear and corresponding recognition accuracy rate was improved significantly.4)The propose of dimension reduction is that keeping fault information more comprehensive and making fault diagnosis result more accurate.KPCA-OLSDA joint dimension reduction method is a certain extent increased the complexity of the algorithm.For as much as possible to keep data intrinsic global and local information,a dimension reduction method for rotating machinery fault diagnosis that uses Global and Local Locality Sensitive Discriminant Analysis(GLLSDA)is proposed.Firstly,method extracts multiple-domain multiple-channel statistical features from vibration signal.And high-dimensional dataset is built.Secondly,the high-dimensional fault dataset is trained and reduced by using the proposed GLLSDA algorithm.Therefore,low-dimensional sensitive feature subset is extracted.Finally,the low dimensional feature subset is input into Weighted K-Nearest Neighbor(WKNN)classifier for fault diagnosis.The rotor system experimental data are applied to verify the effectiveness of new method.