Research On Theory And Methodology Of Manifold Feature Enhancement For Machine Condition Monitoring And Diagnosis

With the research aim of precision state identification and effective fault diagnosis of machinical equipment,this dissertation addresses on the mining capability of manifold learning on the essential feature information.Through analyzing the state feature,pattern characteristics and signal characteristics of condition-based maintenance,we correspondingly proposed three novel model in the field of manifold learning,including reference manifold,deep manifold and sparse manifold,in which the manifold reinforce learning of discriminative features of state,sensitive characteristics of pattern and modal characteristics of signal can be realized.Thus a manifold learning enhancement approach is built for the research system of machine condition monitoring and fault diagnosis.Meanwhile,these three manifold characteristics are separately studied in depth as follows.In the case of state feature enhancement learning,according to the physical truth that the equipment has a long-lasting healthy operation and there would be a big difference between the healthy state and fault state,a novel real-time referenced model,combining the same healthy state and the real-time monitoried samples together,is established to compose the feature space reference manifold by manifold learning.The performance degradation of the state can be described by analyzing the migration and variation of the manifold feature space clustering.The reference manifold feature clustering distribution can highlight the variety information of the monitoried state compared to the health state,and reveal the discriminative features.According to the difference of feature construction and manifold space,this section developed two kinds of manifold space clustering with multivariate statistical features and non-characteristic Rasmussian manifold,respectively.Moreover,a novel spatial clustering chart based on manifold feature clustering is proposed to realize multi-stage assessment analysis of equipment performance degradation and a significant alarm at the early degradation.In the case of pattern characteristic enhancement learning,according to the effect of the reference manifold on the enhancement of discriminative feature,the original features are established to expand the learning process by implementing multi-unit referenced models,then cascade learning can be further achieved by deep learning.In these manners,a deep manifold learning model,similar to neural networks,is built,and the sensitive pattern characteristics are enhanced in this feature expansion and learning process.Utilizing the multi-layer manifold,this section propose a two-layer manifold feature enhancement with multi-uint referenced models,named as fusion feature based on multiple reference model.The proposed approach can effectively enhance the difference among different fault types and improve the sensitivity of pattern characteristic.Furthmore,it shows superiority in fault classification as compared to the traditional manifold feature recognition method.In the case of signal characteristic enhancement learning,with the merits of time-frequency manifold learning in extracting essential information of transient feature,sparse analysis theory is introduced on the basis of complementary advantages,and a novel sparse manifold feature learing model is built.This can give a deep feature mining and relearning for signal sparse mode characteristic,which can overcome the shortcomings of the traditional time-frequency manifold learning methods in transient characteristic distortion and strong noise removal.Based on the idea of sparse manifold analysis,this section proposes manifold mode enhancement with time-frequency sparsely matching and envelope-invariant sparsly pursuit,which can realize the preservation and recovery of the original signal waveform,which is conducive to accurate fault diagnosis of machinical equipment.In summary,this dissertation explores the application of manifold feature enhancement in condition-based maintenance for machinical equipment,including three ascepts:condition monitoring and early warning,fault pattern classification and fault signal diagnosis,proposed reference manifold,deep manifold and sparse manifold for feature enhancement learning,thus a more complete feature enhancement learning was carried out.Compared to the traditional methods,the manifold feature enhancement can realize a more effective,more precise and more sensitive analysis of condition monitoring and fault diagnosis,which is of great significance to the systematic study of condition-based maintenance.