Research On Mainfold Learning-Based Performance Degradation Of Momentum Wheel Bearing

The development of aerospace technology calls for increasingly high reliability for satellites.Momentum wheel(MW)is a typical actuator,which plays a significant role on stabilizing the spacecraft and carrying out on-orbit scheduled missions.It is of great importance for MW to perform well as expected while fault takes place accidently due to complicated orbital environment and unknown circumstance.Bearing malfunction is the main cause of MW failure.This thesis focuses on the research how to figure out the degradation of momentum wheel bearing(MWB)performance from numerous vibration datasheet.A high-dimension feature set is formed by the indexes of multi-domain.Adaptive neighbor locality preserving projection(ANLPP)algorithm and improved discriminant locality preserving projection(IDLPP)algorithm are developed to extract the features of MWB from high-dimension sets.Fuzzy c-means clustering algorithm is used to assess the performance degeneration,and by this means the degree of the health of MWB is calculated.After that,a dynamic ARMA model is built to predict its future performance.The testing datasheet of NASA is used to demonstrate the effectiveness and validity of the proposed method.Firstly,a high-dimension feature set including characteristics in time domain,frequency domain and time-frequency domain is formed.Followed that,test data is pretreated by wavelet filtering.Not all the features of performance degradation of MWB are contained just by the information from either time domain or frequency domain.As such,a high-dimension feature set is formed by multi-domain information,including time domain,frequency domain,and time-frequency domain,to investigate the exact and full-scale characteristics of the performance while MWB gets a change undergoing degradation.Secondly,ANLPP and IDLPP algorithm are proposed to extract the features from high dimension sets.With consideration of its sensitivity to the number of neighbors,LPP algorithm is not receivable..For this sake,the algorithm of adaptive neighbor locality preserving projection(ANLPP)is developed to adjust neighbor number according to sample density to make the result more consistent and closer to a real mainfold structure.However,ANLPP algorithm is a kind of unsupervised method,by which a lot of information of labeled samples can not be integrated in the round.Therefore,an improved algorithm named discriminant locality preserving projection(IDLPP)is proposed.IDLPP algorithm combines information of maximum margin criterion(MMC)and ANLPP,which can distinguish the within-class samples and between-class samples.Simulations demonstrate the superiority and desired performance of the proposed IDLPP in feature extracting.Finally,the performance degradation of MWB is analyzed.The algorithm of fuzzy c-means clustering is applied to find cluster center of normal samples and health index(HI)is introduced to describe the affiliation of each sample to the cluster center.Followed by this,quantitative evaluation of MWB’s health can be carried out.A prediction model using dynamic ARMA algorithm is developed with the help of HI,which takes advantage of selecting model window dynamically.Prediction results show the effectiveness of proposed method.