The Research Of Projective Nonnegative Matrix Factorization And Application In Machine Remain Useful Life Prediction

With the quick emergence of information technology, system complexity risesin every domain of modern industry. To enhance the ability of diagnosis and main-tainability, the importance of remain useful life prediction in prognostics and healthmanagement becomes vital. In this dissertation, a degradation trajectory similaritybased framework is used to improve the accuracy of RUL prediction model. Further-more, based on nonnegative sparse projection principle, a family of algorithms basedon projective nonnegative factorization is proposed for degradation feature extractionand achieved a higher accuracy of RUL prediction, The experimental results show thatwith the proposed algorithm for RUL prediction we can gain a better interpretationwhile pertains good predict result:To improve the efectiveness of feature extraction under non-gaussian and seriescorrelation condition, based on Bayesian sparse coding principle, a new AutoRelevance Determination Projective Nonnegative Matrix Factorization algorith-m is proposed for fnd better basis for fnding feature mode, which has a betterinterpretation.To incorporate neighborhood information for fnding more accurate basis for ex-tracted feature，neighbor preserving embedding is imposed to the original pro-jective nonnegative matrix factorization algorithm, and shows improvement infnding better feature for RUL prediction.To improve efectiveness and interpretability for regression model of health in-dex, a representative based selection strategy for Gaussian process regression modelling based on representor theorem is proposed, shows a faster learningspeed, and a better generalization ability in RUL prediction.