Label consistent semi-supervised non-negative matrix factorization for maintenance activities identification

Abstract

Health prognostic is playing an increasingly essential role in product and system management, for which non-negative matrix factorization (NMF) has been an effective method to model the high dimensional recorded data of the device or system. However, the existing unsupervised and supervised NMF models fail to learn from both labeled and unlabeled data together. Therefore, we propose a label consistent semi-supervised non-negative matrix factorization (LCSSNMF) framework that can simultaneously factorize both labeled and unlabeled data, where the discriminability of label data is preserved. Specifically, it firstly incorporates a class-wise coefficient distance regularization term that makes the coefficients for similar samples or samples with the same label close. Moreover, a label reconstruction regularization term is also presented, as the classification error with coefficient matrix of labeled data is expected as low as possible, which will potentially improve the classification accuracy in maintenance activities identification for industrial remote monitoring and diagnostics. The experiment results on real maintenance activities identification application from PHM 2013 data challenge competition demonstrate that LCSSNMF outperforms the state-of-arts NMF methods and results provided by the competition.