Semi-supervised hierarchical attribute representation learning via multi-layer matrix factorization for machinery fault diagnosis

Data-driven fault diagnosis methods have become a research hotspot recently. However, the following two problems are still barring them from the application: (1) Most of the existing models rely deeply on sufficient labeled samples and neglect the high cost of labeled data collection in reality; (2) The existing models usually focus on the single-level attribute of the sample and ignore the latent hierarchical fault attributes. To address these issues, a novel semisupervised multi-layer non-negative matrix factorization (SMNMF) method is proposed in this study. The fault pattern and severity identification problems are converted into a hierarchical fault attribute representation task, which can reduce the complexity of the classification task and improve the fault diagnosis accuracy. The hierarchical attribute representations of different fault locations and sizes are learned from the time-frequency distribution (TFD) of signals by a newly constructed two-layer non-negative matrix factorization model. The graph-based semi-supervised learning method is adopted to lead the attributes of the hierarchy structure and carry out label propagation from labeled samples to unlabeled samples for more accurate fault diagnosis. The fault diagnosis experiments executed in the aeroengine bearings and a diesel engine demonstrated the feasibility and superiority of the proposed method.

Automated summary

A novel semi-supervised multi-layer non-negative matrix factorization method is proposed in this study for fault diagnosis, which learns hierarchical attribute representations of different fault locations and sizes from the time-frequency distribution of signals and uses semi-supervised learning to improve fault diagnosis accuracy. Experiments in aeroengine bearings and a diesel engine demonstrate the feasibility and superiority of the proposed method.