A new fault isolation approach based on propagated nonnegative matrix factorizations

Abstract

To address the challenging fault isolation problem, this paper proposes a new fault isolation approach based on propagated nonnegative matrix factorizations (PNMFs). PNMFs make significant contributions to the theoretical research on nonnegative matrix factorizations (NMFs)-based fault isolation. Specifically, PNMFs provide a new way to improve the fault isolation performance of NMFs by training a matrix-factorization model using labeled and unlabeled samples where labeled samples are the samples whose categories are already known, and unlabeled samples are the samples whose categories are unknown. Moreover, PNMFs incorporate label propagation theory into NMFs for recognizing unlabeled samples based on labeled samples. As a result, PNMFs change the learning mechanism of NMFs for improving fault isolation performance. To demonstrate the superiority of the PNMFs-based fault isolation approach, a case study on fault isolation for a penicillin fermentation process based on PNMFs and NMFs was implemented. The case study results demonstrate that the proposed PNMFs-based fault isolation approach outperforms the state-of-the-art fault isolation approaches.