Multilevel Fault Information Fusion Based on Deep Block-Wise Semi-Nonnegative Matrix Factorization for Industrial Fault Detection and Diagnosis.

This article proposes a deep block-wise semi-nonnegative matrix factorization (DBSNMF) model for multilevel fault information fusion in industrial fault detection and diagnosis (FDD). DBSNMF first partitions variables into subblocks through semi-nonnegative matrix factorization (Semi-NMF) and Fast Unfolding to extract localized fault-related features. In each variable subblock, a deep Semi-NMF (DSNMF) architecture is built by stacking multilayer Semi-NMF models over concatenated feature and residual subspaces, thereby jointly exploiting information from both spaces. Then, we design a multilevel information fusion approach based on Bayesian inference and a fault-highlighted weighting strategy to fuse layer-, block-, and subspace-level information into a comprehensive fault detection statistic and a unified diagnosis index. Experiments on the continuous stirred tank reactor (CSTR) and Tennessee Eastman (TE) processes show that DBSNMF achieves the highest fault detection rates (FDRs) and the best abnormal-variable diagnosis performance.