Statistical process monitoring based on modified nonnegative matrix factorization

Abstract

As an effective feature extraction and dimension reduction method, nonnegative matrix factorization (NMF) can yield sparse, spatially localized, part-based subspace representations by finding a low-rank matrix approximation from the original data. However, one drawback it has to suffer is its failure to retain the statistical properties of data. In this paper, NMF is modified to take the variations of data into account. This modified NMF (MNMF) is aimed at not only extracting the latent signals which are added together to produce the observed signals, but also capturing the main variations of data. Then MNMF can be used to extract the latent variables in a process and combine them with process monitoring techniques for fault detection. The proposed method was applied to the Tennessee Eastman Process (TEP) to evaluate its monitoring performance, and the experiment results demonstrated its feasibility and availability for process monitoring.