Abstract
In this paper, dynamic predictive monitoring schemes based on dynamic latent variable models are proposed. We consider the most typical case in industrial data where dynamics usually exist in a reduced dimensional subspace. First, using dynamic latent variable models, predictions are made to reduce the dynamic latent space variability and focus on the unpredictable variabilities for process monitoring, leading to reduced control regions without reducing confidence levels. A general expression is developed to decompose the overall uncertainty into dynamic prediction errors and static variabilities. Second, monitoring based on multistep ahead prediction windows are used to generate fault-free predictions whenever a fault is detected. Third, we illustrate that oblique projections in dynamic latent variable models are required to separate static variabilities from dynamic latent variables. A detected fault is further classified as a static fault or a dynamic fault based on different monitoring indices. Case studies on a simulation data set and the Tennessee Eastman data set demonstrate the effectiveness of the predictive monitoring schemes.
Original language | English |
---|---|
Pages (from-to) | 2353-2365 |
Number of pages | 13 |
Journal | Industrial and Engineering Chemistry Research |
Volume | 59 |
Issue number | 6 |
Early online date | 10 Jan 2020 |
DOIs | |
Publication status | Published - 12 Feb 2020 |
Externally published | Yes |
Funding
This work was supported in part by the Shenzhen Committee on Science and Innovations (20160207, 20170155) and the Postdoctoral Fellowship Fund of the Chinese University of Hong Kong, Shenzhen.