Nonlinear and Non-Gaussian Process Monitoring Based on Simplified R-Vine Copula

Abstract

In the field of chemical process monitoring, the vine copula model provides a new idea for describing the interdependence between high-dimensional complex variables, and directly characterizes the correlation without dimensional reduction. However, in actual industrial processes, the number of pair copulas to be optimized and the parameters to be estimated increase rapidly when the dimensionality of the variables is large. This greatly increases the computational load and reduces the detection efficiency. In this paper, a fault diagnosis method based on a simplified R-vine (SRV) model is proposed. Without reducing the precision of the model significantly, the simplified level is set to reduce the complexity of the workload and calculations. The simplified level of an R-vine model is obtained by a Vuong test. Then, the generalized local probability (GLP) of the non-Gaussian state is constructed by using the theory of highest density region (HDR) and a density quantile table. The monitoring results of the Tennessee Eastman (TE) process and a real acetic acid dehydration distillation system show that the proposed SRV approach achieves good performance in monitoring results and computational load for chemical process fault monitoring.