A process
monitoring method based on double-model and multi-subspace
vine copula (DMVC) is proposed in this paper. To improve the fault
detection performance, process variables are divided into two sets
according to their correlation, in which C-vine and D-vine are used
to build models, respectively. The variables with stronger correlation
are selected to build the C-vine model, and those with weak correlation
are used for the D-vine model. In addition, the two models individually
establish three different subspaces according to different training
data information, including marginal distribution subspace (MDS),
dependent structural subspace (DSS), and joint distribution subspace
(JDS). Highest density region (HDR) and generalized local probability
(GLP) are also introduced to establish a robust control domain and
detection index, which makes the method more sensitive to data. The
effectiveness of the proposed method in the field of process monitoring
is verified by numerical simulation and the Tennessee Eastman (TE)
process. Compared with classical multivariate statistical methods,
the DMVC displays superior monitoring performance.