Subspace monitoring based on full variable information

Since data collected from chemical processes always have high dimensions, modeling directly can be very complex. PCA can extract main features of the original data and obtain a more compact representation. However, the traditional PCA process monitoring scheme may cause information loss, since it only preserves principal components with large variance, which will greatly affect the performance of process monitoring. To handle this problem, a novel subspace monitoring method based on full variable information was proposed. Firstly, Based on the similarity level between each variable and principal component subspace (PCS) or residual subspace (RS), the original data space was divided into three low-dimensional subspaces, which preserved the whole process variables. Thus it could use the process information better. Secondly, the monitoring models were established respectively in each subspace, and then the Bayesian inference was introduced to integrate monitoring results of the subspaces. Finally, the feasibility and effectiveness of the FVI method were illustrated through a numerical example and the Tennessee Eastman process.