By combining mechanistic and empirical-based models, a process performance monitoring representation of a dynamic, non-linear process can be developed with the model-plant mismatch forming the basis of the monitoring scheme. In practice, the mechanistic model will not be perfect and therefore the residuals will contain structure. A modified model-based approach, Super Model-Based PCA (SMBPCA), is proposed which incorporates an additional residual modelling stage to remove structure from the residuals. The approach is evaluated on a simulation of a batch process using a number of residual modelling techniques including Partial Least Squares (PLS), dynamic PLS, ARX and dynamic Canonical Correlation Analysis (CCA). The out-of-control average run lengths for these techniques show that the SMBPCA approach gives improved process monitoring and fault detection compared to standard multivariate techniques.
展开▼
