Batch process diagnosis: PLS with variable selection versus block-wise PCR

Data from a batch chemical process have been analysed in order to diagnose the causes of variability of a final quality parameter. The trajectories of 47 process variables from 37 batches have been arranged in a matrix by using alignment methods. Two different approaches are compared to diagnose the key process variables: PLS with variable selection and block-wise PCR. The application of Unfold Partial Least Squares Regression (U-PLS) leads to one significant component. By means of weight plots, the variables most correlated with the final quality are identified. Nevertheless, with observed data, it is not possible to know if correlation is due to causality (and hence related to a critical point) or is due to other causes. Pruning PLS models by using variable selection methods and technical information of the process has allowed the process variables most correlated with the final quality to be revealed. The application of Principal Component Regression to the trajectories of the process variables (block-wise PCR) has given straightforward results without requiring a deep knowledge of the process. The results obtained have been used to propose several hypotheses about the likely key process variables that require a better control, as a previous step to conducting further studies for process diagnosis and optimisation, like experimental designs.

[1]  Theodora Kourti,et al.  Process analysis, monitoring and diagnosis, using multivariate projection methods , 1995 .

[2]  J. Macgregor,et al.  Analysis of multiblock and hierarchical PCA and PLS models , 1998 .

[3]  Theodora Kourti,et al.  Analysis, monitoring and fault diagnosis of batch processes using multiblock and multiway PLS , 1995 .

[4]  L. E. Wangen,et al.  A multiblock partial least squares algorithm for investigating complex chemical systems , 1989 .

[5]  John F. MacGregor,et al.  Analysis and optimization of a polyurethane reaction injection molding (RIM) process using multivariate projection methods , 2003 .

[6]  L. Gleser.  Measurement, Regression, and Calibration , 1996 .

[7]  Age K. Smilde,et al.  PLS discriminant analysis with contribution plots to determine differences between parallel batch reactors in the process industry , 1999 .

[8]  H. Law.  Research methods for multimode data analysis , 1984 .

[9]  S. Wold,et al.  Multi‐way principal components‐and PLS‐analysis , 1987 .

[10]  J. Macgregor,et al.  Monitoring batch processes using multiway principal component analysis , 1994 .

[11]  Theodora Kourti,et al.  Comparing alternative approaches for multivariate statistical analysis of batch process data , 1999 .

[12]  Svante Wold,et al.  Modelling and diagnostics of batch processes and analogous kinetic experiments , 1998 .

[13]  S. Wold,et al.  Some recent developments in PLS modeling , 2001 .

[14]  Svante Wold,et al.  Hierarchical multiblock PLS and PC models for easier model interpretation and as an alternative to variable selection , 1996 .

[15]  John F. MacGregor,et al.  Multivariate SPC charts for monitoring batch processes , 1995 .

[16]  Thomas E Marlin,et al.  Process Control , 1995 .

[17]  Age K. Smilde,et al.  Improved monitoring of batch processes by incorporating external information , 2002 .

[18]  H. Kiers.  Towards a standardized notation and terminology in multiway analysis , 2000 .

[19]  Age K. Smilde,et al.  Three-way analyses problems and prospects , 1992 .

[20]  Jean-Pierre Gauchi,et al.  Comparison of selection methods of explanatory variables in PLS regression with application to manufacturing process data , 2001 .

[21]  Theodora Kourti,et al.  Multivariate dynamic data modeling for analysis and statistical process control of batch processes, start‐ups and grade transitions , 2003 .