Trilinear models for batch MSPC: Application to an industrial batch pharmaceutical process

Abstract

In this paper, PARAFAC and Tucker3 models were compared with the commonly used multiway principal components analysis approach (MPCA) for multivariate process control of an industrial batch antibiotic production process. Two different approaches for on-line monitoring were used: sliding window (multiple models) and global window (single model) monitoring strategies. The later approach requires orthogonality for the time dimension scores. In this context, a modification of the Parafac algorithm was proposed. The Tucker3 and Parafac models as proposed here share an identical structure. Scores (D) and residuals (Q) statistics were used to on-line identify faults. We concluded that Parafac and Tucker3 models outperformed MPCA in terms of detection of faults specially when the statistic for scores is used. All models performed equally well in the residuals statistics. The sliding window strategy proved to be more appropriate to identify faults than the global window strategy. This is, to our best knowledge, the first time such study was performed for an industrial batch antibiotic process.