(233d) Multivariate Statistical Process Control for the Continuous Manufacturing of Solid Oral Dosage Forms

Nowadays a commodity in the industrial software industry, the application of principal components analysis (PCA) to monitor the state of a process with Multivariate Statistical Process Control (MSPC) charts is far from a venture. Its implementation in a commercial plant with a contemporary control and/or automation system requires a modest effort, given a stable process and the availability of a historian data base with abundant data representative of the normal operating conditions (NOC) for the plant. This work presents the development of an MSPC monitoring scheme for a continuous drug product manufacturing process, using information available at the R&D stage. At the time when the process is still being developed, the data representative of the expected NOC may not be abundant. Additionally, the product development efforts may have other experimental priorities, for example running experimental trials at different mass rates and/or configuration of the feeders, mixer, and/or tablet press equipment.

 

The approach presented is based on the assumption that, regardless of the target set point for the main operating conditions (e.g. mass rate), the common cause variability in the process will remain the same. Hence the covariance structure for the variables in the system can be identified by a PCA model on a matrix resulting from the concatenation of data from multiple â??trial runsâ? that were pre-centered around their mean values. An analysis of the loadings vectors reveals variable correlations that matches the expectations from a process engineering standpoint; and the corresponding movement in the scores space reveals directions of expected normal (and tolerated) variability, as well as patterns that can be matched with the startup and the shutdown of the system.

 

Our case study presents the analysis of a model built with the data collected from the system of screw feeders that dispense material to the unit, as well as the analysis of the model obtained with data collected from the tablet press. Furthermore, a method is presented to determine the lag time between a measurement vector from the feeders, and a measurement vector from the tablet press; such that a combined lagged model can be built. The combined model eases the holistic identification of contributions in the case of a disturbance, and the use of the three models (one per system and the combined) in tandem provides a complete monitoring approach for the system as a whole.