Figure 6.51. MPLS-based end-of-batch monitoring results.
and 14 have the high contributions compared to other variables. These contributions are also calculated between 570th and 690th measurements where the out-of-control occurs. Variables 3, 6, 7, 13 and 14 are found to be significant in that case (Figure 6.51(e)). Since the original disturbance was introduced into the variable 3 (glucose feed rate) as a small downward drift, its effect on the other structurally important process variables such as dissolved oxygen concentration (variable 6) and biomass concentration (variable 7) becomes more apparent as the process progresses in the presence of that disturbance. Obviously, culture volume is expected to be directly affected by this disturbance as it is found with SPEx contribution plots. The weight contributions highlight the effect of this change on the process variables that are effective in the quality space. Variables 13 and 14 (heat generated and cooling water flow rate, respectively) being highly correlated with biomass concentration (variable7) also show high contribution. Since MPLS model can be used to predict end-of-batch quality as well, model predictions are compared with actual measurements in Figure 6.51(f). Quality variable 3 is predicted somewhat poorly. This is due to model order, and if the focus is on the prediction, this can be improved by increasing the number of latent variables retained in the MPLS model. End-of-batch quality can be predicted from the start of a new batch, this case is illustrated in Section 6.5.1.
Many chemical processes consist of several distinct processing units. Data from various processing 'stages' carried in processing units and 'phases' for operational or phenomenological regions in single units provide the information about the progress of the batch. As the number of units and phases increases, the complexity of the monitoring problem also increases. The techniques presented in Section 4.5.3 are useful for monitoring these type of processes with some modifications in both data pretreatment and monitoring techniques.
Example. A pharmaceutical granule production process by wet granulation technique following a fluidized-bed drying operation was chosen as a test case in this study. Process variables were broken up into blocks that correspond to specific processing stages. The choice of blocks depends on engineering judgment and the objectives of the study. In this study, blocks are related to particular processing units. Furthermore, because of the different operating regimes occurring in each unit, it is convenient to split the data from a stage into phases (Figure 6.52). This way, the predictive and diagnostic capabilities of the multivariate statistical models can be improved to provide more accurate inferences about the whole process. These
Was this article helpful?