Bioprocess statistical control: Identification stage based on hierarchical clustering

Abstract

Bioprocesses are characterized by the fact that small variations in operating conditions may have a substantial impact on the final batch quality. Therefore, the early detection and isolation of faults allow implementing corrective actions before the effects of deviations from the normal operation have a detrimental effect on production. In this work a new strategy for the statistical monitoring of batch processes is presented, and it is applied to monitor the operation of a fermentation process. The methodology works in the original variable space, therefore it only uses the Hotelling statistic for detection purposes. To determine the set of measurements by which the fault is revealed, the nearest in control neighbor to the observation point is calculated, and the distance between these two points is used to evaluate the contribution of each observation to the inflated statistic. In contrast to the existing latent-variable and original-variable based approaches, a simple hierarchical clustering technique allows to identify the set of suspicious measurements, without assuming the probability density function of the variable contributions. Furthermore, the performance of the proposed identification procedure is compared to the one achieved using other monitoring techniques. A well-known fed-batch fermentation benchmark is employed with this purpose, and the comparison is based on the results of a comprehensive set of simulated fault scenarios.