Data integration for the monitoring of batch processes in the pharmeceutical industry

Abstract

Advances in sensor technology has resulted in large amounts of data being available electronically. However, to utilise the potential of the data, there is a need to transform the data into knowledge to realise an enhanced understanding of the process. This thesis investigates a number of multivariate statistical projection techniques for the monitoring of batch fermentation and pharmaceutical processes. In the first part of the thesis, the traditional performance monitoring tools based on the approaches of Nomikos and MacGregor (1994) and Wold et al. (1998) are introduced. Additionally, the application of data scaling as a data pre-treatment step for batch processes is examined and it is observed that it has a significant impact on monitoring performance. Based on the advantages and limitations of these techniques, an alternative methodology is proposed and applied to a simulated penicillin fermentation process. The approach is compared with existing techniques using two metrics, false alarm rate and out-ofcontrol average run length. A further manufacturing challenge facing the pharmaceutical industry is to understand the differences in the performance of a product which is manufactured at two or more sites. A retrospective multi-site monitoring model is developed utilising a pooled sample variancecovariance methodology of the two sites. The results of this approach are compared with a number of techniques that have been previously reported in the literature for the integration of data from two or more sources. The latter part of the thesis focuses on data integration using multi-block analysis. Several blocks of data can be analysed simultaneously to allow the inter- and intra- block relationships to be extracted. The methodology of multi-block Principal Component Analysis (MBPCA) is initially reviewed. To enhance the sensitivity of the algorithm, wavelet analysis is incorporated within the MBPCA framework. The fundamental advantage of wavelet analysis is its ability to process a signal at different scales so that both the global features and the localised details of a signal can be studied simultaneously. Both existing and the modified approach are applied to data generated from an experiment conducted in a batch mini-plant and that was monitored by both physical sensors and on-line UV-Visible spectrometer. The performance of the integrated approaches is benchmarked against the individual process and spectral monitoring models as well as examining their fault detection ability on two additional batches with pre-designed process deviations.