Spectroscopic and process data fusion : enhanced monitoring of an industrial fermentation
Large scale manufacturing of pharmaceutical products is a highly competitive industry in which technological improvements can maintain fine business margins in the face of competition from those with lower manufacturing overheads. Processes in which pharmaceuticals are produced via fermentation are particularly susceptible to large variability and reduced productivity due to natural variation and limited monitoring and control options. The latest monitoring methods offer the potential to understand causes of variation, improve productivity and as a result maintain the competitive edge. Unfortunately the fermentation environment is not conducive to the implementation of instrumentation. This thesis shows how signals from spectral instruments can be enhanced by other process and spectroscopic measurements, to provide on-line measurements of critical broth concentrations traditionally only available from infrequent off-line analysis. Near infrared (NIR) and Mid infra red (MIR) spectral analysis of fermentation broth can provide measurements of key concentrations throughout a batch. The off-line analysis of broth samples is more straightforward but on-line implementation is possible. In the case of on-line implementation, the quality of information is compromised, placing greater demands on the signal interpretation methods. The objective of the thesis was to understand the causes of process variation and compensate for them during batch progression, consequently on-line implementation was essential. The construction of a robust calibration model for individual instruments is the first step in implementation. The traditional strategy is either to use multivariate techniques such as projection to latent structures (PLS) or wavelength selection through genetic algorithms followed by PLS. An alternative approach is developed where a search strategy identifies a limited number of spectral windows (SWS) that are most descriptive of the concentrations of interest. The benefit of using SWS is that problems associated with over-fitting the calibration model construction data are minimised. This is particularly important in a development environment where the number of batches is limited. The random nature of the search strategy of the SWS algorithm results in a range of calibration models. Multiple calibration models are `stacked' to provide improved accuracy and robustness. It is demonstrated that stacking provides an improved prediction capability compared to selecting the single `best' performing model. Finally, developing calibration models for sub-regions of fermentation operation is contrasted with a global model. The improvement in accuracy of measurements from SWS and stacking is significant but errors in the determination of the concentration of some compounds remained significant. To overcome these offsets, a model relating the calibration residuals to on-line process measurements was constructed using PLS. The model was then used to correct the spectral calibration prediction to result in improved determination of broth concentrations. The fermentation monitoring methodology is demonstrated by application to an industrial antibiotic production process. Corrected predictions of product concentration and broth nutrient levels demonstrate that combining multiple information sources is advantageous from a measurement perspective.