e Pergamon Computers and Chemical Engineering Supplement (1999) S827-S830 © 1999 Elsevier Science Ltd. All rights reserved PH: S0098·1354199/00089·7 Case Study Investigating Multivariate Statistical Techniques for Fermentation Supervision B. Lennox', G.A. Montague", H.G. Hiden+ and G. Komfeld* t Control Technology Centre, School of Engineering, University of Manchester, England + Department of Chemical and Process Engineering, University of Newcastle, England * Biochemie, Kundl, Austria Barry.Lennox@man.ac.uk Abstract This paper describes a case study in which multivariate statistical procedures have been developed to assist in the supervision of an industrial fed-batch fermentation process operated by Biochemie in Austria. The procedures have been developed to enhance the monitoring capabilities of the current system by interfacing directly into the present G2 real-time knowledge based supervisory system. While the G2 rule based system is useful for detecting deviations in single variables, it has been found to be unable to detect some of the more subtle deviations caused by the complex interactions between the process variables. Multivariate statistical techniques have been utilised in this study to provide early indications of deviations from nominal batch behaviour. The cause of these deviations can subsequently be determined by interrogating the information produced by these algorithms. Although the multivariate statistical techniques adopted in this paper are not new, their integration within the industrial supervisory system and the on-line application to the industrial fermentation process is novel. Keywords: Fault detection, Fault diagnosis, Fermentation, Multivariate statistical process control, Principal component analysis Introduction Traditional approaches for satisfying quality control in an industrial fermentation system rely on process operators to ensure that the fermentation progresses around a required trajectory. The operator uses his experience and knowledge of the fermentation process to detect potential problems and make modifications when necessary. The importance of effective operator control cannot be underestimated as the performance of a fermentation is very much dependant upon the ability to keep the system operating smoothly. A fermentation that is free from major upsets is likely to be more productive than one subject to significant disturbances. Therefore, the earlier a potential problem to the system can be detected, the less severe its influence on the system will be and the resulting corrective action will consequently be more restrained. Typically, the process operators develop a 'feel' for whether a fermentation is adhering to normality by comparing the current batch operation with what they perceive to be normal. This approach tends to make limited use of an important resource, namely the historical batch information that is routinely gathered and logged. A more rigorous approach to monitoring the fermentation process would be to compare individual variables with historical records and determine if any deviations exist. A rule-based structure could then be developed which brings any deviations to the attention of the process operators. A number of fermentation companies are now exploiting this type of analysis procedure using real- time knowledge based systems, such as G2 from Gensym. Unfortunately, this approach to process monitoring fails to detect deviations that may be the result of the combined effects of multiple variables. Recent studies have demonstrated that such combined deviations can be detected through the use of multivariate statistical procedures, such as Principal Component. Analysis (PCA) and Projection to Latent Structures (PLS). The remainder of this paper will demonstrate how these techniques have been applied to this fermentation process and an example of the ability of the procedures to correctly detect and classify a process disturbance will be provided. Multi-way Principal Component Analysis and Projection to Latent Structures Multivariate statistical process control (MSPC) has received considerable interest in recent years. Its name reflects an association with univariate statistical process control (SPC) and much of the technology finds some counterpart therein. Traditional SPC methods are based upon the charting of individual process variables and acting when necessary, to correct observed deviations from steady state (Wetherill and Brown, 1991). In certain systems this type of analysis may be all that is required, however, fermentation processes pose several problems that make univariate SPC inappropriate, such as: • batch operation means that steady state is not achieved.