Search

The analysis of batch process data can provide insight into the process operation, and there is a vast amount of historical data available for data mining. Empirical modelling utilising this data is desirable where there is a lack of understanding regarding the underlying process (Formenti et al. 2014). This may be the case for fed-batch fermentation processes, where mechanistic modelling is challenging due to non-linear dynamics, and non-steady state operation. There is also a lack of sensors for key parameters which are considered to define the quality of the batch, such as product concentration (Nomikos and MacGregor 1995). Multivariate analysis is a powerful tool for investigating large data sets by identification of trends in the data. However, there are also challenges associated with the application of multivariate analysis tools to batch process data. This is due to issues related to the different batch lengths, different data sampling intervals, noise in the measurements, and both online and offline data. The importance of the pre-processing stages are often underappreciated (Gurden et al. 2001). In this work, a 30 batch dataset from a production process operating at Novozymes A/S is analysed by multivariate analysis with the aim of predicting the final product concentration, which is measured offline at the end of each batch. Many modelling iterations were required using different pre-processing methods, in order to extract the trends from the data set. The final model gave an average prediction error of 7.6%. The success of the final regression model was heavily dependent on the decisions made in the pre-processing stages, where the issues of different batch lengths, different measurement intervals, and variable scaling are considered. Therefore a methodology is presented for future application of multivariate methods to industrial scale process data to cover these considerations.

Batch production processes pose specific challenges for process monitoring and control. This isdue tomany reasons includingnon-linear behaviour, and arelativelypoor understanding of thesystem dynamics[1].It is therefore challenging for theprocess engineer to optimise the operationconditions, due to a lack of available process models, and complex interactions between variableswhich are not easy to define, especiallyacross scales and equipment.There is however a vastamount of batch process datagenerated, which canbe investigated with the aim of identifyingdesirable process operating conditions, and thereforeareas offocus for optimising the processoperation.This requires multivariate methods which canutilise the complexdatasetswhich areroutinely collected, containing online measured variables and offline sample data.Fermentation processes are highly sensitive to operational changes, as well as between batchvariations, and are therefore an interesting application of multivariate methods. The processdynamics are governed by the combination of process variables, and cannot be fully characterisedby individual variables alone[2]. There is also a lack of sensors for key variables which areconsidered todefine the operation[3], which makestraditional modelling a challenge.Although multivariate techniques are routinely used for chemometric applications, theirapplication to batch processes islesscommon due to the additional challenges associated withuneven batch lengths and less reproducible data, which has naturally greater variability, as well ashigh measurement noise. This requires additional preprocessing stagesin order to extract theinformation within such a dataset.A 30 batch dataset from a production process operating at Novozymes A/Sis analysed bymultivariate analysis with the aim of predicting the final product concentration, which is measuredoffline at the end of each batch. By creating a model for product concentration, it is possible toanalysethe model results and interpret this to guideprocess optimisation effortstowardsachieving a greaterproduct concentration.

Multivariate analysis allows process understanding to be gained from the vast and complex datasets recorded from fermentation processes, however the application of such techniques to this field can be limited by the data pre-processing requirements and data handling. In this work many iterations of multivariate modelling were carried out using different data pre-processing and scaling methods in order to extract the trends from the industrial data set, obtained from a production process operating in Novozymes A/S. This data set poses challenges for data analysis, combining both online and offline variables, different data sampling intervals, and noise in the measurements, as well as different batch lengths. By applying unfold principal component regression (UPCR) and unfold partial least squares (UPLS) regression algorithms, the product concentration could be predicted for 30 production batches, with an average prediction error of 7.6%. A methodology is proposed for applying multivariate analysis to industrial scale batch process data.