Monitoring fermentation processes using in‐process measurements of different orders

BACKGROUND: In‐process monitoring of fermentation processes (at‐line, on‐line or in‐line measurements) is essential to control productivity and ensure high product quality. A number of different monitoring techniques are available for this purpose and one possible categorization among this variety of techniques is based on the different structures generated by the measurements and the potential of these with respect to holding and extracting process information. In this study lactic fermentation processes is monitored by different techniques (brix, pH, NIR‐ and fluorescence‐spectroscopy) providing different data structures (zero‐, first‐ and second‐order). Multivariate data analysis (PCA and PARAFAC) was applied on the first‐ and second‐order data sets, and the different measurement signals or derivatives of these were combined by a multiblock strategy. The aim of this work is to present and clarify the advantages and variations of the different data structures. RESULTS: The zero‐order pH and brix measurements (a commonly used measure for total sugar content in wine fermentations) decreased in a smooth and logical pattern from 6.4 to 4.4 and from 10.5% to 6.2%, respectively – provided valuable critical quality attributes, communicating the fermentation process is progressing over time in accordance with biological and engineering intuition. The first‐order NIR measurements modelled with PCA showed an increasing trend over time on PC1. This increasing trend corresponds to the lactic bacterial growth. This trend could be distinguished by statistical modelling from a second trend (PC2), reproducible for all production batches. Based on the second‐order fluorescence measurements modelled by PARAFAC and its statistical uniqueness properties, three distinctive fluorescence compounds were found to vary over process time. Most probably these three compounds represent riboflavin, tryptophan and lumichrome or NADH. Using multiblock PCA the combined sensor signals identified two distinguished, reproducible time profiles for all batch runs. CONCLUSIONS: The most interpretable chemical information was obtained by fluorescence spectroscopy due to the uniqueness properties of second‐order measurements. The first‐order technique NIR spectroscopy also provided valuable process information, though the process trends can only be interpreted indirectly and if interfering species had been encountered they could not have been modelled. The multiblock data set provided by zero‐, first‐ and second‐order measurements recorded over time highlighted important relationships among the different variables that provide chemical information when multivariate data analysis is applied. Although, first‐ and second‐order measurements seem to obtain more information than the zero‐order measurements, it is important to keep in mind that zero‐order measurements can provide valuable information about the process, especially in combination with different sensors. © 2014 Society of Chemical Industry