Beer fermentation: Monitoring of process parameters by FT-NIR and

multivariate data analysis

Silvia Grassi
a
, Jos Manuel Amigo
b,
, Christian Bge Lyndgaard
b
, Roberto Foschino
a
, Ernestina Casiraghi
a
a
Department of Food, Environmental and Nutritional Sciences (DeFENS), Universit degli Studi di Milano, via Celoria 2, 20133 Milano, Italy
b
Department of Food Science, Faculty of Sciences, University of Copenhagen, Rolighedsvej 30, DK-1958 Frederiksberg C, Denmark
a r t i c l e i n f o
Article history:
Received 6 November 2013
Received in revised form 15 January 2014
Accepted 18 January 2014
Available online 27 January 2014
Keywords:
Beer fermentation
Quality control
NIR
FT-NIR
PCA
PLS
LWR
a b s t r a c t
This work investigates the capability of Fourier-Transform near infrared (FT-NIR) spectroscopy to moni-
tor and assess process parameters in beer fermentation at different operative conditions. For this purpose,
the fermentation of wort with two different yeast strains and at different temperatures was monitored
for nine days by FT-NIR. To correlate the collected spectra with Brix, pH and biomass, different multivar-
iate data methodologies were applied. Principal component analysis (PCA), partial least squares (PLS) and
locally weighted regression (LWR) were used to assess the relationship between FT-NIR spectra and the
abovementioned process parameters that dene the beer fermentation. The accuracy and robustness of
the obtained results clearly show the suitability of FT-NIR spectroscopy, combined with multivariate data
analysis, to be used as a quality control tool in the beer fermentation process. FT-NIR spectroscopy, when
combined with LWR, demonstrates to be a perfectly suitable quantitative method to be implemented in
the production of beer.
2014 Elsevier Ltd. All rights reserved.
1. Introduction
Beer is, arguably, the most widespread alcoholic beverage in the
world and its success is related, among other factors, to its suitabil-
ity for large scale production. Beer production was originally
handcrafted and the control of the fermentation was exerted by
skilled brew masters able to evaluate the progress of the process
with empirical sensory evaluations (Boulton & Quain, 2006, chap.
6). In the past decades, thanks to the rise of industrial beer produc-
tion, the subjective evaluation of the process has been outdated by
initial control parameters, being the regulation of wort composi-
tion and yeast strain the main ones (Bamforth, 2006, chap. 11). De-
spite the precise assessment of the previous mentioned parameters
the brewing industry still has to face different variation in the cycle
of the fermentation that might hamper the quality of the nal
product. Therefore, large scale production of beer leads to the
necessity of quality assurance systems to guarantee the fullment
of the required quality consistency of the nal product, forcing the
breweries to dene and control one or several parameters during
the process to meet the proper standards. Although each company
denes its own specications, some are required by law. In partic-
ular, for what regards the nal product, alcohol content or Plato
degree per hectolitre are the legal parameters upon which beer is
taxed (Directive 92/84/EEC, 1992). For this motive, an unavoidable
parameter to be controlled during the rst fermentation of beer is
specic gravity (SG), correlated with sugar concentration in the
wort and, therefore, alcohol content in the nal product. The indus-
trial trend is to monitor the fullment of SG to the dened speci-
cation through the use of Quality Control Charts, which allow a fast
visual check of the trend of the system being controlled through
mean and control limits statistically dened. However, with this
univariate quality assurance strategy is difcult to date back to
the source of the no standard behaviour leading to the abnormality
in the batch (Kourti, 2005). The main problem is that a failure in a
parameter evaluated might be originated from several correlated
variables, often non-measured, governing such a complex bioproc-
ess as it is beer rst fermentation (Kourti, 2006).
Recently, companies started to look for methods providing com-
prehensive information of the on-going process in order to assure
an effective control at all stages. The implementation of Fourier-
Transform near-infrared (FT-NIR) probes has been proved to offer
detailed information in real-time in several food processes, allow-
ing the assurance of meeting the quality parameters to the dened
specications (Huang, Yu, Xu, & Ying, 2008). Despite being a prom-
ising technique, FT-NIR has several drawbacks to be considered as
well. For instance, the spectral signal generally obtained in the fer-
mentation of beer is strongly characterised by the overlapping of
http://dx.doi.org/10.1016/j.foodchem.2014.01.060
0308-8146/ 2014 Elsevier Ltd. All rights reserved.

Corresponding author. Tel.: +45 353 32570.

E-mail address: [email protected] (J.M. Amigo).
Food Chemistry 155 (2014) 279286
Contents lists available at ScienceDirect
Food Chemistry
j our nal homepage: www. el sevi er . com/ l ocat e/ f oodchem
the bands due to the heterogeneous and complex composition of
the medium. In fact, the wort is mainly composed by fermentable
sugars (i.e. maltose, maltotriose, glucose, sucrose, fructose), non-
fermentable sugars (i.e. dextrins) and nitrogen sources in the form
of proteins, peptides and aminoacids (Lewis & Young, 2002, chap.
2). All the mentioned compounds absorb in the NIR region
(120004000 cm
1
), as they contain CAH, NAH, OAH and C@O
bonds, giving an overlapped signal. Moreover, water is the major
constituent of the wort (ca. 92%) and it strongly characterises
NIR spectral information with its peaks around 6900 cm
1
(OAH
rst overtone) and around 5300 cm
1
(combination of the asym-
metric stretch and bending of the water molecule) (Workman &
Weyer, 2008, chap. 6).
Chemometrics has been applied in the last 30 years to over-
come the mentioned drawbacks and to extract relevant informa-
tion from FT-NIR data (Bock & Connelly, 2008). Principal
component analysis (PCA) and Partial Least Square (PLS) regression
became of interest, as they are able to take advantage of the struc-
tures in highly overlapping and co-linear data (Martens & Ns,
1989). Both methods are linear. Nevertheless, in some cases, the
correlation between the spectral signal and the reference measure
has a non-linear nature. In these cases, methods for linear model-
ling of nonlinear surfaces are needed. Cleveland and Devlin (1988)
proposed locally weighted regression (LWR) as methodology to
full this goal and Ns and Isaksson (1992) and Naes, Isaksson,
and Kowalski (1990) applied the strategy to infrared spectroscopy.
To what brewing monitoring concerns, FT-NIR spectroscopy
combined with Chemometrics has been applied to simultaneous
determination of compounds in brewing nal product (Duarte,
Barros, Almeida, Spraul, & Gil, 2004; Ghasemi-Varnamkhastia
Rodriguez-Mendeza, Gomes, Ugulino Arajo & Galvo, 2012); Inon,
Llario, Garrigues, & de la Guardia, 2005; Lachenmeier, 2007 and for
beer authentication (Di Egidio, Olivieri, Woodcock, & Downey,
2011; Engel, Blanchet, Buydens, & Downey, 2012). Nevertheless,
little research exists on the use of FT-NIR for monitoring and
assessment of changes in relevant physico-chemical parameters
in the beer fermentation process (McLeod et al. 2009).
Therefore, much more intensive work is needed to really assess
the usefulness and power of FT-NIR and Chemometrics in model-
ling, controlling the beer fermentation process. This work investi-
gates the capability of FT-NIR spectroscopy to monitor wort
fermentation conducted at different operative conditions (temper-
ature and yeast type), deepening into the robust correlation
between NIR spectra collected during fermentation and different
key factors in beer manufacturing. For this purpose, PCA was
applied to the FT-NIR spectra of a set of six different beer fermen-
tation conditions, to ascertain whether beer spectral changes may
be correlated with the biotransformation progress, both biomass
increase and chemical modications. Furthermore, PLS and LWR
models were calculated to correlate sugar content (Brix), pH and
biomass to the spectral information in order to develop a rapid
method to detect the changes of these parameters in real-time dur-
ing the fermentation process.
2. Materials and methods
2.1. Materials
Standard commercial wort Highland heavy Ale wort was pur-
chased (Muntons plc, Suffolk, UK) and reconstituted according to
instructions of the manufacturers in 2 L of heat treated water for
each trial, agitated at 200 rpm for 2 h to saturated level. Two differ-
ent Saccharomyces cerevisiae strains, one for British Ale (WLP005)
and the other for Belgian Ale (WLP570) (WhiteLab Inc., San Diego,
CA, USA), were separately inoculated into the wort, reaching
approximately the concentration of 7 10
6
CFU mL
1
. The prepa-
ration of the inoculum was performed as follow: one fresh colony
for each strain was aerobically propagated in 10 mL YEPD broth
(Yeast Extract 10 g L
1
, Peptone 20 g L
1
, Dextrose 20 g L
1
, pH
6.4) at 20 C for 24 h; afterwards, the suspension was transferred
in 100 mL of fresh broth with a magnetic stirrer (200 rpm) for
24 h at 20 C, then transferred in 400 mL YEPD and incubated for
24 h at 20 C. The harvested cells were pitched in the wort and
gently agitated to create a homogeneous distribution before shar-
ing out about 300 mL into sterile glass bottles, one for each sam-
pling point, closed with airlock caps. The cell concentration for
each strain was determined using a calibration curve obtained by
correlating plate counts on YEPD agar and optical density values
measured at 620 nm.
2.2. Fermentation trials
The experiments were performed according to a factorial design
with the two different S. cerevisiae strains (WLP005 and WLP570)
and three different fermentation temperatures (19, 21 and 24 C),
all replicated twice, giving a total amount of 6 different experi-
ments for each beer type.
Samples were collected in triplicate right after pitching (0 h,
starting time) and then every 22 h until the 9th day of fermenta-
tion, using two different sampling methods: directly from superna-
tant and after centrifugation for 15 min at 3000g.
2.3. pH,Brix and biomass determination
pH was measured at each sampling point by a Portamess

911
pH-meter (Knick Elektronische Messgerte GmbH & Co. KG, Berlin,
Germany).
Total sugar content was measured by a refractometer
(Bellingham and Stanley RMF 340, Kent, UK).
The changes in biomass were determined by optical density
measured at 620 nm with a UVvisible spectrometer (Hemlett
Packard 8453, Agilet Technologies, Waldbronn, Germany).
2.4. FT-NIR spectroscopy
FT-NIR spectra were collected in transmission mode, with a
Bomen QFA Flex FT-NIR spectrometer (Q-Interline A/S, Roskilde,
Denmark) equipped with a 1 mm path length cuvette. The data
were collected in the range 120004000 cm
1
, with a resolution
of 16 cm
1
and 128 scans for both background and samples.
2.5. Data processing
The whole wavenumber range (120004000 cm
1
) was reduced
in order to eliminate useless or saturated variables from spectra.
For PCA and regressions of chemical indexes the nal two spectral
ranges selected were from 7500 to 5500 cm
1
and from 4700 to
4350 cm
1
. They are characterised by the principal absorption
bands of the compounds involved in the biotransformation. Only
for the regression models for biomass prediction the spectra range
used was 10,500 to 5500 cm
1
and from 4700 to 4350 cm
1
, in
order to keep the information about baseline changes which are
correlated to the yeast growth in the media.
Except for biomass regression, the spectra were pre-processed
to minimise the effect of baseline shifts and noise and to highlight
modications due to the chemical composition. Thus, Standard
Normal Variate (SNV), rst derivative Savitzky Golay (7 wave-
lengths gap size and 2nd order polynomial) was performed. Before
data analysis with PCA, the nal pre-processed spectra were
mean centred. In the same way, before regression models, the
280 S. Grassi et al. / Food Chemistry 155 (2014) 279286
pre-processed spectra were mean centred and the variable to pre-
dict auto-scaled.
All models were developed by the PLS_Toolbox (Eigenvector
Research Inc., Wenatchee, WA) working under MatLab v. 7.4 (The
MathWorks, Natick, MA).
PCA and PLS models are quite common in literature, and further
information can be found elsewhere (Kramer, 1998; Kramer,
Workman, & Reeves, 2004).
Briey, PCA is a suitable tool for simplication, data reduction,
outlier detection and pattern recognition (Jolliffe, 1986), highlight-
ing the variance of the spectra. PLS, on the contrary, is a standard
method to solve multivariate regression problems, maximising
the correlation between the NIR spectra and the parameters to
quantify (Naes, Isaksson, Fearn, & Davies, 2002).
LWR, though, is not frequently presented as regression tech-
niques applied to spectroscopy.
Therefore, the main steps characterising the analysis are
reported (Bevilacqua, Bucci, Materazzi, & Marini, 2013) as they
were applied in this work:
1. Denition of the number of nearest neighbours (local points)
close to the prediction spectrum to be used to build local cali-
bration models.
2. Find for every new object the samples closest to it in the local
calibration model (nearest neighbours).
3. Build a local calibration model using the nearest neighbours
only; assign the weights (a) of the neighbours in the local model
according to their Euclidean distance to determine the close-
ness to the dependent variable and the auto-scaled distance
in the principal components space.
4. Prediction (by PLS regression) of the new sample by applying
the local calibration model developed.
3. Results and discussion
3.1. pH and Brix degree
pHvalues obtained at the inoculation time (t = 0) were very sim-
ilar for all the samples, having an average value of 5.64 0.07. This
conrms the reproducibility of the wort reconstitution protocol
(Fig. 1a and b). Just after the fermentation started, a fast drop in
pH values was observed for all the trials performed within 22 h of
fermentation. This decay in the pH, though, was different (as ob-
served in Fig. 1a and b) depending of the yeast strain inoculated.
The trials performed with WLP005 strain showed a minimum pH
value at 24 C after 72 h (4.45 0.02) (Fig. 1a), followed by a modest
rise up to 4.52 0.03 at the end of the monitoring (t = 216 h). This
behaviour of the pH after the fermentation mid-point was previ-
ously reported by Berner and Arneborg (2001).
Fig. 1. pH and Brix proles for the three tested temperatures: blue squares for trails performed at 19 C, green circles for the one at 21 C and red triangles for the one at
24 C. pH evolution of trials conducted with WLP005 strain (a), and WLP570 (b). Soluble solid content of trials conducted with WLP005 strain (c), and WLP570 (d). (For
interpretation of the references to colour in this gure legend, the reader is referred to the web version of this article.)
Table 1
Two way analysis of variance for Brix and pH value at the end-points to assess the inuence of the three temperatures and the two yeast strains tested (a = 0.05).
Source of variation SS DF MS Effect of factor (n.s.)
Brix Temperature 0.016 2 0.008 0.043
Yeast strain 0.112 1 0.112 0.60
pH Temperature 0.008 2 0.004 0.101
Yeast strain 0.015 1 0.015 0.366
SS = sum of square; DF = degree of freedom; MS = mean square.
S. Grassi et al. / Food Chemistry 155 (2014) 279286 281
Concerning the trials inoculated with WLP570 strain, the pH
decay continued until the last sampling point (t = 216 h) (Fig. 1b).
More precisely, it is possible to observe the temperature effect
for the trials incubated at 19 C (the slope of the pH curve is lower
than for the trials performed at higher temperatures in Fig. 1a and
b). Nevertheless, the nal pH values were comparable at all the
temperature tested (4.41 0.13).
Fig. 1(c and d) shows as well the Brix changes for all the trials
performed. As observed for pH, the Brix at starting time were the
same for all the trials (9.37 0.07 Bx) corresponding to a sugar
content of approx. 97.50 g/L. The total sugar content showed a fast
decrease in all the trials performed within three days of fermenta-
tion. The sugar consumption was faster when WLP005 strain was
used, reaching 6.55 0.05 Bx after 46 h at 24 C (Fig. 1c) and
remaining constant after 142 h of fermentation, due to the charac-
teristic high occulation of the yeast cells and the consequent low
sugar consumption. The main difference was observed when wort
inoculated with WLP005 was incubated at 19 C; in this case the
sugar content decreased with a lower speed but reaching at the fer-
mentation end-point the same level of the other trials performed
(5.19 0.04 Bx). The sugar depletion in the fermentations con-
ducted with WLP570 strain showed a lower and more constant
slope than those inoculated with WLP005 (Fig. 1d). This is probably
due to the ability of WLP570 yeast cells to remain in suspension in
the wort during the fermentation time.
To assess if the end points of the trials performed were signi-
cantly different according to the temperature and yeast used the
mean of Brix and pH values at the nal points were subjected to
a two way analysis of variance (ANOVA). Table 1 reports the results
obtained. At 95% of condence level there were found no signi-
cant differences in the end points.
As described beforehand, the trends of the pH and Brix could
be different according to the yeast used, but evaluating their values
at end point is not possible to nd out differences due to the strain
inoculated or the operating temperature. In this regard pH and
Brix measurement alone could not be an exhaustive tool to mon-
itor the overall fermentation performance as is not possible to nd
out differences in the processes due to failure of temperature con-
trol or due to the yeast strain governing the fermentation.
3.2. FT-NIR spectroscopy
Fig. 2a shows an example of the spectra collected directly
from the supernatant for the fermentation conducted at 19 C
Fig. 2. FT-NIR spectra collected directly from the supernatant for the fermentation conducted at 19 C with WLP570. Raw spectra in the region 75005500 and
47004350 cm
1
(a), detail of the region 47004350 cm
1
(b). Spectra transformed with SNV and rst derivative (7 pt, 2nd polynomial order) (c), detail of the region
47004350 cm
1
after pre-treatments (d).
282 S. Grassi et al. / Food Chemistry 155 (2014) 279286
with WLP570. Here it is possible to observe an increase in
absorption and baseline drift due to the scattering effect caused
by the yeast growth in the media. The main peak around
6900 cm
1
is related to OAH rst overtone of water (Workman
& Weyer, 2008, chap. 6), which is the main compound present
in the fermenting wort (around 90%). Other small features are
present in the 47004350 cm
1
region, being related to the sug-
ars consumption and ethanol production (McLeod et al., 2009).
Fig. 2b shows in detail the second spectral area, being able to
observe the characteristic peak of ethanol around 4420 cm
1
and its evolution with time.
The pre-processed spectra obtained at 19 C with WLP570 are
shown in Fig. 2c. Main changes are highlighted in the region
75005500 cm
1
. In particular changes are associated with water
OAH absorption (70006500) and with CAH methyl associated
with OAH as RAOHCH3 (5925 cm
1
) (Workman & Weyer, 2008,
chap. 6). Moreover in the region 47004350 cm
1
ethanol and sug-
ars modications (CAH combination bands, and OAH stretch over-
tone) are evident.
The increase of ethanol content according to fermentation time
is well highlighted in Fig. 2d. The highlighted wavelength range
(47004350 cm
1
) recorded at the beginning of the fermentation
(light gray colour) does not have the characteristic ethanol
peaks due to the OAH stretching combined with OAH bending
(Workman & Weyer, 2008, chap. 6); whereas it is possible to ob-
serve the appearance of the peaks and their increment with time.
3.3. Principal component analysis
Fig. 3a shows the score plot of the rst and the second principal
components, PC1 and PC2, respectively, with samples coloured
according to the fermentation time. PC1 explains 63.21% of the to-
tal variance and clearly describes the evolution of the spectra
according to the fermentation time. It is noteworthy that this clear
denition of the evolution of fermentation by PC1 is repeated for
all the fermentations, no matter the sampling method used, the
yeast or the temperature tested, denoting the precision and robust-
ness of FT-NIR spectra in reporting fermentation evolution under
different conditions. Samples collected at t = 0 (red triangles) have
positive PC1 values and are grouped together, in agreement with
the results obtained by pH and Brix measurements. With the pro-
gress of the fermentation time samples have lower PC1 scores,
with a good separation between each sampling time.
The main peaks responsible of samples distribution along the
PC1 (Fig. 3d) are due to the inuence of ethanol (peaks around
60005500 cm
1
and 4400 cm
1
), sugars (rst combination region
CAH around 4500 cm
1
) and the consequently change in water
(OAH rst overtone at 6900 cm
1
).
On the other hand, PC2 clearly separates the samples according
to the sampling method used (Fig. 3b). All the samples collected
after wort centrifugation have negative PC2 values; while most
of the samples collected directly from the supernatant have posi-
tive values, which generally increase with the time progress.
Fig. 3. Principal component analysis of all the FT-NIR spectra collected. Score plots: (a) PC1 vs. PC2 with samples coloured according to fermentation time (red 4= 0 h, orange
s= 22 h, purple h= 46 h, blue = 70 h, light blue } = 142 h, green = 216 h), (b) samples coloured according the sampling method used (blue h= centrifuged and green
= not centrifuged), (c) samples coloured according to the strain inoculated (h for WLP005 and for WLP570). Loadings plot: (d) represents the PC1 loading (gray line) and
PC2 loading (black-dashed line). (For interpretation of the references to colour in this gure legend, the reader is referred to the web version of this article.)
S. Grassi et al. / Food Chemistry 155 (2014) 279286 283
This outcome can be associated with the physical effect of
increasing of the concentration of suspended cells in the medium
(Beauvoit, Liu,Kang, Kaplan, Miwa & Chance, 1993). However, there
are some of the samples collected from the supernatant which are
characterised by low PC2 values. Looking at Fig. 3c is possible to
nd an explanation about the supernatant samples having low
PC2 values with the progress of time. The samples collected
directly from the supernatant of the wort fermented with
WLP005 (h) have PC2 values increasing until t = 70 h, thereafter
the values decrease up to reaching the same value of the samples
collected just after the starting time. This phenomenon can be as-
cribed to the occulating nature of this strain. Indeed, the loading
prole of PC2 (Fig. 3d, ()) is highly inuenced by scattering effect
in the regions 65005500 and 50004500 cm
1
which was not
completely corrected by SNV pre-treatment.
To better understand the different behaviours observed accord-
ing to the yeast strains inoculated, PCA models on the single strain
were performed. Fig. 4a reports the scores plot of the PC1 vs. PC3
performed by selecting from the dataset only the experiments per-
formed with WLP005.
As observed in the global PCA (Fig. 3) this strain is characterised
by occulation behaviour, which causes the precipitation of the
cells on the bottom of the ask after 70 h. In this case the scatter
effect and changes in the water peak (6900 cm
1
) is described
mainly by the PC3. High values of PC3 correspond to a high number
of yeast cells in the media, i.e. high scattering effect. Moreover the
third component is inuenced by changes in the OAH rst over-
tone (6900 cm
1
), as observed in the corresponding loadings plot
(Fig. 4c).
Fig. 4b shows the scores plot referred to the PCA performed by
selecting just the WLP570 strain. In this case the scattering effect is
described by the PC2 and two groups are well dened according to
the sampling method used (Fig. 4b). All the spectra collected from
centrifuged samples have negative PC2 values whereas samples
collected directly from the supernatant are characterised by posi-
tive PC2 values increasing during fermentation time due to the
higher concentration of yeast in the media causing scattering
effect. In both individual PCA (Fig. 4a and b) the PC1 described
the time progress related to the main modication occurring dur-
ing the biotransformation progress no matter the sampling method
used. In particular the regions effecting the samples distribution
are related to sugars (45004300 cm
1
) and ethanol (60005500
and 4200 cm
1
) (Fig. 4C and d).
3.4. Quantitative determination of solid content, pH and biomass
PLS analysis was employed as initial method to estimate Brix,
pH and biomass for all the spectra collected. PLS calibration models
were constructed using the whole dataset comprising 6 experi-
ments run at different temperatures (19, 21, 24 C) and with differ-
ent inoculated strains (WLP005 and WLP570), replicated twice and
collected with two different sampling methods (centrifuge and
Fig. 4. Principal component analysis of the FT-NIR spectra collected from the trials performed with WLP005 strain (a and c) and with WLP570 (b and d). (a) PC1 vs. PC3 of
WLP005 with samples coloured according the sampling method used (blue h= centrifuged or green = not centrifuged). (b) PC1 vs. PC2 of WLP570 with samples coloured
according the sampling method used (blue h= centrifuged or green = not centrifuged). (c) PC1 (solid line) and PC3 (dashed line) loadings plot of WLP005 samples. (d) PC1
(solid line) and PC2 (dashed line) loadings plot of WLP570 samples. (For interpretation of the references to colour in this gure legend, the reader is referred to the web
version of this article.)
284 S. Grassi et al. / Food Chemistry 155 (2014) 279286
from the supernatant) with a total of 497 spectra. Before regression
analyses different spectral regions were selected following the sug-
gestions made by McLeod et al. (2009). From 7500 to 5500 cm
1
and from 4700 to 4350 cm
1
for Brix and pH and from 10,500 to
5500 cm
1
and from 4700 to 4000 cm
1
for biomass. Different
pre-treatments were tested (SNV and rst derivative) always fol-
lowed by mean centering. A double cross-validation in both PLS
and LWR models. A double cross-validation procedure was per-
formed following these steps:
(1) Select one of the batches (whole fermentation) as external
test set.
(2) The remaining batches conformed the training/calibration
set.
(3) Perform the calibration model using leaveone batch out
cross validation methodology.
(4) Predict the test set batch with the calibration model.
(5) Select another batch to be considered as external test set and
come back to step 2.
This ow-chart was repeated as many times as batches com-
prise the full dataset. In our case, twelve times. The gures of merit
reported are the mean value of the gures of merit calculated in
every iteration. In this way, more robust and reliable models can
be constructed compared to a single cross-validation procedure,
especially in the cases where there are not many samples available.
The gures of merit of the PLS models are reported in Table 2. The
number of latent variables (LVs) was selected by inspecting root
mean square error of calibration (RMSEC) and cross-validation
(RMSECV), selecting the number of LVs that led to a minimum of
both RMSEC and RMSECV and, most important, containing infor-
mation about the chemical features of the samples. Moreover,
the root mean square error of prediction (RMSEP) and the coef-
cient of determination (R
2
) were also reported for all the models.
Inspecting the regression vectors of the models for Brix and pH
it was noticed that the variables in the region 4500 to 4000 cm
1
were the one explaining most of the variance of the model, no mat-
ter the pre-treatment used. This nding conrmed what has been
stated by McLoad et al. (2009). This region, related to sugar and
ethanol content, was selected for further PLS analysis (Table 2).
For what concern the model for biomass no pre-treatment was
used because the main idea was to describe the dispersion of the
cells in the fermenting beer. The region from 10,500 to
5500 cm
1
was the one remarkable in the corresponding regres-
sion vectors, mainly describing the slope changes in this region,
i.e. the scattering effect of the yeast in the media. With regard to
the Total Solid content (Brix) excellent models were obtained
for all the pre-treatments and the variable selected tested. In par-
ticular, the PLS model obtained with the spectra pre-treated with
SNV and using the selected spectral region presented coefcient
of determination (R
2
) in calibration, cross-validation and predic-
tion of 0.975, 0.971 and 0.970, respectively and a small error
(RMSE) in calibration, cross validation and prediction (0.227,
0.250 and 0.230, respectively). This highlights the extreme impor-
tance that data exploration step has when dealing with FT-NIR,
showing that the small spectral area between 4500 and
4000 cm
1
is, indeed, the most informative one. The use of rst
derivative in addition to the SNV caused a general slight improve-
ment of the gures of merit, assessing that the best models were
those with minimum pre-processing.
For what concerns the models obtained for pH, the results were
of lower quality. The best performance was obtained by calculating
the model with spectra transformed with SNV, with variable selec-
tion, using 5 LVs. This lack of good performance for pH and its log-
arithmic behaviour suggested that there might be a nonlinear
relationship between the spectra and the pH.
With reference to biomass the models obtained were good,
especially when the region 10,5005,5000 cm
1
was selected (in
Table 2
Figures of merit of the PLS and LWR models obtained for total solid content (Brix), pH and biomass.
Calibration Cross-validation Prediction
Model Parameter Range Pre-treat Variable
selection
LV Global
PC
Local
points
Local
LV
a Iteration R
2
RMSEC R
2
RMSECV R
2
RMSEP
PLS Brix 4.9
9.4
SNV None 4 0.969 0.255 0.964 0.275 0.964 0.249
SNV 4500
4000 cm
1
4 0.975 0.227 0.971 0.250 0.970 0.230
SNV + d1 None 5 0.977 0.220 0.970 0.253 0.970 0.237
SNV + d1 4500
4000 cm
1
3 0.968 0.256 0.964 0.272 0.964 0.253
pH 5.71
4.27
SNV none 4 0.773 0.201 0.752 0.208 0.743 0.205
SNV 4500
4000 cm
1
4 0.872 0.105 0.839 0.168 0.827 0.166
SNV + d1 None 5 0.865 0.153 0.818 0.179 0.811 0.175
SNV + d1 4500
4000 cm
1
4 0.796 0.190 0.756 0.207 0.746 0.203
Biomass
*
0.14
3.19
None None 5 0.911 0.267 0.894 0.291 0.829 0.262
None 10,500
5500 cm
1
5 0.922 0.249 0.908 0.271 0.848 0.240
LWR Brix 4.9
9.4
SNV + d1 3 50 4 0.00 5 0.987 0.128 0.960 0.291 0.962 0.266
3 50 4 0.50 5 0.987 0.126 0.960 0.294 0.960 0.269
3 50 4 0.75 5 0.988 0.141 0.906 0.289 0.961 0.259
pH 5.71
4.27
SNV + d1 3 50 4 0.00 5 0.987 0.055 0.917 0.122 0.913 0.117
3 50 4 0.50 5 0.987 0.046 0.923 0.116 0.921 0.112
3 50 4 0.75 5 0.974 0.063 0.915 0.123 0.910 0.119
0.14
3.19
None 3 50 3 0.00 5 0.956 0.184 0.910 0.271 0.851 0.216
Biomass
*
None 3 50 3 0.50 5 0.951 0.166 0.910 0.270 0.852 0.211
None 3 50 3 0.75 5 0.953 0.170 0.905 0.273 0.852 0.212
SNV = standard normal variate, d1 = rst derivative (window size 7, second polynomial order), global PC = number of principal components in the global model, local
LV = number of latent variables in the local model, a = weighting of y-distances in selection of local points, iteration = iterations in determining local points, R
2
= coefcient of
determination, RMSEC = root mean square error of calibration, RMSECV = root mean square error of cross-validation, RMSEP = root mean square error of prediction.
*
O.D. (optical density) for biomass at 620 nm. The bold values denote the models chosen as the models with the best performance.
S. Grassi et al. / Food Chemistry 155 (2014) 279286 285
prediction the R
2
was 0.848 and the RMSEP = 0.240). However, LV1
against LV2 plot revealed the inuence of the different experimen-
tal conditions (strain inoculated) which can effect in the models
results in the same extent as a nonlinear relationship.
LWRPLS regression results are summarised in Table 2. For all
the models reported the number of local point (neighbours) was
set to 50; the number of global component set to 3; the number
of local latent variables set to 3 or 4; and the interactions set to
5. The alpha parameter, the relative weight of the Euclidian dis-
tance to prediction in the selec tion of the subset of calibration
spectra, was tested at different level (0, 0.5 and 0.75). For pH
and biomass, the best models were achieved with a = 0.50, in-
deed the coefcients of determination were higher than 0.95 in
calibration, 0.91 in cross-validation and 0.85 in prediction. For
what concern Total Solid Content (Brix) the best model was ob-
tained with a = 0.75 with R
2
in prediction of 0.961. The LWR root
mean errors in prediction were lower than 0.054 in the case of
pH and 0.029 for biomass if compared with the ones obtained
with PLS methodology. For what concern Brix the LWR did
not improve the quality of the gure of merit obtained by PLS
models.
4. Conclusions
In this study the application of FT-NIR spectroscopy as a method
to monitor beer manufacturing parameters, i.e. soluble solid con-
tent (Brix), pH and biomass, with a simple centrifugation and
without any sample preparation was investigated.
Different multivariate data techniques were applied on the
spectroscopic data. PCA results demonstrated that is possible to
follow biomass evolution, i.e. concentration of suspended cells in
the medium, and the evolution of fermentation from a chemical
point of view, no matter the sampling method adopted.
The use of PLS provided acceptable models for Brix, pH and
biomass determination, even though the results suggested a possi-
ble nonlinear relationship between the spectra and the parameter
investigated. The better results obtained with LWRPLS technique
instead of the linear PLS conrmed the nonlinearity relationship
and permitted to achieve precision and robustness models to
determine Brix, pH and biomass, no matter the sampling method
used.
The results obtained in this research suggest the robust and reli-
able possibility that the implementation of NIR instruments, with
the proper conguration, provides for on-line industrial brewing
systems to monitor Brix, pH and biomass evolution during the
beer fermentation. Supplying industries with a systemgiving infor-
mation in real-time would allow the assurance of meeting the
main parameters to the dened specications and the action in
case of non standard trends detection.
Acknowledgements
The authors thank Michael Crafack, University of Copenhagen,
Department of Food Science Food Microbiology for providing
the yeast strains used and the material for the performance of
the microbiological analysis.
References
Bamforth, C. W. (2006). Brewing, new technology. Cambridge: Woodhead Publishing.
Beauvoit, B., Liu, H., Kang, K., Kaplan, P. D., Miwa, M., & Chance, B. (1993).
Characterization of absorption and scattering properties for various yeast
strains by time-resolved spectroscopy. Cell Biophysics, 23, 91109.
Berner, T. S., & Arneborg, N. (2001). The role of lager beer yeast in oxidative stability
of model beer. Letters in Applied Microbiology, 54, 225232.
Bevilacqua, M., Bucci, R., Materazzi, S., & Marini, F. (2013). Application of near
infrared (NIR) spectroscopy coupled to chemometrics for dried egg-pasta
characterization and egg content quantication. Food Chemistry, 140, 726734.
Bock, J. E., & Connelly, R. K. (2008). Innovative uses of Near-Infrared spectroscopy in
food processing. Journal of Food Science, 73(7), R91R98.
Boulton, C., & Quain, D. (2006). Brewing yeast and fermentation. Oxford: Blackwell
Science.
Cleveland, W. S., & Devil, S. J. (1988). Locally weighted regression: An approach to
regression analysis by local tting. Journal of the American Statistical Association,
83(403), 596610.
Di Egidio, V., Olivieri, P., Woodcock, T., & Downey, G. (2011). Conrmation of brand
identity in foods by near infrared transectance spectroscopy using
classication and class-modelling chemometric techniques The example of
a Belgian beer. Food Research International, 44, 544549.
Directive 92/84/EEC: OJ L 316, 1992. 31.10.1992 Bull. 101992.
Duarte, I. F., Barros, A., Almeida, C., Spraul, M., & Gil, A. M. (2004). Multivariate
analysis of NMR and FTIR data as a potential tool for the quality control of beer.
Journal of Agricultural and Food Chemistry, 52, 10311038.
Engel, J., Blanchet, L., Buydens, L. M. C., & Downey, G. (2012). Conrmation of brand
identity of a Trappist beer by mid-infrared spectroscopy coupled with
multivariate data analysis. Talanta, 99, 426432.
Ghasemi-Varnamkhastia, M., Mohtasebi, S. S., Rodriguez-Mendeza, M. L., Gomes, A.
A., Ugulino Arajo, M. C., & Galvo, R. K. H. (2012). Screening analysis of beer
ageing using near infrared spectroscopy and the Successive Projections
Algorithm for variable selection. Talanta, 89, 286291.
Huang, H., Yu, H., Xu, H., & Ying, Y. (2008). Near infrared spectroscopy for on/in-line
monitoring of quality in foods and beverages: A review. Journal of Food
Engineering, 87(3), 303313.
Inon, F. A., Llario, R., Garrigues, S., & de la Guardia, M. (2005). Development of a PLS
based method for determination of the quality of beers by use of NIR spectral
ranges and sample-introduction considerations. Analytical and Bioanalytical
Chemistry, 382, 15491561.
Jolliffe, I. T. (1986). Principal component analysis. Springer-Verlag. Journal of Food
Science, 73(7), R92R98.
Kourti, T. (2005). Application of latent variable methods to process control and
multivariate statistical process control in industry. International Journal of
Adaptive Control and Signal Processing, 19(4), 213246.
Kourti, T. (2006). The process analytical technology initiative and multivariate
process analysis, monitoring and control. Analytical and Bioanalytical Chemistry,
384, 10431048.
Kramer, R. (1998). Chemometric techniques for quantitative analysis. New York:
Marcel Dekker Inc.
Kramer, R., Workman, J., Jr., & Reeves, J. B. I. I. I. (2004). Qualitative analysis. In C. A.
Roberts, Jr., J. Workman, Jr., & J. B. Reeves, III (Eds.), Near-infrared spectroscopy in
agriculture. Madison, Wis: American Society of Agronomy Inc., Crop Science
Society of America Inc., and Soil Science Society of America Inc., Publishers.
Lachenmeier, D. W. (2007). Rapid quality control of spirit drinks and beer using
multivariate data analysis of Fourier transform infrared spectra. Food Chemistry,
101, 825832.
Lewis, M. J., & Young, T. W. (2002). Brewing (2nd ed.). New York: Klumer Academic/
Plenum Publishers.
Martens, H., & Ns, T. (1989). Multivariate calibration. Chichester: Wiley.
McLeod, G., Clelland, K., Tapp, H., Kemsley, E. K., Wilson, R. H., Poulter, G., et al.
(2009). A comparison of variate pre-selection methods for use in partial least
squares regression: A case study on NIR spectroscopy applied to monitoring
beer fermentation. Journal of Food Engineering, 90, 300307.
Naes, T., & Isaksson, T. (1992). Locally weighted regression in diffuse Near-infrared
transmittance spectroscopy. Applied Spectroscopy, 46(1), 3443.
Naes, T., Isaksson, T., Fearn, T., & Davies, T. (2002). A user-friendly guide to
multivariate calibration and classication. Chichester: NIR Publications.
Naes, T., Isaksson, T., & Kowalski, B. (1990). Locally weighted regression and scatter
correction for near-infrared reectance data. Analytical Chemistry, 62(7),
664673.
Workman, J., & Weyer, L. (2008). Practical guide to interpretative Near-Infrared
spectroscopy. Boca Raton: CRC Press.
286 S. Grassi et al. / Food Chemistry 155 (2014) 279286