LWT - Food Science and Technology 77 (2017) 298e307

Contents lists available at ScienceDirect

LWT - Food Science and Technology

journal homepage: www.elsevier.com/locate/lwt

In-line characterization of ground oilseeds concentration in

solid-liquid dispersions in the food industry
Marco Marconati a, *, Adrian Haiduc b, Ste
phane Berrut b, Federico Mora b,
Andrea C. Santomaso a, Mauro Cavinato b
a  Degli Studi di Padova, Department of Industrial Engineering, via Marzolo 9, 35131 Padova, Italy
Universita
b
Nestl
e Product Technology Centre, Route de Chavornay 3, 1350 Orbe, Switzerland

a r t i c l e i n f o a b s t r a c t

Article history: Near infrared spectroscopy is proposed as a method for continuous characterization of ground oilseed
Received 18 May 2016 concentration in diluted solid liquid dispersions. The effect of operational parameters, such as flow rate
Received in revised form and temperature on the stability and reliability of the in-line measurement was investigated. Moreover,
8 November 2016
in order to widen the applicability of the approach to different materials, the variety of oilseed, the
Accepted 19 November 2016
Available online 22 November 2016
particle size distribution, and the chemistry of the suspending liquid carrier were all changed, and their
influence assessed through principal component analysis. Simple multivariate techniques were effi-
ciently used to regress spectral features against known values of dispersed solids concentration. The
Keywords:
Near infrared spectroscopy
calibration, obtained after internal cross validation, proved to give values of coefficient of determination
Chemometrics always higher than 0.95, ratio of performance to deviation higher than 4, and comparably reduced root
Particle size distribution mean squared error of cross validation. Instantaneous concentration prediction were characterized by
Solid-liquid dispersion oscillations much dependent on the dispersed solids particle size, being their amplitude lower in the case
Mixing homogeneity of finer particles. Those oscillations were characterized through the coefficient of variation that, at steady
state, gave values small enough to constitute an effective parameter to tackle mixing end-point, hence
constituting a novel way to assess homogeneity during powder incorporation in liquids.
© 2016 Elsevier Ltd. All rights reserved.

1. Introduction
Seeds constitute a common, affordable, and nutrient food whose
consumption has been steadily increasing over the past decades
(Day, 2013). This represents a significant opportunity for the
development of new methods capable of efficiently extracting and
delivering the most important nutrients of human diet from oil-
seeds, cereals, grains, and legumes. Mechanical and chemical
extraction of lipid fraction is a non-trivial operation whose yield has
to be carefully monitored. Depending on the kind of seed, after
initial mechanical pressing, solid contain in raw vegetable oil is
Abbreviations: CoV, coefficient of variation; CV, cross validation; DoE, design of
experiment; LV, latent variable; MSECV, mean squared error of cross validation;
NIR, near infrared; PC, principal component; PCA, principal component analysis;
PSD, particle size distribution; RMSECV, root mean squared error of cross valida-
tion; RPD, ratio of performance to deviation; SG, Savitzky-Golay; SNV, standard
normal variate; SSE, summation of squared errors; STD, standard deviation.
* Corresponding author. University of Surrey, Department of Chemical and
Process Engineering, Guildford, Surrey, GU2 7XH United Kingdom.
E-mail address: m.marconati@surrey.ac.uk (M. Marconati).
typically in the range of 0.02 ÷ 0.05 kgSOLIDS/kgOIL, but can reach
values up to 0.15 (Williams, 2010). Residual sediments can nega-
tively impact on further production steps or have a detrimental
effect on taste and body of the final product (Hamm, Hamilton, &
Calliauw, 2013, pp. 127e152; Johnson, 2008), therefore solid-
liquid separations steps such as sedimentation, filtration or
centrifugation are required to purify the raw pressed oil (Mohos,
2010, pp. 449e455). During soluble coffee production, for
instance, effective and fast filtering techniques have to be employed
to avoid over extraction, keeping dispersed solids weight fraction
always below a few percents (Clarke, 1985). Other than extraction
selected ground oilseeds, in particular nuts, can also be directly
incorporated into the final product stream in food industry to
improve taste, texture, and appearance: examples include almonds,
hazelnuts, cocoa and coffee seeds (Manley, 2011). In these cases,
precise powder dosing and dispersion homogeneity achievement
deeply impact on the quality of the final product and often repre-
sent key parameters to match customer expectancies. In chocolate
confectionery, for instance, cocoa powder agglomeration, caused by
oil capillary bridges, can be used to adjust the texture of the product

http://dx.doi.org/10.1016/j.lwt.2016.11.053
0023-6438/© 2016 Elsevier Ltd. All rights reserved.
 M. Marconati et al. / LWT - Food Science and Technology 77 (2017) 298e307
(Wollgarten, Yuce, Koos, & Willenbacher, 2016).
All considered, the need for the development of new, fast, and
reliable in-line concentration measuring techniques is well justi-
fied. Several technologies have been industrialized and are today
available to characterize solid-liquid dispersions, examples
including ultrasonic cells, photometric analyzers, mass flow meters,
viscometers and conductimeters. Nonetheless, not many in-line
analytical techniques proved to be fully adequate in real time
monitoring the dispersed powder weight fraction, as most tech-
nologies are affected by variations in the physical conditions of the
analyte, such as changes in the suspended particle size distribution
(PSD) (Fagan, Cullen, & O'Donnel, 2009). By virtue of that, chemical
analytical tools seem more suitable to correlate variations in sus-
pended solids concentration and, over the past years, a new gen-
eration of in-line near-infrared (NIR) diffuse-reflectance probes has
gained much attention (Huang, Yu, Xu, & Ying, 2008).
NIR spectroscopy applied to food industry has an extensive
background (Lohumi, Lee, Lee, & Cho, 2015; Porep, Kammerer, &
Carle, 2015) and it has become a popular tool for off-line analysis
of both granular materials and clear liquids (Agelet & Hurburgh,
2014). Conversely, characterization of diluted dispersions appear
to be the subject of a limited number of studies: spectroscopic
analysis in those conditions presents some unique challenges,
which include probe fouling and nonlinear scattering phenomena
caused by the presence of bubbles, agglomerates and powder
lumps (Brimmer, DeThomas, & Hall, 2001). Nonetheless, some
works have been published concerning spectroscopic analysis on
suspensions and emulsions, focusing, in particular, on crystalliza-
tion processes, determination of mixing end-point and liquid
aeration (Abebe, Wang, Li, Roberts, & Lai, 2008; Kaddour, Morel, &
Cuq, 2008; Tamburini, Marchetti, & Pedrini, 2014). Additionally, a
number of papers deal with the use of in-line spectroscopy to
assess dough properties and track flour wet agglomeration (Jirsa,
Hruskova, & Svec, 2008; Mandato, Taliani, Aït-Kaddour, Ruiz, &
Cuq, 2013). In view of this, the present study proposes a develop-
ment in the direction of assessing whether in-line measurement of
suspended ground oilseeds can be inferred through NIR spectros-
copy. Moreover, the research aims at including also the role of
temperature and flow rate, which represent key operational vari-
ables in any industrial environment, yet are seldom considered in
literature.
2. Materials & methods
2.1. Materials
In order to increase the effectiveness of the study and widen the
base of its possible applications, two different water-based sus-
pending liquid carriers, together with two varieties of the same
oilseed, namely variety V1 and variety V2, were used in the exper-
iments. The particular oilseed species tested is not deemed relevant
Table 1
Two varieties, V1 and V2, of the same oilseed genus, characterized by different lipid content, were crushed using an industrial mill and incorporated into two water-based liquid
carriers. Samples from each ground were characterized by laser diffraction (Mastersizer 3000E, Malvern Instruments Ltd, Malvern, UK) the level of comminution of seed variety
V2 was changed so that the effect of particle size distribution (PSD) in the near infrared measurement could be inferred. The laser diffraction analysis was repeated twice per
each batch of powder. Average and absolute error of 10th, 50th, and 90th percentiles are reported below.
Powder Variety PSDa
V1 Fine
V2 Fine
Medium
Coarse
a
PSD, particle size distribution.
b
Volume-based values.
299
for the applicability of the experimental approach to other types of
oilseeds. The lipid fraction of both varieties accounts for less than
0.25 kg/kg of the seeds dry weight, and is mainly composed of
triacylglycerols, sterols and tocopherols, which represent typical
compounds found in most edible vegetable oils (Pomeranz, 1991,
pp. 248e298). The main difference between the two varieties
tested is found in the lipid fraction: seed variety V1 has double oil
content compared to seed variety V2. Proteins and organic acids,
account for less than 0.15 kg/kg of the seeds dry mass, while the
remaining portion is given by carbohydrates (0.40 kg/kg), minerals,
and other minor components.
The two seed varieties were finely ground to similar values of
particle mean diameter with an industrial mill. The investigation of
PSD effect on NIR measurements was possible varying the level of
comminution of seed variety V2. The grounds were characterized
through a laser diffraction analysis (Mastersizer 3000E, Malvern
Instruments Ltd, Malvern, UK) repeated twice, mean values being
reported in Table 1.
Concerning the liquid phase, it was decided to firstly run the
experiments using distilled water (liquid A) in order to ensure
consistency among the experiments, as variations in the liquid
phase composition and ionic strength have reported to affect NIR
spectra (Hashimoto, Sugimoto, Suehara, & Kameoka, 2011). Next,
the experiments were run in a more complex liquid matrix (liquid
B) consisting of water containing a significant quantity of soluble
solids (0.55 ± 2 kg/kg). Liquid B is both denser and much thicker
than liquid A (r ¼ 1250 ± 60 kg/m3, m0 ¼ 0.80 ± 0.01 Pa s at 25  C)
and is deemed representative, in terms of viscosity, of some pro-
cessing conditions such as soaking of soybean during soy milk
production or preparation of linseed mucilage (Golbitz, 1992;
Khatib, Herald, & Boyer, 2002). The choice of using this complex
liquid matrix, characterized by a slightly variable soluble solids
content was intended to assess the effectiveness of NIR spectros-
copy in coping with real production batches for which slightly
variations in composition can be expected. This constituted an
additional degree of freedom to assess the extent of probe sensi-
tiveness to changes in liquid media properties. For consistency, the
investigated range of dispersed powder weight fraction in the solid
liquid dispersions spanned values between 0 and 0.075 kg/kg, with
increments of 0.015 kg/kg, in both liquids A and B. Values in this
range are of relevant importance in oilseed processing, examples
being found in the production of soymilk and tofu (Golbitz, 1992).
2.2. Equipment
Spectroscopic measurements were taken in diffuse reflectance
mode using a fixed diode array near infrared spectrometer, model
SentroPAT FO, coupled with a SentroProbe DR LS probe tip (Sen-
tronic GmbH, Dresden, Germany). Wavelengths used in the
regression (1200 ÷ 2100 nm) are acquired with a resolution of 2 nm
and an absolute wavelength accuracy of ±1 nm. The sample
10th percentileb 50th percentileb 90th percentileb
(mm) (mm) (mm)
3±1 35 ± 5 151 ± 8
8±1 45 ± 7 173 ± 11
19 ± 3 337 ± 12 730 ± 28
27 ± 6 597 ± 31 1251 ± 43
300 M. Marconati et al. / LWT - Food Science and Technology 77 (2017) 298e307
Fig. 1. Schematic drawing of the experimental setup used during the study. The rig is
composed of a stirred tank (a) and an external jacketed pipework (b) along which the
solid-liquid dispersion is continuously recirculated by means of a positive displace-
ment pump (c). Temperature is controlled with external heat exchanger unit. A ther-
mocouple (d), placed in proximity of the near infrared (NIR) probe, was used to record
the temperature and ensure the absence of fluctuations around its set-point. The NIR
in-line probe (e) was installed in the vertical section of the external tubing, more than
25 pipe diameters away from the discharge section of the pump to allow for the full
development of the laminar velocity profile. Dispersed phase consists of two varieties
of ground oilseed (V1 and V2), different by lipids content, and its concentration in the
dispersion was known a priori by mass balance.
acquisition rate was about 1 spectra per second, as resulting from
the average of 50 spectra taken with an integration time of 0.01 s.
The choice of those operating settings was the result of a pre-
liminary investigation.
In-line sampling for a given suspended phase concentration was
possible using a continuous recirculation rig composed of a jac-
keted stirred tank and an external piping through which a positive
displacement pump continuously recirculated the dispersion
(Fig. 1). Temperature control was ensured by means of an external
heat exchanger unit (accuracy ±0.5  C in temperature range
25 ÷ 55  C) and temperature readings were taken in proximity of
the NIR probe slot using a thermocouple. Infrared measurements
were carried out through the 20 cm2 sapphire window probe tip
located centrally within the pipe section to avoid fouling. Flow rate
was adjusted by means of the pump gearbox and different motor
speed was used during the experiments, namely LOW (160 ± 10 L/
h) or HIGH (450 ± 20 L/h). Working at those volumetric flow rates
results in laminar regime within the pipe, as the calculated Rey-
nolds number falls below 2000.
Table 2
In-line near infrared (NIR) spectroscopy was used to infer dispersed ground oilseed concentration in diluted solid-liquid dispersions
(<0.075 kgSOLIDS/kgTOT) continuously recirculated in a batch pilot unit. Investigated experimental variables include variation of flow rate, tem-
perature, particle size distribution (PSD) and variety of the ground oilseed. NIR spectra samples at steady state were used in multivariate analysis
to assess the sensitivity of the analyzer to process conditions and develop a calibration in respect of dispersed ground oilseed concentration.
Liquid carrier
Powder variety
Number of steady state samples
Dispersed ground oilseed weight fraction
PSDa
Temperature range
Flow rate
a
PSD, particle size distribution.
b
The corresponding percentiles of the distribution are reported in Table 1.
2.3. Operating procedure
Ground solids were incorporated in small discrete steps
(þ0.015 kg/kg per addition step) into the recirculating liquid carrier
consistently leaving 15 min for complete powder incorporation.
Pump speed and/or temperature were changed following the
experimental variables (Table 2). Prior to each NIR measurement,
the dispersion was allowed to achieve steady state, a condition
clearly identifiable looking at the stable profiles of absorbance.
Average spectra, obtained after a 2-min continuous sampling at
steady state (in the following referred to as steady state samples)
were regressed against the reference values of suspended solid
concentration. These latter quantities were computed through
mass balance, carefully weighting the suspended and suspending
phases prior the experiment. Moreover, any liquid contamination
during tests was avoided as the rig was completely washed and
drained between experiments. This consistently followed experi-
mental procedure ensured a maximum error on reference values of
concentration below 104 kg/kg even when accounting for possible
water evaporation (e.g. liquid A trials). On the whole, a total of 320
steady state samples were processed. Among these, 170 were
collected working with liquid A dispersions and the remaining 150
with liquid B.
3. Data analysis
Near infrared light has enough energy to promote changes in
vibration modes of molecules. Overtones and combination tones of
stretching and bending vibrations of OeH, NeH, SeH and CeH
leave the most relevant bands in the NIR spectrum. The relative
weight amount of a specific chemical groups within a sample can
theoretically be inferred through the Beer's law (Griffiths & Dahm,
2008). This relation states that non-dimensional absorbance is
linearly proportional to the weight fraction of the absorber. Even if
theoretically exact only for non-scattering samples taken in trans-
mission mode, Beer's law can be applicable to diffuse reflectance
measurements (Dahm & Dahm, 2001), where absorbance at each
wavelength can be computed through Eq. (1), in which reflectance
is given by the normalized ratio of the backscattered fraction (IR) of
the incoming radiation (I0).


1 IO


Absorbancejl ¼ log ¼ log
(1)
Reflectancejl IR l
Nonetheless, due to the complexity of light scattering theories,
absorbance is not always the most appropriate quantity of choice to
regress concentration of NIR absorbers, as other formulations are
Liquid A Liquid B
V1 and V2 V1 and V2
170 150
0 ÷ 0.075 kg/kg 0 ÷ 0.075 kg/kg
Fineb Fineb
Mediumb
Coarseb
25 ÷ 40  C 25 ÷ 50  C
HIGH (450 ± 20 L/h) HIGH (450 ± 20 L/h)
LOW (160 ± 10 L/h) LOW (160 ± 10 L/h)
 M. Marconati et al. / LWT - Food Science and Technology 77 (2017) 298e307
proposed (Griffiths & Dahm, 2008). For instance, considering just
reflectance might lead to significant improvements in enhancing
spectral features, as proved in the present study.
Absorption bands of most chemical compounds in the NIR re-
gion are typically broad, leading to spectra that often lack a clear
detailed structure, consequently multivariate (multiple wave-
length) data analysis tools are required to extract the desired
chemical information from spectral data sets. Raw spectral features
are enhanced by correcting the baseline and improving the reso-
lution of weak peaks through filtering and differentiation (Dahm &
Dahm, 2001). In the present study, simple spectra preprocessing
techniques based on mean centering, standard normal variation
(SNV) and 1st order differentiation resulted adequate to meet the
objective. Slightly more refined techniques, often used in NIR off-
line analysis such as 2nd order differentiation, were tried, but
rejected due to the high experimental noise experienced.
Operationally, the whole data analysis was implemented in
MATLAB (The MathWorks Inc., Natick, MA, USA), and it consisted of
three main steps: 1) visual analysis of raw spectra to determine the
filtering window required; 2) spectra preprocessing through mean
centering or SNV, followed by Savitzky-Golay (SG) 0th order
filtering, or SG 1st order derivation (Rinnan et al., 2009); 3) launch
of multivariate algorithms to inspect the presence of outliers, to
identify the role of operating variables and to obtain a calibration in
respect of suspended solids concentration. In this latter step well-
established algorithms, such as Principal Component Analysis
(PCA) and Partial Least Squares Regression (PLSR), were used. The
former minimizes multicollinearity performing a linear decompo-
sition of pre-processed spectra. That is achieved seeking new
orthogonal directions in space, termed principal components (PCs),
which aim at explaining in the best way the data observed vari-
ability. Each PC, defined as the eigenvector of the preprocessed data
variance-covariance matrix, orderly explains a fraction of the
observed variance of the original data set, the greatest part being
explained by the first PC (i.e. PC1). On the other hand, PLSR
simultaneously decomposes both the predictors (NIR spectra) and
the responses y (i.e. dispersed solids weight fraction) through a set
of components called latent variables (LVs). Whilst, the first LVs
better explain most of the relation between predictors and re-
sponses, higher LVs tend instead to model random fluctuations
therefore contributing to statistical over fitting (Wold, Sjostroma, &
Eriksson, 2001).
Calibration performances over the training subset were evalu-
ated through the coefficient of determination (R2 statistics) as
defined by Eq. (2) (Draper & Smith, 1998, pp. 135e140), in which y is
the set of reference values, being y their mean, and y b is the
regression-predicted set of values.
 
 2
y  b
y 
 
R2 ¼ 1    (2)
 2
y  y 
Although the R2 statistic does not constitute an absolute crite-
rion for comparing different calibrations, when its value is lower
than 0.95 the robustness of the regression is considered inadequate
to withstand continuous operation in a factory environment
without the need for frequent adjusting and recalibration (Brimmer
et al., 2001).
Additionally, being R2 sensitive to over-fitting (Romia &
Bernardez, 2009), the criterion of optimal choice for the number
of latent variables in the regression model was rather based upon
minimization of the mean squared error of cross validation
(MSECV). Using cross validation (CV), the original data set is split
into a calibration set, on which the PLSR is build, and an internal
301
validation set. The error of prediction on the latter subset is
computed through Eq. (3): in which nCV is the number of samples
validated by CV and SSECV is the summation of squared errors of
internal prediction.
1
MSECV ¼ SSECV (3)
nCV
The root mean square of MSECV (RMSECV), sharing the same
unit of measurement of the response, has the physical meaning of a
standard deviation of prediction and therefore represents a rele-
vant performance index. Moreover, upon the definition of RMSECV,
the computation of a ratio of performance to deviation (RPD) on the
CV set is straightforward, as shown in Eq. (4) (Zude, 2008, pp.
215e216). This parameter gives a useful indication of the perfor-
mance of the regression over its operational range, higher RPD
being desirable. Even though, there is no univocal agreement on the
range the RPD shall assume for a calibration to be considered
satisfactory, it has become common to consider a fit to be accept-
able when RPD>2 (Bellon-Maurel, Fernandez-Ahumada, Palagos,
Roger, & McBratney, 2010).
STDð yÞ
RPD ¼ (4)
RMSECV
In performing the CV two-third of the total sample population
were used as a training subset, the remaining as a validation set
(Kramer, 1998, pp. 3e25). For better statistical relevance and con-
sistency in the results, it was convenient to repeat the random data
splitting several times to obtain average values of MSECV and RPD.
This was done by setting 10 Monte-Carlo repetitions in MATLAB,
even though different methods can be found elsewhere (Sileoni,
Marconi, Perretti, & Fantozzi, 2013).
Other than computing the mean values of each steady state
samples, the PLSR coefficients allow for real-time concentration
predictions. The in-line concentration values, within every steady
state sample, were subsequently used to study the homogeneity of
the dispersion through the introduction of the coefficient of vari-
ation (CoV) of the i-th steady state sample. That indicator is defined
as the ratio of STD and average powder weight fraction (Eq. (5)).
 
STD b
yi
CoVi ¼ (5)
yi
Values of CoV between 0.03 and 0.06 () are often used to quantify
mixing end point homogeneity in food industry (Slettengren,
Heunemann, Imbach, Wethli, & Windhab, 2016).
4. Results and discussion
4.1. Spectral features
The regression analysis was developed averaging 120 spectra
taken during a 2-min in-line reading per each steady state sample.
Typical examples of those raw recorded spectra, in the case of the
two liquid carriers, at different pump speeds and dispersed powder
weight fractions, are illustrated in Fig. 2. Shape and noise well
distinguish the spectra of the two liquid dispersions: the low op-
tical density of liquid A led to signal saturation at low dispersed
powder concentrations; conversely, instrumental noise was less
important for the much thicker liquid B, whose low and better
defined spectral features were preserved using smaller SG filtering
window (23 points instead of 25). Variations in dispersed powder
weight fraction (0 ÷ 0.075 kg/kg) also have a remarkable influence
on the shape of NIR spectra: increasing powder weight fraction
302 M. Marconati et al. / LWT - Food Science and Technology 77 (2017) 298e307

Fig. 2. Examples of near infrared (NIR) absorbance spectra, resulting from averaging 2-min in-line sampling of continuous recirculating solid-liquid dispersions in which the
dispersed phase consisted of two different varieties of ground oilseeds (V1 and V2) and the suspending phase was either distilled water (liquid A) or a water-based solution with a
variable content of soluble solids (liquid B). Spectra of powder variety V2 dispersed in liquid A are shown at the top, whilst spectra of powder V1 dispersed in liquid B are displayed at
the bottom. For each liquid, profiles at different dispersed phase weight fraction (light grey 0.00, dark grey 0.03, and black 0.06 kgSOLIDS/kgTOT) and flow rate (continuous line:
160 ± 10 L/h, dotted line: 450 ± 20 L/h) are plotted. Absorbance peaks at 1450 and 1900 nm are due to water OH absorption, while the CH absorbance region is visible at around
1750 nm.

resulted in a downshift in absorbance, resulting from the increased
backscattered signal collected (Eq. (1)). Consequently, spectral
definition is enhanced, as the water eOH broad peaks (at around
1450 and 1900 nm) emerged from the baseline of the pure liquids,
and the oily nature of the powder is eventually captured (eCH peak
at around 1750 nm).
Downward spectral shifts also highlight the nonlinear impact of
flow rate (LOW-MAX) and temperature (25 ÷ 50  C) in respect of
dispersed powder concentration. The dependence of absorbance on
flow rate, observed in Fig. 2, stemmed primarily from the presence
of variations in scattering pattern homogeneity, possibly due to the
increased number of moving particles in the same unit time, or
internal fluid dynamics phenomena (i.e. vortex shedding in prox-
imity of the probe tip). Moreover, liquid B high viscosity did not
allow for an efficient de-aeration when the pump speed was
increased, resulting in a noticeable shift on NIR spectra that is not
instead observed in clear liquid A. Temperature is known to have
also an important role in NIR analysis, especially in the case of
aqueous samples: the spectral peaks related to polar groups, such
as eOH and eNH, are, in fact, highly affected by thermal variations.
Consequently, both longitudinal and vertical shifts are observed in
spectra: in particular the eOH peaks (at 1450 and 1900 nm) slightly
move to lower wavelengths due to loosen strength of hydrogen
bonds (Czarnik-Matusewicz & Pilorz, 2006; Wülfert, Kok, & Smilde,
1998). Finally, the effect of increasing and broadening the dispersed
phase PSD resulted in spectra with similar average shape, but
noticeably higher standard deviation, possibly caused by the loss of
homogeneity in the scattering pattern.
4.2. Principal component analysis
PCA assess the effect of operational variables, comparing the
behavior of single spectra to their sample-averaged values. The
analysis was performed on processed (i.e. smoothed and/or
derived) spectra and it was found that classification performances
slightly differ in respect on the pre-processing method used. Whilst
first derivation usually performed better than 0th order smoothing,
mean centering and SNV processing led to contrasting results:
higher amount of experimental noise from liquid A trials resulted in
the slightly loss of classification performance when running
normalization. Conversely, SNV processing increased the spectral
definition of liquid B experiments. In any case, clustering by con-
centration is observed on the score plot, with the first PC always
explaining more than 96% of the data-set covariance, which con-
firms that most of the variability among spectra is caused by
powder concentration, the effect of flow rate variations being
limitedly captured by higher PCs (Fig. 3). One of the main reasons
for classification by concentration is to be attributed to the differ-
ences in the amount of light scattering centers. That is particularly
true in the case of liquid A samples, in which the weight given to the
 M. Marconati et al. / LWT - Food Science and Technology 77 (2017) 298e307 303

scattering region is reflected by the high values of low-wavelengths

associated loadings (Fig. 4). The PC loadings plot also allows to
appreciate the higher importance given to variations in water
content, rather than powder -hence lipids-content: the PC loadings
for both dispersions, in fact, weight more the eOH region than the
lipid region (at around 1750 nm) of the spectra. Accordingly, that is
due to the fact the overall quantity of lipids in the analyte is still
much lower (<0.018 kg/kg) than the amount of water. PCA allows to
spot the presence of possible outliers, identified as points in the
score plot located far off from the center of the ellipse of confidence
of each steady state reading. Interestingly, it was observed that
more stable measurements (i.e. less dispersion among clusters)
could be achieved working with more concentrated dispersions.
Furthermore, different experiments and repetitions were per-
formed to investigate the role of PSD at constant dispersed powder
concentration on the clustering performance in the score plot, an
example being reported in Fig. 5. In the case presented, steady state
samples are compared at identical thermal and flow conditions, but
different grinding level of dispersed phase. It can be seen that using
broader PSD resulted in a significant increase of data point
dispersion that could be quantified by the ratio of areas of 95%
confidence ellipses.

Fig. 3. Score plot of the principal component analysis (PCA). The samples used in the
multivariate analysis consist of mean centered and smoothed near-infrared (NIR)
spectra of powder variety V2 dispersed in liquid A (distilled water). The recirculation
unit was operated at constant temperature (30 ± 0.5  C), but at different dispersed
powder weight fractions (0, 0.03 and 0.06 kgSOLIDS/kgTOT) and flow rates (160 ± 10 L/h
and 450 ± 20 L/h). Both clustering by concentration (▽ ¼ 0.00, ¼ 0.03,; ¼ 0.06
kgSOLIDS/kgTOT) and flow rate (upward triangles: 160 ± 10 L/h, downward triangles:
450 ± 20 L/h) are observed, however the latter parameter has a stronger impact on the
first Principal Component (PC1), as confirmed by the high value of its associated
variance (in brackets: 97.97% for PC1 vs 1.56% for PC2).
4.3. Regression
PLSR was used to linearly correlate in the vector of known
concentrations with the average processed spectra. The analysis
was launched on 81 steady state samples for liquid A and 136 for
liquid B. Temperature and flow rate were changed, but only a single
variety of powder was used for calibration purposes: V2 for liquid A
and V1 for B, respectively. That was purposely done to show and
compare the regression predictions on samples of diverse powder
variety.
The importance of spectra preprocessing in NIR analysis is well-

Fig. 4. Loading plot from principal component analysis performed on near-infrared (NIR) spectra of solid-liquid dispersions of fine ground oilseed variety V2 in liquid A (distilled
water). The recirculation unit was operated at constant temperature (30 ± 0.5  C), but at different dispersed powder weight fractions (0, 0.03 and 0.06 kgSOLIDS/kgTOT) and flow rates
(160 ± 10 L/h and 450 ± 20 L/h). The loadings of the first two principal components (PC1 and PC2) are respectively represented with a continuous and a dashed line. The associated
variance of the first two PC almost entirely explains most of the covariance between the different samples. The third PC loading (dotted line) accounts for less than 0.2% of the
observed variability in the data-set and is characterized by much higher fluctuations. Classification of samples along the PC1 is mainly achieved in the low-wavelength scattering
region of the NIR spectrum, which is characterized by the highest values of component loadings. Water content variations among the samples leave a more important mark on the
loading plot than lipids, as confirmed comparing the absolute values associated with the OH absorption bands (at around 1450 and 1900 nm) with the CH first overtone (at
around 1750 nm).
304 M. Marconati et al. / LWT - Food Science and Technology 77 (2017) 298e307

Fig. 5. Score plot of principal component analysis (PCA) launched over absorbance
spectra from three steady state samples of ground oilseed V2 dispersed in liquid A. The
data-set consist of in-line readings acquired at identical temperature (30 ± 0.5  C),
powder weight fraction (0.03 kgSOLIDS/kgTOT) and flow rate (160 ± 10 L/h) but the level
of comminution and the width of the particle size distribution (PSD) of the raw oilseed
was changed. Fine (B), Medium (;) and Coarse ( ) grounds, leave a different pattern
on the PCA score plot, as the smaller and narrower the distribution the better the
clustering.
reviewed in literature and the use of second order derivation is
common in off-line NIR spectroscopy, however the comparable
high experimental noise typical of the in-line setup negatively
affected instantaneous concentration predictions, thereby only
smoothing and first derivation were used. Regression performance
in respect of the pre-processing method used are reported in
Table 3. There the high value of the R2 statistics (always higher than
94%) can be appreciated despite the number of LVs was always
intentionally top-limited to 7 to prevent over-fitting. Interestingly,
liquid B exhibited similar performance figures to liquid A, but only
after spectra had been converted from absorbance to reflectance
(Eq. (1)). If classical absorbance spectra were instead used, the R2 of
the regression would be top limited to 0.96 with the same amount
of LVs (7). This highlights the importance of choosing the right kind
of processing for samples with different optical properties. The
minimum of MSECV was within the user-imposed maximum
number of components for liquid A, while it fell well above for
liquid B, as exemplified in Fig. 6. Cross validation over samples of
same powder gender and particle size always resulted in values of
MSECV lower than 4$105 kg2/kg2, hence RMSECV below 0.007 kg/
kg. Moreover, the ratio of performance to deviation always resulted
to be greater than 4, which is deemed adequate for an industrial
application of the device.
Other than absolute concentration measurement, the opportu-
nity of using the in-line probe to assess the dispersion homogeneity
was investigated. After calibration, real-time prediction of powder
weight fraction can be plotted as a function of data point. These
plots enable to spot the time trend of predicted concentrations
within every steady state and/or transient trial, computation of STD
and CoV being straightforward. It was observed that the trend of
STD was only slightly affected by the magnitude of dispersed
powder weight fraction, therefore CoV values steadily decreased as
powder weight fraction increased (Eq. (5)). Moreover, for a given
powder concentration, lower values of STD (and CoV) were ob-
tained working with finer PSD: this is confirmed by looking at Fig. 7
which shows the STD on samples of different PSDs in identical flow
rate and thermal conditions. With the sampling rate of the in-line
analyzer of approximately 1 Hz, values of STD fell below 2$103
for all the grounds but the coarsest, whose steady state predictions
remained more volatile. Corresponding CoV values are always
deemed suitable to assess dispersion homogeneity even at the low
concentration limit (0.015 kg/kg) when working with fine powders.
The effect of powder variety was assessed comparing PLS pre-
dictions over powder variety V2 and V1 samples, as illustrated in
Fig. 8. In liquid A the amount of powder V2 is well predicted
(±0.003 kg/kg) using the PLSR developed on the same powder
variety. However, a concentration over-estimation is experienced if
the model is instead used to predict the weight fraction of powder
variety V1 (Fig. 8b). The explanation is to be found in the different
quantity of proteins and lipids in the two varieties (i.e. V1 has
double oil content compared to V2), therefore specific calibrations
are required when working with the different varieties.
5. Conclusions
In this study NIR spectroscopy was used to assess concentration
and homogeneity of two varieties of the same oilseed genus in two
water-based dispersions: the in-line analyzer proved to ensure a
fast sampling rate with good precision and accuracy that represent
key features to improve process control during powder

Table 3
Partial least squares regression (PLSR) was used to correlate near infrared (NIR) spectral features with corresponding values of dispersed ground oilseed weight fraction in the
different solid-liquid dispersions. Reference values were obtained from mass balance with an absolute error lower than 104 kg/kg. Cross validation was used to build the
regression model and associated errors were consistent among 10 different random ways of splitting the original data-set. The complexity of the regression model was chosen
accordingly to the trend of the mean squared error of cross validation (MSECV) and the number of latent variables (LVs) was top-limited to 7. Regression performance,
characterized through the coefficient of determination of the calibration subset (R2), and the ratio of performance to deviation (RPD), is dependent upon the preprocessing
method used and the type of NIR spectra considered. Figures of the two extreme conditions, namely 0th order smoothing, and first derivative on standard normal variate (SNV)
spectra, are listed below.

Liquid Carrier Powder variety Nr of samples Spectra processing method Nr of LVsa MSECVb,d (kg2/kg2) R2 RPD c,d

5
A V2 81 0th order smoothing on mean-centered absorbance spectra 5 1.2$10 0.99 7.3
A V2 81 1st derivative on SNV processed absorbance spectra 4 3.4$105 0.97 4.4
B V1 136 0th order smoothing on mean-centered reflectance spectra 7 6.3$105 0.94 3.5
B V1 136 1st derivative on SNV processed reflectance spectra 7 3.2$105 0.98 4.8
a
LVs, latent variables.
b
MSECV, mean squared error of cross validation.
c
RPD, ratio of performance to deviation.
d
Mean values obtained after 10 repetitions of different ways the original data-set could be split during cross validation. Standard deviation of MSECV resulted to be always
below 0.4$105 kg2/kg2.
 M. Marconati et al. / LWT - Food Science and Technology 77 (2017) 298e307 305

Fig. 6. Partial least squares regression (PLSR) was the algorithm of choice to correlate known values of dispersed ground oilseeds concentration with the near infrared (NIR) spectra
acquired with the in-line probe. In order to improve the robustness of the regression, cross validation (CV) was used: the original data set of spectra was randomly assigned to a
validation set (one-third of the original number of samples) and a calibration set (the remaining two-thirds. The number of component (LV) of the PLSR model was chosen according
to the mean squared error of prediction calculated on the cross validation subset (MSECV). The trend of MSECV for both liquid A (solid black line) and liquid B (dotted grey line)
dispersions is illustrated above, together the error bars, which represent the standard deviation (STD) of MSECV on 10 repetitions of possible ways in which the original data-set
could be split during CV. Preprocessing of raw spectra relied on standard normal variate algorithm and 1st order differentiation for liquid A, mean centering and filtering were
instead used with liquid A.

incorporation in liquids, or other operations that involve the

Fig. 7. Standard deviation (STD) of instantaneous predicted concentration of ground
oilseed in solid-liquid dispersions at steady state can be used to assess homogeneity and
mixing end-point. The values of standard deviation, together with the corresponding
error bars, are reported as a function of the actual dispersed phase concentration for
different levels of comminution of oilseed variety V2 (B ¼ fine ground, ▽ ¼ medium
ground, and ▵ ¼ coarse ground). The data-set consisted of 10-points window averaged
predictions of instantaneous concentration from near infrared (NIR) spectra of powder
variety V2 in liquid A (distilled water) taken with a sampling rate of approximately
1 Hz at constant temperature (30 ± 0.5  C) and flow rate (160 ± 10 L/h).
presence of suspended solids in a liquid matrix. Although only
simple spectra preprocessing techniques were employed, the PLSR
figures featured values of the R2 statistics higher than 95%,
maximum values of RMSECV below 7$103 kg/kg, and RPD higher
than 4. This set of positive results pave the way for the application
of diffuse NIR spectroscopy even in low concentrated dispersions.
The study demonstrates that NIR is sensitive to all the operational
parameters investigated, chief among which, the role of variations
in suspended particle PSD might constitutes a significant insight.
Further studies shall be addressed to correlate increase in average
particle size and width of the size distribution with increasing os-
cillations in the instantaneous predicted concentration observed in
the present study. The experimental campaign also highlighted a
remarkable effect of both temperature and flow rate on the PCA
score plot. Nonetheless, those still remain factors of secondary
impact compared to variations in suspended solids concentration,
even though more advanced spectra preprocessing techniques
might be tested to optimally compensate for both thermal and flow
rate shifts.
Following the satisfactory outcomes of the study, future de-
velopments will require to introduce additional predictors in the
PLSR, such as temperature and weight percentage of soluble solids
in the case of liquid B. Additionally, investigation of a wider and
higher range of powder concentrations could be a subject of in-
terest to spread the applicability of NIR spectroscopy in other fields
of oilseed processing, thereby including oily liquid carriers other
than the water-based carrier here studied.
306 M. Marconati et al. / LWT - Food Science and Technology 77 (2017) 298e307

Fig. 8. The multivariate regressions, used to correlate the quantity of ground solids in the solid-liquid dispersion, were specifically developed over a single type of liquid carrier and
a single variety of oilseed. Regression of near infrared (NIR) spectra of dispersed ground oilseed variety V2 in liquid A (distilled water) gives precise and accurate measurements, as
confirmed comparing real-time predictions from the NIR probe (continuous line) to experimental values of seed concentration (dotted line) illustrated in the figure to the left.
However, if the same vector of regression parameters is instead used in predicting the concentration of seed variety V1 in liquid A, in the same conditions of flow rate and
temperature, accuracy is no longer guaranteed (figure to the right). In this case, the values of prediction are overestimated proportionally to the dispersed ground oilseed con-
centration, while the measurements still preserve a good degree of accuracy.

Appendix A. Supplementary data
Supplementary data related to this article can be found at http://
dx.doi.org/10.1016/j.lwt.2016.11.053.
References
Abebe, S. B., Wang, X. Z., Li, R., Roberts, K. J., & Lai, X. (2008). The information
content in NIR spectral data for slurries of organic crystals. Powder Technology,
179(3), 176e183.
Agelet, L. E., & Hurburgh, C. R. (2014). Limitations and current applications of near
infrared spectroscopy for single seed analysis. Talanta, 121, 288e299.
Bellon-Maurel, V., Fernandez-Ahumada, E., Palagos, B., Roger, J.-M., & McBratney, A.
(2010). Critical review of chemometric indicators commonly used for assessing
the quality of the prediction of soil attributes by NIR spectroscopy. Trends in
Analytical Chemistry, 29(9), 1073e1081.
Brimmer, P. J., DeThomas, F., & Hall, J. W. (2001). Method development and
implementation of near-infrared spectroscopy in industrial manufacturing
processes. In P. Williams, & K. Norris (Eds.), Near-infrared technology in the
agricultural and food industry (2 nd ed., pp. 207e210). St. Paul, MN: AACCI Press.
Clarke, R. J. (1985). Water and mineral contents. In R. J. Clarke, & R. Macrae (Eds.),
Coffee chemistry (pp. 42e81). London, UK: Elsevier Applied Science.
Czarnik-Matusewicz, B., & Pilorz, S. (2006). Study of the temperature-dependent
near-infrared spectra of water by two-dimensional correlation spectroscopy
and principal components analysis. Vibrational Spectroscopy, 40(2), 235e245.
Dahm, D. J., & Dahm, K. D. (2001). The physics of near-infrared scattering. In
P. Williams, & K. Norris (Eds.), Near-infrared technology in the agricultural and
food industry (2 nd ed., pp. 1e17). St. Paul, MN: AACCI Press.
Day, L. (2013). Proteins from land plants potential resources for human nutrition
and food security. Trends in Food Science & Technology, 32(1), 25e42.
Draper, N. R., & Smith, H. (1998). Applied regression analysis (3rd ed.). New York, NY:
Wiley.
Fagan, C. C., Cullen, P. J., & O'Donnel, C. P. (2009). Monitoring and control of mixing
operations. In P. J. Cullen (Ed.), Food Mixing: Principles and applications (pp.
107e120). Oxford, UK: Wiley-Blackwell.
Golbitz, P. (1992). Establishment of oriental food enterprises. In Proceedings of the
world conference on oilseed technology and utilization (pp. 261e267). Budapest,
Hungary: AOCS Press.
Griffiths, P. R., & Dahm, D. J. (2008). Continuum and discontinuum theories of
diffuse reflection. In D. A. Burns, & E. W. Ciurczak (Eds.), Handbook of near-
infrared analysis (pp. 22e62). Boca Raton, FL: CRC Press.
Hamm, W., Hamilton, R. J., & Calliauw, G. (2013). Edible oil processing (2nd ed.).
Oxford, UK: Wiley-Blackwell.
Hashimoto, A., Sugimoto, Y., Suehara, K., & Kameoka, T. (2011). Influences of pH and
temperature on infrared spectroscopic features of brewed coffee. Procedia Food
Science, 1, 1132e1138.
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 Engi-
neering, 87(3), 303e313.
Jirsa, O., Hruskova, M., & Svec, I. (2008). Near-infrared prediction of milling and
baking parameters of wheat varieties. Journal of Food Engineering, 87(1), 21e25.
Johnson, L. A. (2008). Recovery, refining, converting, and stabilizing edible fats and
oils. In C. C. Akoh, & D. B. Min (Eds.), Food lipids (pp. 206e221). Boca Raton, FL:
CRC Press.
Kaddour, A., Morel, M. H., & Cuq, B. (2008). Description of batter mixing using near
infrared spectroscopy. Journal of Cereal Science, 48(3), 698e708.
Khatib, K. A., Herald, T. J., & Boyer, J. E. (2002). Physicochemical characteristics of
soft tofu formulated from selected soybean varieties. Journal of Food Quality,
25(4), 289e303.
Kramer, R. (1998). Chemometric technique for qualitative analysis. New York, NY:
Marcel Dekker.
Lohumi, S., Lee, S., Lee, H., & Cho, B. (2015). A review of vibrational spectroscopic
techniques for the detection of food authenticity and adulteration. Trends in
Food Science & Technology, 46(1), 85e98.
Mandato, S., Taliani, C. C., Aït-Kaddour, A., Ruiz, T., & Cuq, B. (2013). In-line moni-
toring of durum wheat semolina wet agglomeration by near infrared spec-
troscopy for different water supply conditions and water addition levels. Journal
of Food Engineering, 119(3), 533e543.
Manley, D. (2011). Dried fruits and nuts as biscuit ingredients. In D. Manley (Ed.),
Manley's technology of biscuits, crackers and cookies (pp. 200e208). Cambridge:
Woodhead Publishing.
Mohos, F. (2010). Confectionery and chocolate engineering. Oxford, UK: Wiley-
Blackwell.
Pomeranz, Y. (1991). Functional properties of food components (2nd ed.). San Diego,
CA: Academic Press.
Porep, J. U., Kammerer, D. R., & Carle, R. (2015). On-line application of near infrared
(NIR) spectroscopy in food production. Trends in Food Science & Technology,
46(2), 211e230.
Rinnan, A., Nørgaard, L., van den Berg, F., Thygesen, J., Bro, B., & Engelsen, S. (2009).
Data pre-processing. In D. Sun (Ed.), Infrared spectroscopy for food quality
analysis and control (pp. 29e50). Cambridge, MA: Academic Press.
Romia, M. B., & Bernardez, M. A. (2009). Multivariate calibration for quantitative
analysis. In D. Wen (Ed.), Infrared spectroscopy for food quality analysis and
control (pp. 51e82). Cambridge, MA: Academic Press.
Sileoni, V., Marconi, O., Perretti, G., & Fantozzi, P. (2013). Evaluation of different
validation strategies and long term effects in NIR calibration models. Food
Chemistry, 141(3), 2639e2648.
Slettengren, K., Heunemann, P., Imbach, S., Wethli, C., & Windhab, E. J. (2016).
Production and evaluation of water based powder-liquids and powder-liquid
mixing. Journal of Food engineering, 169, 278e283.
 M. Marconati et al. / LWT - Food Science and Technology 77 (2017) 298e307 307

Tamburini, E., Marchetti, M. G., & Pedrini, P. (2014). Monitoring key parameters in
bioprocesses using ner-infrared technology. Sensors, 14(10), 18941e18959.
Williams, M. A. (2010). Separation technologies in oilseed processing. In S. Rizvi
(Ed.), Separation, extraction and concentration processes in the food, beverage and
nutraceutical industries (pp. 396e429). Cambridge, UK: Woodhead Publishing.
Wold, S., Sjostroma, M., & Eriksson, L. (2001). PLS-regression: A basic tool of che-
mometrics. Chemometrics and intelligent laboratory systems, 58(2), 109e130.
Wollgarten, S., Yuce, C., Koos, E., & Willenbacher, N. (2016). Tailoring flow behavior
and texture of water based cocoa suspensions. Food Hydrocolloids, 52, 167e174.
Wülfert, F., Kok, W., & Smilde, A. (1998). Influence of temperature on vibrational
spectra and consequences for the predictive ability of multivariate models.
Analytical Chemistry, 70, 1761e1767.
Zude, M. (2008). Optical monitoring of fresh and processed agricultural crops. Boca
Raton, FL: CRC Press.