This document describes research using a spectroscopic sensor and neural network model to monitor droplet size distributions (DSDs) in metal working fluid (MWF) emulsions. The sensor measured light absorption and scattering spectra of MWF samples. A neural network model was trained using spectroscopic data and reference DSD measurements. The model accurately estimated DSDs for new samples, distinguishing monomodal and bimodal distributions. This technique could monitor MWF emulsion aging and destabilization in industrial processes.
2. coefficient, which can be estimated from the Mie Theory. In this
way, it is possible to associate both phenomena and to obtain the
DSD from light absorption and scattering spectra with appropriate
techniques
I ¼ I0 expðÀtðlÞlÞ ð1Þ
tðlÞ ¼ Np
p
4
Z1
0
Qextðl; xÞx2
fðxÞ dx ð2Þ
Celis and Garcia‐Rubio[6,8]
and Elicabe and Garcia‐Rubio[9]
have
used algorithms to estimate the DSD in emulsions and dispersions
based on the optical properties of its components and on
spectroscopic measurements, by applying inversion methods.
This approach enables the real‐time estimation of the DSD, and
the in‐line monitoring of this emulsion property. However, these
models are not suitable for emulsions with high droplet
concentration due to multiple scattering effects, in which light
that is scattered by one droplet is affected by the light scattered by
other droplets before being detected by the sensor, leading to
inaccurate results.
Neural Network Models
An alternative approach that can be applied to emulsions under
high droplet concentration is based on pattern recognition
techniques. In this case, the spectral data measured by the
turbidimetric sensor is associated with the corresponding DSD by
means of a previously calibrated multivariate model. Among
different techniques that can be applied, nonlinear models such as
neural networks have been successfully applied by one of the
present authors in place of light scattering models to estimate size
distributions in concentrated solid–liquid suspensions.[10,11]
Figure 3 illustrates the structure of a commonly applied neural
network, that is a three‐layer feed‐forward neural network, used in
this study. The input to a neuron j in the network, consists of the
weighted sum Sj of outputs from neurons i (i ¼ 1, 2, …, q), Xi
(Equation 3). The weights, Wi,j, are model parameters that are
fitted to each specific system. The last input, Xqþ1, with value equal
to 1, is a bias
Sj ¼
Xq
i¼1
Wi;jXi þ Wqþ1;j ð3Þ
The output from neuron j is a response function Oj ¼ f(Sj), in
which f(Sj) can consist of different mathematical forms, but in most
cases is a sigmoidal function (Equation 4):
fðSjÞ ¼
1
1 À eÀSj
ð4Þ
The fitting of a neural network is divided in two parts: training,
which consists of the fitting of the parameters or weights, and
validation. In the training step, measured values of the system
outputs corresponding to known values of inputs are presented to
the network, and the best set of weights is selected so that a
minimum squared error E is achieved. E is defined in Equation (5),
where yk is the experimental value of output k and Ok is the
calculated value of output k. The fitting consists of presenting the
neural network to the set of experimental pairs of inputs and
outputs. At each presentation of the data set the weights to each
neuron (Equation 3) are changed according to the backpropagation
algorithm,[12]
so as to minimise the error E
E ¼
X
all
observations
Xp
k¼1
ðy
ðmÞ
k À O
ðmÞ
k Þ2
ð5Þ
1 (bias)
1 (bias)
Figure 3. Illustration of a feed‐forward neural network.
0
0,2
0,4
0,6
0,8
1
1,2
200 400 600 800 1000
OpticalDensityinAU
Wavelength in nm
Figure 2. Example of a light absorption and scattering spectrum.
0
2
4
6
8
10
12
14
16
18
20
0.01 0.1 1 10 100 1000
f(x)in%/mµ
New Fluid
New Fluid
in Use
Aged Fluid
Diameter, x in µm
Figure 1. Illustration of droplet size distributions obtained by the authors
with metal working fluid emulsions under different conditions.
VOLUME 92, FEBRUARY 2014 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING 319
3. The second part of the fitting consists of the model validation.
The calculated outputs are compared with experimental values for
new observations that have not been used in the training step, in
order to check if the model is able to predict the desired results.
The computational programs for neural network model fitting,
validation and simulations were developed at home in the
Chemical Engineering Department, University of São Paulo.
MATERIALS AND METHODS
Commercial metalworking fluid, Kompakt YV Neu, was obtained
from Jokisch GmbH (Oerlinghausen, Germany). The chemical
composition of this fluid was not specified by the producer, but it
contains 40% oil phase. Emulsions were prepared by dilution of
the metalworking fluid with deionised water to reach the desired
MWF concentrations (3.5–5.2%). Artificial aging was promoted by
adding to the emulsions 0–0.3% of CaCl2 (CaCl2·2H2O, purity of
99.5%), from Grüssing GmbH.
Absorbance measurements were performed with a UV–Vis–NIR
spectrometer, model HR2000 þ ES, from OceanOptics, with DH
2000‐BAL (200–1100 nm) light source, and a deep probe with
6.35 mm diameter, 127 mm of length and 2 mm optical path, which
enables in‐line monitoring.
Experimental values of the droplet size distribution, adopted as
references in the neural network training, were based on
measurements in a Malvern Mastersizer 2000 laser diffractometer,
with particle size detection range from 0.02 to 2000 mm.
RESULTS
Data Treatment
Figure 4 shows typical spectra obtained in the experiments with
emulsion samples at different times. The observed values of optical
density at different wavelengths represent the sum of absorption
and scattering by the droplets.
Due to the large number of absorbance–wavelength data
contained in each observation, a previous statistical analysis was
carried out in order to reduce the number of input variables to the
neural network, based on the relative importance of each input
variable (optical density at a given wavelength) on the variance of
the data. The selection of the most important input variables was
based on a principal component analysis (PCA). This technique
consists of transforming the original variables of a multivariate
system into non‐correlated new variables (components) that are
linear combinations of the original variables. Thus, from a number
n of original variables xj (j ¼ 1, …, n) a smaller number of p non‐
correlated components ei (i ¼ 1, …, p) are calculated, which are
linear combinations of the original variables with the form:
ei ¼ wi1x1 þ þwijxj þ þwinxn, in which the terms wij are the
loadings, or weights, of variable xj on the component ei and are
computed so that each component represents the maximum of the
system variability in decreasing order. The technique is used to
reduce the number of variables involved in an analysis, and to
detect underlying relationships among groups of variables.
Descriptions of the method are presented in books on multivariate
statistical analysis.[13]
The weights correspond to the eigenvectors
of the covariance matrix of the original variables. Components are
ordered according to the decreasing value of variances, represented
by their eigenvalues. Numerical differences of variables were
eliminated by working with standardised variables (zero mean,
standard deviation equal to 1). Thus, computations were based on
the correlation matrix. Result interpretation was based on the
absolute value of the weights wij.[14]
Based on this analysis, for the set of observations in the present
study, it was possible to reduce the number of input variables from
the original ca. 400 variables (since the resolution of the
spectrometer was about 2 nm, and the measured range was from
ca. 200 to 1000 nm) to only three most important wavelengths: 460,
695 and 943 nm. These variables represent ca. 98% of the total
observed variance in the data set.
Neural Network Fitting
A three‐layer feed‐forward neural network like the one presented
in Figure 3 was used to fit the experimental data. A total of 7
variables were used as inputs: absorbance values selected by PCA
at 460, 695 and 943 nm, concentration of oil, water and CaCl2, and
the time interval between addition of salt to the emulsion and each
measurement (aging time). Thus, the inputs consist of the
formulation of each sample, and the resulting optical density at
different times after salt addition. As outputs of the neural network
17 sizes classes were selected, from 0.04 to 10 mm, as multiples offfiffiffi
2
p
. This number of size classes was arbitrarily adopted in order to
reconstruct the DSD of the samples with appropriate resolution.
Thus, since the number of inputs and outputs is defined by the
specific characteristics of the system, then the only degree of
freedom was the number of neurons in the hidden layer (Figure 3).
In the fitting step, for each value of this number the minimum value
of the error (Equation 5) was recorded. The best fitting was
obtained with six neurons in the hidden layer, after 500 000
presentations of the data set to the neural network. Figures 5 and 6
show representative results obtained in the fitting and validation of
the model, for samples with different aging times and consequently
different DSD. Good agreement between the calculated and
experimental values was obtained for samples with monomodal
and bimodal distributions, with different proportions of each
droplet population.
These results indicate that the treatment of the optical density
data measured by the UV/Vis spectroscopic sensor by means of a
neural network model is able to provide accurate values of the
droplet size distribution for new and aged MWF, presenting
monomodal and bimodal DSD, respectively.
In case of unstable emulsions, the proposed method can be
applied as a monitoring technique before phase separation takes
place, since phase separation changes the concentration of the
emulsion components and affects the light scattering pattern,
generating false results.
0,0
0,5
1,0
1,5
2,0
2,5
3,0
400 500 600 700 800 900 1000
OpticalDensity(AU)
Wavelength(nm)
Figure 4. Absorbance spectra obtained with MWF emulsions.
320 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING VOLUME 92, FEBRUARY 2014
4. Application to the Monitoring of a Destabilisation Experiment
In order to test the spectroscopic sensor coupled with the neural
network model, an experiment was carried out in which the aging
of a MWF sample was monitored over time. Thus, 0.3% of CaCl2
was added to a MWF emulsion with concentration of 4%. The
aging was monitored over time with the spectrometer and the deep
probe, and samples were collected at desired times for droplet size
distribution measurement with the Malvern Mastersizer 2000
diffractometer for comparison with the results provided by the
sensor. The spectroscopic data consisted of the optical density at
460, 695 and 943 nm. The other input variables were the MWF
formulation, and time after addition of the salt. Thus, by running
the neural network model with data generated by the spectrometer
the DSD could be estimated over the destabilisation time.
DSD plots obtained by applying the technique to the MWF
destabilisation experiment are shown in Figure 7 for different times
after addition of CaCl2 to the emulsion. The DSD plots show that,
after salt addition, a second population is formed with much larger
droplets than in the original population. The volumetric fraction of
this population increases with time, and tends to reach appreciable
values. The location of the maximum in the DSD of this second
population also tends to increase with time, while this tendency is
not so clearly observed in the initial population. The plots also
show that the DSD curves calculated by the model are similar to
those produced by laser diffractometry. Thus, the sensor was able
to retrieve the DSD with good accuracy as well as to monitor the
time evolution of the aging process in terms of the DSD, which
changes from monomodal to bimodal distribution.
These results point out the potential of this technique for
monitoring such emulsions with possible applications in similar
systems. Under the conditions tested, the results were not affected
by multiple scattering, suggesting that this technique can even be
used in more concentrated emulsions, if the model is fitted to
representative experimental data.
CONCLUSIONS
The fitting of a multivariate model based on neural network to
associate UV/Vis optical density spectral data with the droplet size
distribution of metal working fluid emulsions has shown good
agreement for monomodal and bimodal distributions, typical of
new and aged emulsions, respectively. Since this approach is based
on an empirical model, the application of the method to other
emulsions involves a preliminary calibration step, that is the fitting
of a neural network model under the specific conditions of each
use, in order to obtain a valid model.
The number of input variables to the model was significantly
reduced by carrying out a principal component analysis, aimed at
selecting the most important variables in terms of their contribu-
tion to the variance of the data. This approach enabled the fitting of
0,0
1,0
2,0
3,0
4,0
5,0
6,0
7,0
8,0
9,0
10,0
0,01 0,10 1,00 10,00
f(x)
Diameter, x (μm)
Training
Experimental
Calculated
0,0
1,0
2,0
3,0
4,0
5,0
6,0
7,0
8,0
9,0
10,0
0,01 0,10 1,00 10,00
f(x)
Diameter, x (μm)
Training
Experimental
Calculated
B
A
Figure 5. (A) and (B) Neural network fitting results for samples with
monomodal and bimodal droplet size distributions, for observations used in
model fitting.
0,0
1,0
2,0
3,0
4,0
5,0
6,0
7,0
8,0
9,0
10,0
0,01 0,10 1,00 10,00 100,00
f(x)
Diameter, x (μm)
Validation
Experimental
Calculated
0,0
1,0
2,0
3,0
4,0
5,0
6,0
7,0
8,0
9,0
10,0
0,01 0,10 1,00 10,00
f(x)
Diameter,x (μm)
Va
B
A
lidation
Experimental
Calculated
Figure 6. (A) and (B) Neural network fitting results for samples with
monomodal and bimodal droplet size distributions, for observations used to
validate the model.
VOLUME 92, FEBRUARY 2014 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING 321
5. a neural network with a relatively small number of inputs, which
resulted in a model with a relatively small number of parameters.
The model was applied to an aging experiment and was able to
detect changes in the DSD profile over time from monomodal to
bimodal distribution with good accuracy. If no phase separation
occurs, the technique is apparently not affected by high
concentration of droplets, which causes multiple scattering effects
in such systems.
Thus, based on the results shown, the optical sensor coupled
with the neural network model can be used to monitor changes in
the emulsion structure based on the changes in the DSD. Such a
sensor can be applied in industrial processes involving emulsions,
providing real‐time information on the DSD based on in‐line
measurements by the spectroscopic sensor.
NOMENCLATURE
E squared error
ej non‐correlated components of the transformation of the
original variables in PCA
f(x) droplet size distribution function
I measured light intensity
I0 intensity of the light source
l optical path
0,0
2,0
4,0
6,0
8,0
10,0
0,01 0,10 1,00 10,00 100,00
f(x)
Diameter, x
A D
B E
C F
(μm)
Starting time
Experimental
Calculated
0,0
2,0
4,0
6,0
8,0
10,0
0,01 0,10 1,00 10,00 100,00
f(x)
(μm)
After 8 minutes
Experimental
Calculated
0,0
2,0
4,0
6,0
8,0
10,0
0,01 0,10 1,00 10,00 100,00
f(x)
(μm)
After 20 minutes
Experimental
Calculated
0,0
2,0
4,0
6,0
8,0
10,0
0,01 0,10 1,00 10,00 100,00
f(x)
(μm)
After 30 minutes
Experimental
Calculated
0,0
2,0
4,0
6,0
8,0
10,0
0,01 0,10 1,00 10,00 100,00
f(x)
(μm)
After 80 minutes
Experimental
Calculated
0,0
2,0
4,0
6,0
8,0
10,0
0,01 0,10 1,00 10,00 100,00
f(x)
(μm)
After 1040 minutes
Experimental
Calculated
Diameter, x
Diameter, xDiameter, x
Diameter, x
Diameter, x
Figure 7. Droplet size distribution of artificially destabilised MWF estimated by the sensor, compared with results obtained with the laser diffractometer. (Plot
A) before addition of CaCl2; (plots B–F) at different times after CaCl2 addition.
322 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING VOLUME 92, FEBRUARY 2014
6. Np total particle number per unit volume
Oj objective function calculated in the output of the neuron j
Ok calculated value of output k
Qext extinction coefficient
Sj weighted sum of inputs of the neural network
Wi,j weights of the neural network
wi,j weights of the linear combination in PCA
x droplet size
Xqþ1 bias
Xi inputs of the neural network
xj original variables
yk experimental value of output k
l wavelength
t turbidity
ACKNOWLEDGEMENTS
This study is part of a joint project between the Universities of São
Paulo and Bremen, within the BRAGECRIM program (Brazilian
German Cooperative Research Initiative in Manufacturing). The
authors would like to thank FAPESP, CAPES, FINEP and CNPq
(Brazil), and DFG (Germany) for the financial support.
REFERENCES
[1] M. A. El Baradie, J. Mater. Process Technol. 1966, 56, 786.
[2] J. F. G. De Oliveira, S. M. Alves, Produção 2007, 17, 129.
[3] F. Klocke, G. Eisenblätter, Ann. CIRP 1997, 46, 519.
[4] I. D. Morrison, S. Ross, Colloidal Dispersions—Suspensions,
Emulsions and Foams, 1st edition, Wiley‐Interscience, New
York 2002, p. 656.
[5] J. Deluhery, N. Rajagopalan, Colloids Surf. A Physicochem.
Eng. Aspects 2005, 256, 145.
[6] M. Celis, L. H. Garcia‐Rubio, J. Dispersion Sci. Technol. 2008,
29, 20.
[7] D. Bohren, C. F. Huffman, Absorption and Scattering of Light
by Small Particles, John Wiley & Sons, New York 1983, p. 519.
[8] M.‐T. Celis, L. H. Garcia‐Rubio, Indus. Eng. Chem. Res. 2004,
43, 2067.
[9] G. E. Elicabe, L. H. Garcia‐Rubio, Adv. Chem. Ser. 1990, 227,
83.
[10] R. Guardani, C. A. O. Nascimento, R. S. Onimaru, Powder
Technol. 2002, 126, 42.
[11] C. A. O. Nascimento, R. Guardani, M. Giuletti, Powder
Technol. 1997, 90, 89.
[12] D. Rummelhart, J. McClelland, Parallel Distributed Process-
ing Explorations in the Microstructure of Cognition, 1986, 1,
MIT Press, Cambridge, MA p. 567.
[13] R. A. Johnson, D. W. Wichern, Applied Multivariate
Statistical Analysis, Prentice‐Hall, Upper Saddle River 2002,
p. 800.
[14] I. T. Jolliffe, Principal Component Analysis, Springer, New
York 1986, p. 502.
Manuscript received December 26, 2012; revised manuscript
received February 19, 2013; accepted for publication March 07, 2013.
VOLUME 92, FEBRUARY 2014 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING 323