Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

LWT - Food Science and Technology 63 (2015) 1309e1316

Contents lists available at ScienceDirect

LWT - Food Science and Technology


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

Influence of temperature, frequency and moisture content on honey


viscoelastic parameters e Neural networks and adaptive neuro-fuzzy
inference system prediction
Mircea Oroian
Faculty of Food Engineering, Stefan cel Mare University of Suceava, Suceava, Romania

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

Article history: The aim of this study is to evaluate the influence of temperature, moisture and frequency on nine honeys
Received 16 August 2014 from viscoelastic (complex viscosity, h*, loss modulus, G00 , and storage modulus, G0 ) point of view using
Received in revised form artificial neural networks (ANN) and adaptive neuro-fuzzy inference system (ANFIS). The temperature
9 April 2015
has ranged between 5 and 40  C, the moisture content 16.04 and 17.82% and the frequency 0.1 and 10 Hz.
Accepted 18 April 2015
Artificial neural networks (Multilayer perceptron e MLP, Probabilistic neural network e PNN, Radial
Available online 27 April 2015
basis function network e RBF and Recurrent network e RN) have been used to evaluate their model of
prediction usefulness. Keeping into account the statistical parameters values, it seems that the ANNs
Keywords:
Honey
methodology predicts better the evolution of viscoelastic parameters of honey in function of tempera-
Viscoelasticity ture, frequency and moisture content than ANFIS.
ANN © 2015 Elsevier Ltd. All rights reserved.
ANFIS

1. Introduction frequency, temperature and moisture content on the honey visco-


elastic parameters has been reported.
Honey is a natural sweetener, being a supersaturated sugar so- Artificial neural network (ANN) is a system of information
lution which is composed, in its majority, by a carbohydrate processing, which is inspired by the biological nervous system
mixture (Kabbani, Sepulcre, & Wedekind, 2011). Honey is a hy- (brain). The objective of a neural network is to compute output
groscopic material; that is, it has excellent water absorbing prop- values from input values by some internal calculation (Khajeh &
erties. Thus, honey changes its moisture content according to the Barkhordar, 2013; Sadrzadeh, Mohammadi, Ivakpour, & Kasiri,
surrounding atmosphere. This characteristic is important for honey 2008). Unlike other analytical methods, where prior knowledge
storing because it absorbs water when exposed to high relative of relationships among process parameters is required, ANN draws
humidity (RH) (Camara & Laux, 2010). on previously gathered information and utilizes this when
The rheological parameters of honey are influence by temper- analyzing new data input (Jain, 2010). It is particularly useful in
ature, moisture content and presence of crystals and colloids in the managing uncertainties and non-linear data relationships. In the
product (Yanniotis, Skaltsi, & Karaburnioti, 2006). Many papers area of food quality control, ANN has been successfully applied to
have been published regarding the influence of temperature at predict food parameters (Fan et al., 2013; Khajeh & Barkhordar,
specific moisture content (Yanniotis et al., 2006),  Brix concentra- 2013; Singh, Ruhil, Jain, Patel, & Patil, 2009; Tulbek et al., 2003).
tion (Oroian, Amariei, Escriche, & Gutt, 2013; Oroian et al., 2014), Over the last years, the fuzzy logic and fuzzy inference system
chemical composition (Oroian et al., 2014). Al-Manasneh, Rababah, have been applied to identify and model complex non-linear sys-
and Ma'Abreh (2013) applied the neural network approach to tems (Karaman & Kayacier, 2011). To correctly describe the complex
evaluate the effect of temperature, shear rate and water content of and non-linear systems, fuzzy inference systems can be efficiently
the viscosity of honey. However, no other study on the influence of used with the precise rules for the prediction of certain parameters
(Ga€nzle, Kilimann, Hartmann, Vogel, & Delgado, 2007). The fuzzy
inference system can be applied to predict non-linear evolution of
food parameters such as physico-chemical or rheological properties
(Abu Ghoush, Samhouri, Al Holy, & Herald, 2008). The fuzzy
E-mail address: m.oroian@fia.usv.ro.

http://dx.doi.org/10.1016/j.lwt.2015.04.051
0023-6438/© 2015 Elsevier Ltd. All rights reserved.
1310 M. Oroian / LWT - Food Science and Technology 63 (2015) 1309e1316

inference system has been applied to predict physico-chemical and towards zero. Lower error, however, does not always mean a better
rheological characteristics of molasses (Karaman & Kayacier, 2011) network. It is possible to overtrain a network.
and mechanical properties of tapioca starch-poly (lactic acid) Cross validationxe “Cross Validation” is a highly recommended
nanocomposite foams (Lee, Hana, & Jones, 2008). criterion to stop the training of a network. Although highly rec-
The aim of this study is to predict the influence of temperature, ommended, it is not required. One will often want to try several
moisture content and frequency on the viscoelastic parameters of networks by using only training data in order to see which one
honey (complex viscosity (h*), loss modulus (G00 ) and storage works best, and then use cross validationxe “Cross Validation” for
modulus (G0 )), using the artificial neural network and adaptive the final training. When using cross validationxe “Cross Validation”,
neuro-fuzzy inference system. the next step is to decide how to divide your data into a training set
and a validation set, also called the test set. The network is trained
2. Materials and methods with the training set, and the performance checked with the test
set. The neural network will find the inputeoutput map by
2.1. Materials repeatedly analyzing the training set (NeuroDimension 6.0 Trial
version).
Nine different types of honey samples purchased from the local The cross validation idea is to split the training set into two: a set
market of Valencia, Spain, have been used for the analysis. The of examples to train with, and a validation set. The agent trains
rheological parameters of honeys can be influenced by the presence using the new training set. Prediction on the validation set is used
of crystals and air bubbles (Bhandari, Arcy, & Chow, 1999; Mossel, to determine which model to use. The error of the training set gets
Bhandari, D'Arcy, & Caffin, 2000). Before being used they were smaller as the size of the tree grows. The idea of cross validation is
warmed up to 55  C to dissolve any crystals, and kept in flasks at to choose the representation in which the error of the validation set
30  C to remove air bubbles that could interfere with rheological is a minimum. In these cases, learning can continue until the error
studies (Oroian, 2012). of the validation set starts to increase. The validation set that is used
as part of training is not the same as the test set. The test set is used
to evaluate how well the learning algorithm works as a whole
2.2. Moisture content determination
(Poole & Mackworth, 2010).
The moisture contents of honey samples were obtained by
measuring the refractive index at 20  C using a digital refractom- 2.4.1. Artificial neural networks
eter (Leica Mark II Plus). The water content and  Brix concentration The viscoelastic parameters of honeys were modeled using four
were determined based on a Chataway Table (Bogdanov, 2002). artificial neural networks such as: Multilayer perceptron e MLP,
Probabilistic neural network e PNN, Radial basis function network
e RBF and Recurrent network e RN, respectively.
2.3. Viscoelastic measurement

The dynamic rheological properties of honey samples were 2.4.1.1. Multilayer perceptron e MLP. The MLP can be considered a
obtained with a RheoStress 1 rheometer (Thermo Haake, Germany) good method for distinguishing better data that are not linearly
at different temperatures (5, 10, 15, 20, 25, 30 and 40  C), using a separable (Sarle, 1994). As seen in Fig. 1, the input data are trans-
parallel plate system (Ø 60 mm) at a gap of 500 mm. A batch of each mitted from the input (independent) layers, through the hidden
composition was prepared and at least two measurements were layers, to the output (dependent) layer in a feedforward approach.
performed on each batch, using a fresh sample for each measure- Different scaling functions could be used to transform the variables
ment. After loading the sample, a waiting period of 5 min was used at input nodes (Chatterjee & Hadi, 2006). Each neuron of the hid-
to allow the sample to recover itself and to reach the desired den layer receives input data a computes output data using a
temperature. In order to determine the linear viscoelastic region, transfer (or kernel) function (Leung, Chen, & Daouk, 2000). The
stress sweeps were run at 1 Hz first. Then, the frequency sweeps transfer function is typically a Gaussian function, defined by
were performed over the range f ¼ 0.1e10 Hz at 1 Pa stress. The 1 Pa
stress was in the linear viscoelastic region. Rheowin Job software (v.
2.93, Haake) was used to obtain the experimental data and to
calculate storage (or elastic) modulus (G0 ), loss (viscous) modulus
(G00 ), and complex viscosity (h*).

2.4. Statistical analysis

The ANNs and ANFIS were developed using the Neurosolutions


6 trial version (NeuroDimension, Inc., USA). The system is
composed of three inputs (temperature e T, frequency e f, and
moisture content e M) and three outputs (complex viscosity e h*,
loss modulus e G00 , and storage modulus e G0 ). Each model applied
for predicting viscoelastic parameters of honeys has been checked
to achieve its suitability using the mean squared error (MSE) and
mean absolute error (MAE). The viscoelastic data (complex vis-
cosity, loss modulus and storage modulus) were divided into three
groups: one group for training (60% of the data), one group for
cross-validation (15% of the data) and the last one for testing (25%
of the data). Fig. 1. Schematic diagram of a multilayer perceptron neural network model: T e
During training, the input and desired data will be repeatedly temperature, f e frequency, M e moisture content, h* e complex viscosity, G00 e loss
presented to the network. As the network learns, the error will drop modulus, G0 e storage modulus.
M. Oroian / LWT - Food Science and Technology 63 (2015) 1309e1316 1311

standard deviation which is known as a smoothing factor. As re-


ported by Jeyamkondan et al. (2001), lower smoothing factor
closely matches (over-fit) the network predictions to the target
outputs in the training data sets, but it will not interpolate well.
However, a more relaxed fit can be expected from a higher
smoothing factor, but the network may lose finer details (under-fit)
(Gosukonda, Mahapatra, Liu, & Kannan, 2015).

2.4.1.2. Probabilistic neural network e PNN. A probabilistic neural


network has similarity to back propagation model in the way they
forward. But PNN has dissimilarity in learning procedure. The ar-
chitecture of PNN consists of an input layer, pattern layer, sum-
mation layer and the output layer. The pattern layer has similarity
to radial basis network; however, summation layer has similarity of
competitive network. The pattern layer has the same number of
neurons as the input sample numbers and summation layer has the Fig. 3. Topology of radial basis function neural network: Rbf e radial basis function, Lf
e linear function, T e temperature, f e frequency, M e moisture content, h* e complex
same number of neurons as the target class. The architecture of
viscosity, G00 e loss modulus, G0 e storage modulus.
PNN used is presented in Fig. 2.

2.4.1.3. Radial basis function network e RBF. The RBF neural net-
works are an important and popular model of feedforward neural accurate predictions. Even if the non-linearity used by each unit
networks, for which fast, linear learning algorithms exist. The RBF is quite simple, iterating it over time leads to very rich dynamics.
neural networks have a single hidden layer, where the nodes are The standard RNN is formalized as follows: given a sequence of
Gaussian kernels, and a linear output layer, as illustrated in Fig. 3. input vectors (x1 … XT), the RNN computes a sequence of hidden
In contrast to the MLP structure, the Radial Basis Function (RBF) states (h1 .. hT) and a sequence of outputs (o1 … oT) by iterating
provides a flatter architecture due to the difference by which its' some equations (Sutskever, Martens, & Hinton, 2011).
incorporate the nonlinear information. The use of radial activa-
tion functions provides a nonlinear method of interpolating be-
tween numbers of different regions in the information space. RBF 2.4.1.5. Adaptive neuro-fuzzy inference system (ANFIS). The concept
networks are trained rapidly, usually orders of magnitude faster of fuzzy logic has two different meanings. In a narrow sense it is an
than MLP, while exhibiting none of its training pathologies such extension of multivalued logic and relates to classes of objects with
as paralysis or local minima problems. An RBF is a function which unsharp boundaries in which membership is a matter of degree.
has in-built distance criterion with respect to a center. One major The second meaning is that of an if-then rule of fuzzy rule, which
difference from MLP is that RBFs utilize a local learning strategy has a long history of use in artificial intelligence (Khoshnevisan,
vs. MLP's global learning, resulting thus in a higher rate of ac- Rafiee, Omid, & Mousazadeh, 2014).
curacy and faster training times (Panagou, Kodogiannis, & Nychas, The ANFIS architecture of the 1st order TakagieSugeno infer-
2007). ence system is presented in Fig. 5, where, the ANFIS structure
consists of five layers: the fuzzification layer I, the rule layer II, the
2.4.1.4. Recurrent neural network e RNN. A recurrent neural normalization layer III, the defuzzification layer IV and the output
network is a straightforward adaptation of the standard feed- layer V. In order to convert crisp data into fuzzy data, they are
forward neural network to allow it to model sequential data. At passed through the first layer which contains membership func-
each timestep, the RNN receives an input, updates its hidden tions (MFs). The output of layer 2 is referred to as the firing strength
state, and makes a prediction (Fig. 4). The RNN's high dimen- of the rule and is calculated by mathematical multiplication. In
sional hidden state and nonlinear evolution endow it with great layer 3 the normalization of the firing strengths is performed and
expressive power, enabling the hidden state of the RNN to inte- the output of this layer is called the normalized firing strengths. The
grate information over many timesteps and use it to make output of layer 4 is comprised of a linear combination of the inputs
multiplied by the normalized firing strength and layer 5 is the
simple summation of the outputs of layer 4 (Bektas Ekici & Aksoy,
2011).
The mean squared error (MSE) of an estimator measures the
average of the squares of the “errors”, that is, the difference be-
tween the estimator and what is estimated. The MSE was calculated
using the equation (1):

n  2
1X
MSE ¼ Yi;real  Yi;pred (1)
n i¼1

The MAE was computed using the equation (2):

n  
1X  
MAE ¼ Yi;real  Yi;pred  (2)
n i¼1

Fig. 2. Architecture of PNN: T e temperature, f e frequency, M e moisture content, h* where Yi,exp and Yi,cal are the experimental and calculated re-
e complex viscosity, G00 e loss modulus, G0 e storage modulus. sponses, and n is the number of the experimental run.
1312 M. Oroian / LWT - Food Science and Technology 63 (2015) 1309e1316

Fig. 4. Recurrent neural network architecture.

Fig. 5. ANFIS architecture.

3. Results and discussion and elastic modulus against frequency at 5, 10, 15, 20, 25, 30 and
40  C. The figures are presented in Fig. 6. The complex viscosity
3.1. Water content of honey samples showed a constant evolution with the frequency, exhibiting a
Newtonian behavior. The Newtonian behavior of honeys has been
The moisture content of the honey samples studied ranged observed by other scientists as well (Ahmed, Prabhtu, Raghavan, &
between 16.04 and 17.82%. The moisture content of honeys met the Ngadi, 2007; Kabbani et al., 2011; Oroian, 2012, 2013; Oroian et al.,
threshold requirement of moisture content required by the Codex 2014; Schellart, 2011). The loss modulus and the elastic modulus
Alimentarius (max. 20%) (Codex Alimentarius, 1993). have increased their magnitude with frequency. The honey loss
modulus had a greater magnitude than the elastic modulus,
3.2. Rheological measurements exhibiting a solid-like behavior. All the three rheological parame-
ters have been influenced negatively by temperature increase. The
The tests to achieve the viscoelastic parameters have been car- loss modulus was observed to be significantly higher (P < 0.05) as
ried out in the limit of linear viscoelastic region to determine the compared to the storage modulus; the frequency range employed
frequency dependence on the complex viscosity (h*), loss modulus confirming the viscous nature of honey. The viscous modulus of
(G00 ) and storage modulus (G0 ). The rheological behavior of honey honey samples systematically increased depending on angular
was investigated by plotting the complex viscosity, loss modulus frequency while inconsistent elastic modulus data were registered

Fig. 6. Storage modulus (G0 ), loss modulus (G00 ) and complex viscosity (h*) of honey as temperature function.
M. Oroian / LWT - Food Science and Technology 63 (2015) 1309e1316 1313

by the rheogram. Smaller magnitude of elastic modulus indicates reduce model complexity. This can be of special importance with
weak particleeparticle interactions and there was no network the presence of highly colinear inputs. The presence of high colin-
formation in honey samples (Ahmed et al., 2007). earity means that some model inputs are not really helping in
improving model performance. Importance analysis showed that
3.3. Viscoelastic parameters modeling using the artificial neural all three inputs (frequency, temperature and moisture content)
network methodology were important at different levels. Of the three inputs used in
viscoelastic parameter prediction, temperature had the highest
The artificial neural network methodologies have been used to importance with a relative weight ranging between 53.21 and 87.02
estimate honey complex viscosity, loss modulus and storage %, which essentially means that 53.21e87.02% of variation in pre-
modulus in function of temperature, frequency and moisture con- dicting the viscoelastic parameters was accounted for temperature.
tent. To achieve the suitable ANN, the next types of methodologies In the case of the other two inputs, they had different magnitudes.
were investigated: multilayer perceptron (MLP), probabilistic Frequency ranged between 1.08% in the case of complex viscosity,
neural network (PNN), radial basis function network (RBF) and 12.93% in the case of storage modulus and 20.87% in the case of loss
recurrent neural network (RNN). The optimum method to predict modulus. Low relative weight of frequency in the complex viscosity
viscoelastic parameter evolution has been chosen using some sta- evolution is normal due to the Newtonian behavior of honey (the
tistical parameters such as: MSE, MAE and R2. A suitable model complex viscosity does not change its magnitude with frequency).
should have minimum MSE and MAE and maximum coefficient of In the case of moisture content, relative weight ranged between
regression. For each model the data were analyzed as follows: 0.05% in the case of storage modulus, 15.91% in the case of loss
training (60% of the experimental data), cross validation (15% of the modulus and 16.16% in the case of complex viscosity.
experimental data) and testing (25% of the experimental data). The
statistical coefficients (MSE, MAE and R2) are presented for training, 3.4. Performance of ANFIS
cross validation and testing data of each model. The results ob-
tained for each model studied are presented in Tables 1e3. MLPs are Viscoelastic parameters of honey have been very well predicted
a suitable model for predicting honey complex viscosity, loss using the artificial neural network methodology. However, the
modulus and storage modulus. The regressions coefficients of the main drawback of ANNs is the black-box nature of those networks.
ANNs studied decreased as follows: MLP > PNN > RBF > RNN. The The ANFIS (Adaptive Neuro-Fuzzy Inference System) model in-
smallest MSE and MAE have been observed in the case of MLP tegrates adaptable fuzzy inputs with a modular neural network to
networks. Fig. 7 shows the evolution of experimental data vs. rapidly and accurately approximate complex functions. Fuzzy
predicted data using the most suitable model for predicting inference systems are also valuable as they combine the explan-
viscoelastic parameters with ANN (MLP-2-B-L for complex viscosity atory nature of rules (membership functions) with the power of
and loss modulus and MLP-1-B-L for storage modulus). As regards “black box” neural networks. The results of ANFIS are summarized
the MSE and MAE values for training, cross validation and testing in Table 4. Two membership functions (MF): Gaussian and Bell
for a model, the great differences between their values are made up have been studied in this article. The number of membership
by the amount of data used for each one, so the greater the amount functions applied has ranged up to five. A number of MF between
of the data is, the smaller the MSE and MAE values are; the obtained 3 and 5 was applied for each method (Gaussian or Bell). The
model is significantly suitable to the experimental data. In the case suitable method was chosen using the MSE, MAE and R2. It is
of loss modulus (G0 ) the MSE and MAE values are so high due to the observed that five is the best number of membership functions to
parameter evolution; in the case of this parameter it can be get the highest predictability of viscoelastic parameters; MSE, MAE
observed that temperature or moisture content do not have a great and R2 had better values as compared to others, the same obser-
influence on the magnitude of the parameter (Oroian, Amariei, vation was made by Al-Manasneh et al. (2013). It is well known
Escriche, & Gutt, 2013). that the increase of MF number will increase the performance of
The importance analysis was performed for all three inputs to the fuzzy model. In the case of the two membership functions,
investigate the weight of each input parameters when the visco- Gaussian is a better model than Bell because the statistical pa-
elastic parameter model was designed (the analysis was performed rameters computed are better. The statistical model parameters of
only for the suitable model to make parameter prediction). This the ANFIS model are presented in Table 4. The highest regression
step can be a useful one before designing the model and will serve coefficients (R2) are noticed in the case of complex viscosity, fol-
as a screening tool for omitting unimportant inputs in order to lowed by loss modulus and storage modulus, for each viscoelastic

Table 1
ANN statistical parameters for complex viscosity.

Model name Training Cross validation Testing


2 2
MSE R MAE MSE R MAE MSE R2 MAE

1 MLP-1-B-L 10.1 0.997 1.4 37.2 0.991 3.5 1.7 0.999 0.9
2 MLP-1-B-M 25.9 0.993 2.9 13.5 0.997 2.3 7.1 0.996 2.1
3 MLP-1-O-M 14.4 0.996 2.1 49.1 0.991 4.6 2.6 0.999 1.2
4 MLP-2-B-L 6.2 0.998 1.2 160.8 0.997 6.1 1.4 0.999 0.8
5 MLP-2-B-M 89.1 0.976 4.9 39.3 0.996 4.9 11.2 0.998 3.1
6 MLP-2-O-M 23.3 0.994 2.1 52.2 0.989 4.7 2.1 0.999 0.9
7 PNN-0-N-N 5.6 0.998 0.9 315.6 0.994 9.9 2.4 0.998 0.8
8 RBF-1-B-L 111.3 0.968 7.3 167.1 0.971 9.7 67.1 0.955 5.6
9 RBF-1-B-M 1232.6 0.669 23.7 1387.1 0.732 26.1 424.2 0.860 18.7
10 RBF-1-O-M 96.4 0.973 5.3 110.9 0.974 6.1 20.3 0.986 2.8
11 RNN-1-B-L 3439.1 0.302 41.6 3397.8 0.551 41.1 1710.6 0.546 31.6
12 RNN-1-B-M 171.6 0.950 7.9 187.1 0.976 8.7 52.1 0.974 6.1
13 RNN-1-O-M 109.3 0.970 6.8 164.3 0.962 9.2 54.6 0.970 5.2

The shaded entries in table presents the suitable ANN methodology for predicting the viscoelastical parameters.
1314 M. Oroian / LWT - Food Science and Technology 63 (2015) 1309e1316

Table 2
ANN statistical parameters for loss modulus (G00 ).

Model name Training Cross validation Testing

MSE R2 MAE MSE R2 MAE MSE R2 MAE

1 MLP-1-B-L 18,710.8 0.994 61.1 43,273 0.990 104 3809 0.997 34.6
2 MLP-1-B-M 149,045.4 0.952 232.8 144,242 0.964 247 62,928 0.960 207.1
3 MLP-1-O-M 17,130.5 0.995 59.2 34,802 0.997 86 4147 0.997 40.9
4 MLP-2-B-L 4054.3 0.999 29.1 25,858 0.993 58 1620 0.999 19.4
5 MLP-2-B-M 306,428.9 0.900 279.6 336,918 0.893 297 109,198 0.906 216.8
6 MLP-2-O-M 54,680.5 0.983 110.2 73,031 0.987 144 10,643 0.992 63.1
7 PNN-0-N-N 4578.3 0.999 27.8 242,890 0.994 177 2723 0.998 24.4
8 RBF-1-B-L 283,694.5 0.915 324.5 378,957 0.891 448 108,261 0.950 230.8
9 RBF-1-B-M 1,529,348.0 0.193 608.1 1,514,344 0.267 610 587,794 0.253 482.7
10 RBF-1-O-M 584,595.2 0.797 337.4 872,469 0.716 625 161,437 0.856 241.6
11 RNN-1-B-L 2,848,377.0 0.440 1262.1 3,303,008 0.040 1316 1,657,050 0.443 1110.4
12 RNN-1-B-M 350,250.5 0.887 387.1 477,083 0.891 507 186,725 0.878 340.9
13 RNN-1-O-M 326,072.9 0.912 436.4 288,174 0.922 451 201,726 0.917 399.6

The shaded entries in table presents the suitable ANN methodology for predicting the viscoelastical parameters.

Table 3
ANN statistical parameters for storage modulus (G0 ).

Model name Training Cross validation Testing

MSE R2 MAE MSE R2 MAE MSE R2 MAE

1 MLP-1-B-L 879 0.994 8.2 75,278.0 0.188 37.0 560.7 0.914 6.4
2 MLP-1-B-M 23,645 0.824 50.9 25,697.4 0.901 52.7 2731.1 0.780 38.2
3 MLP-1-O-M 6798 0.960 24.5 20,253.9 0.990 36.7 1343.1 0.787 16.2
4 MLP-2-B-L 872 0.994 7.3 28,925.5 0.968 26.1 588.1 0.913 6.3
5 MLP-2-B-M 64,223 0.481 99.5 65,365.3 0.519 106.3 7360.1 0.465 77.8
6 MLP-2-O-M 13,892 0.942 21.2 23,677.8 0.993 31.1 975.1 0.849 11.1
7 PNN-0-N-N 56,731 0.698 30.7 63,719.9 0.687 42.6 2243.3 0.711 12.3
8 RBF-1-B-L 38,586 0.745 115.5 82,540.5 0.609 203.3 14,153.7 0.467 84.2
9 RBF-1-B-M 75,406 0.146 102.6 72,216.3 0.112 85.6 9074.5 0.097 61.7
10 RBF-1-O-M 61,351 0.393 54.2 57,710.8 0.516 76.5 2468.7 0.617 32.6
11 RNN-1-B-L 195,459 0.107 272.2 242,254.8 0.080 306.8 121,527.3 0.282 253.8
12 RNN-1-B-M 94,959 0.113 121.9 84,427.5 0.138 91.9 15,159.6 0.054 84.4
13 RNN-1-O-M 47,768 0.613 110.8 48,046.1 0.588 116.1 12,132.0 0.582 95.4

The shaded entries in table presents the suitable ANN methodology for predicting the viscoelastical parameters.

Fig. 7. Experimental vs. predicted data using MLP e ANN (a) and ANFIS (b) for: honey complex viscosity (h*), loss modulus (G0 ) and storage modulus (G00 ) e rhombus e training,
triangle e cross validation, square e testing.
M. Oroian / LWT - Food Science and Technology 63 (2015) 1309e1316 1315

Table 4
ANFIS statistical parameters.

Model name Training Cross validation Testing

MSE R2 MAE MSE R2 MAE MSE R2 MAE

Complex viscosity (h )
*

Bell membership function 299.5 0.913 10.82 658.04 0.838 20.76 79.62 0.948 7.51
Gaussian membership function 31.02 0.991 4.13 115.3 0.982 8.09 22.99 0.984 3.33

Loss modulus (G00 )


Bell membership function 4,777,480 0.838 349 580,353 0.816 465.06 158,709 0.871 276.2
Gaussian membership function 248,712.2 0.921 286.7 663,756 0.857 586.9 88,376 0.938 201.7

Storage modulus (G0 )


Bell membership function 52,704.3 0.517 74.04 111,803 0.435 245.2 17,973 0.327 79.66
31,598.1 0.750 68.97 66,938 0.505 134.72 6261 0.518 54.85

parameter whereas the best coefficients are noticed in the case of Bogdanov, S. (2002). Harmonised methods of the international honey commission.
Liebefeld, CH-3003 Bern, Switzerland: Swiss Bee Research Centre, FAM.
Gaussian membership function. MSE and MAE low values are
Camara, V. C., & Laux, D. (2010). Moisture content in honey determination with a
noticed in the case of complex viscosity followed by loss modulus shear ultrasonic reflectometer. Journal of Food Engineering, 96(1), 93e96.
and storage modulus. Fig. 7 shows the experimental data vs. Chatterjee, S., & Hadi, A. S. (2006). Regression analysis by example (4th ed.). Hobo-
predicted ones using the ANFIS architecture e Gaussian mem- ken, New Jersey, USA: John Wiley and Sons.
Codex Alimentarius. (1993). Standard for honey. Ref. no. CL 1993/14, SH. Rome:
bership function. Codex Alimentarius Commission FAO/WHO.
The importance analysis was performed for all three inputs to Fan, F. H., Ma, Q., Ge, J., Peng, Q. Y., Riley, W. W., & Tang, S. Z. (2013). Prediction of
investigate the weight of each input parameter when the visco- texture characteristics from extrusion food surface images using a computer
vision system and artificial neural networks. Journal of Food Engineering, 118,
elastic parameters model was designed (the analysis was per- 426e433.
formed only for the suitable model for parameters prediction) for Ga€nzle, M. G., Kilimann, K. V., Hartmann, C., Vogel, R., & Delgado, A. (2007). Data
ANFIS too. Of the three inputs used in viscoelastic parameters mining and fuzzy modelling of high pressure inactivation pathway of Lacto-
coccus lactis. Innovative Food Science and Emerging Technology, 8, 461e468.
prediction, moisture content displayed the highest importance Gosukonda, R., Mahapatra, A. L., Liu, X., & Kannan, G. (2015). Application of artificial
having a relative weight ranging between 46.25 and 64.89%, which neural network to predict Escherichia coli O157:H7 inactivation on beef surfaces.
essentially means that 46.25e64.89% of variation in the prediction Food Control, 47, 606e614.
Jain, A. K. (2010). Data clustering: 50 years beyond K-means. Pattern Recognition
of viscoelastic parameters was accounted for temperature. In the Letters, 31, 651e666.
case of the other two inputs, they had different magnitudes. The Jeyamkondan, S., Jayas, D. S., & Holley, R. A. (2001). Microbial growth modelling
frequency ranged between 19.41% in the case of storage modulus, with artificial neural networks. International Journal of Food Microbiology, 64(3),
343e354.
27.04% in the case of loss modulus and complex viscosity respec-
Kabbani, D., Sepulcre, F., & Wedekind, J. (2011). Ultrasound-assisted liquefaction of
tively. In the case of temperature, the relative weight ranged be- rosemary honey: influence on rheology and crystal content. Journal of Food
tween 15.89% in the case of storage modulus and 26.7% in the case Engineering, 107, 173e178.
of loss modulus and complex viscosity. Karaman, S., & Kayacier, A. (2011). Effect of temperature on rheological charac-
teristics of molasses: modeling of apparent viscosity using adaptive neuro-
fuzzy inference system (ANFIS). LWT e Food Science and Technology, 44,
4. Conclusions 1717e1725.
Khajeh, M., & Barkhordar, A. (2013). Modelling of solid-phase tea waste extraction
for the removal of manganese from food samples by using artificial neural
The honeys studied exhibited a Newtonian behavior. The network approach. Food Chemistry, 141, 712e717.
viscoelastic parameters studied are strongly influenced by the in- Khoshnevisan, B., Rafiee, S., Omid, M., & Mousazadeh, H. (2014). Environmental
impact assessment of tomato and cucumber cultivation in greenhouses using
crease of moisture content and temperature. The loss and storage life cycle assessment and adaptive neuro-fuzzy inference system. Journal of
modulus magnitude were strongly influenced by the magnitude of Cleaner Production, 73, 183e192.
frequency, while the complex viscosity was not because of the Lee, S. Y., Hana, M. A., & Jones, D. D. (2008). An adaptive neuro fuzzy inference
system for modeling mechanical properties of tapioca starchepoly(lactic acid)
Newtonian behavior of the material. The ANNs and ANFIS were nanocomposite foams. Starch, 60, 159e164.
found to predict well the complex viscosity, loss modulus and Leung, M. T., Chen, A., & Daouk, H. (2000). Forecasting exchange rates using general
storage modulus in function of temperature, moisture content and regression neural networks. Journal of Computers & Operations Research, 27,
1093e1110.
frequency. However, the most suitable model to predict viscoelastic
Mossel, B., Bhandari, B., D'Arcy, B., & Caffin, N. (2000). Use of Arrhenius model to
parameters is the MLP-ANN one. predict rheological behaviour in some Australian honeys. LWT e Food Science &
Technology, 33, 545e552.
Oroian, M. (2012). Physicochemical and rheological properties of Romanian honeys.
References Food Biophysics, 7(4), 296e307.
Oroian, M. (2013). Measurement, prediction and correlation of density, viscosity,
Abu Ghoush, M., Samhouri, M., Al Holy, M., & Herald, T. (2008). Formulation and surface tension and ultrasonic velocity of different honey types at different
fuzzy modeling of emulsion stability and viscosity of a gumeprotein emulsifier temperatures. Journal of Food Engineering, 119(1), 167e172.
in a model mayonnaise system. Journal of Food Engineering, 84, 348e357. Oroian, M., Amariei, S., Escriche, I., & Gutt, G. (2013). A viscoelastical model for
Ahmed, J., Prabhtu, S. T., Raghavan, G. S. V., & Ngadi, M. (2007). Physico-chemical, honeys using the timeetemperature superposition principle (TTSP). Food and
rheological, calorimetric and dielectric behavior of selected Indian honey. Bioprocess Technology, 6(9), 2251e2260.
Journal of Food Engineering, 79(4), 1207e1213. Oroian, M., Amariei, S., Escriche, I., & Gutt, G. (2013). Rheological aspects of Spanish
Al-Manasneh, M. A., Rababah, T. M., & Ma'Abreh, A. S. (2013). Evaluating the honeys. Food and Bioprocess Technology, 6(1), 228e241.
combined effect of temperature content on wild-flower honey viscosity using Oroian, M., Amariei, S., Escriche, I., Leahu, A., Damian, C., & Gutt, G. (2014). Chemical
adaptive neural fuzzy inference system and artificial neural networks. Journal of composition and temperature influence on the rheological behaviour of honeys.
Food Process Engineering, 36, 510e520. International Journal of Food Properties, 17(10), 2228e2240.
Bektas Ekici, B., & Aksoy, U. T. (2011). Prediction of building energy needs in early Panagou, E. Z., Kodogiannis, V., & Nychas, G. J. E. (2007). Modelling fungal growth
stage of design by using ANFIS. Expert Systems and Applications, 38, 5352e5358. using radial basis function neural networks: the case of the ascomycetous
Bhandari, B., Arcy, B. D., & Chow, S. (1999). Rheology of selected Australian honeys. fungus Monascus ruber van Tieghem. International Journal of Food Microbiology,
Journal of Food Engineering, 41(1), 65e68. 117, 276e286.
1316 M. Oroian / LWT - Food Science and Technology 63 (2015) 1309e1316

Poole, D., & Mackworth, A. (2010). Artificial intelligence: Foundations of computa- Singh, R. R. B., Ruhil, A. P., Jain, D. K., Patel, A. A., & Patil, G. R. (2009). Prediction of
tional agents. Cambridge University Press. sensory quality of UHT milk e a comparison of kinetic and neural network
Sadrzadeh, M., Mohammadi, T., Ivakpour, J., & Kasiri, N. (2008). Separation of lead approaches. Journal of Food Engineering, 92, 146e151.
ions from wastewater using electrodialysis: comparing mathematical and Sutskever, I., Martens, J., & Hinton, G. (2011). Generating text with recurrent neural
neural network modeling. Chemical Engineering Journal, 144, 431e441. networks. In The 28th international conference on machine learning, Washington, USA.
Sarle, W. S. (1994). Neural networks and statistical models. In Proceedings of the Tulbek, M. C., Panigrahi, S., Borhan, S., Boyacioglu, M. H., Boyacioglu, D., & Clifford, H.
nineteenth annual SAS Users Group international conference (pp. 1538e1550). SAS (2003). Prediction of alkaline noodle, sensory attributes by multiple regression and
Institute. neural network models. In Proceedings of the IFT annual meetings, Chicago, USA.
Schellart. (2011). Rheology and density of glucose syrup and honey: determining Yanniotis, S., Skaltsi, S., & Karaburnioti, S. (2006). Effect of moisture content on the
their suitability for usage in analogue and fluid dynamic models of geological viscosity of honey at different temperatures. Journal of Food Engineering, 72(4),
processes. Journal of Structural Geology, 33(6), 1079e1088. 372e377.

You might also like