Printed

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

Journal of Food Engineering 51 (2002) 239±248

www.elsevier.com/locate/jfoodeng

Prediction of pores formation (porosity) in foods during drying:


generic models by the use of hybrid neural network
M.A. Hussain a, M. Sha®ur Rahman b,*
, C.W. Ng a

a
Department of Chemical Engineering, Faculty of Engineering, University of Malaya, Malaysia
b
Department of Bioresource and Agricultural Engineering, and Department of Food Science and Nutrition, College of Agriculture,
Sultan Qaboos University, P.O. Box 34, Al-Khod 123, Muscat, Oman
Received 14 October 2000; accepted 7 March 2001

Abstract
General porosity prediction models of food during air-drying have been developed using regression analysis and hybrid neural
network techniques. Porosity data of apple, carrot, pear, potato, starch, onion, lentil, garlic, calamari, squid, and celery were used to
develop the model using 286 data points obtained from the literature. The best generic model was developed based on four inputs as
temperature of drying, moisture content, initial porosity, and product type. The error for predicting porosity using the best generic
model developed is 0.58%, thus identi®ed as an accurate prediction model. Ó 2001 Elsevier Science Ltd. All rights reserved.

Keywords: Generic model; Hybrid neural network; Air drying; Porosity; Density; Thermal conductivity

2=3
1. Introduction relation when it was plotted as a function of e0 . Simi-
larly in the case of potato, a relationship between loss
Pores or bubbles occur in diverse food products, and modulus and porosity was observed (Scanlon, Day, &
the issues of pores or bubbles in foods are generic, such Povey, 1998). In case of extruded starch, porosity is
as carbonated drinks, beers, ice creams, cakes, biscuits, highly correlated with shear strength or other mechan-
breads, breakfast cereals, extruded and dried products. ical properties (Bhatnagar & Hanna, 1997; Hutchinson,
The prediction of pore formation in foods during pro- Siodlak, & Smith, 1987; Gogoi, Alavi, & Rizvi, 2000).
cessing is needed for process design, in estimating other Porosity and crispiness of snack foods by extrusion or
properties, and characterizing the quality of a product frying are highly related. Porosity plays the most im-
(Rahman, 2001). Simultaneous heat and mass transfer portant role for agglomerate strength of dried foods.
are involved in food processing operations, such as The lower the porosity or in other words, the higher the
drying, smoking, blanching, baking, frying, evaporative apparent density of the agglomerate, the stronger the
cooling, extrusion and cooking. In these processes, agglomerate. In addition to porosity, adhesion or
foods undergo volume changes either by shrinkage due binding forces of particles in the matrix (i.e. structure)
to moisture loss or by expansion due to gas generation also a€ects its mechanical strength (Pietsch, 1999). The
or pore formation. variation in porosity, average pore size and pore size
Porosity has a direct e€ect on the other physical distribution has signi®cant e€ects on the textural char-
properties, such as mass di€usion coecient, thermal acteristics of the dried foods (Huang & Clayton, 1990).
conductivity, and thermal di€usivity. Mechanical and Thus it is important to predict the formation of pores in
textural properties of food are also correlated to the foods during drying due to its importance in quality
porosity. Vincent (1989) found that torsional sti€ness characterization and process design.
(0.5±7 MPa) varied with the porosity (0.83±0.54) in case The process of pore formation during drying is very
of fresh apple. They found that sti€ness showed a linear complex. In most cases, porosity is predicted by empir-
ical correlations as a function of water content. Exam-
* ples are linear (Madamba, Driscoll, & Buckle, 1994;
Corresponding author. Tel.: +986-515-236, fax: +986-513-418.
E-mail addresses: azlan@fk.um.edu.my (M.A. Hussain), sha-
Palipane, Driscoll, & Srzednicki, 1992), quadratic (Pal-
®ur@squ.edu.om (M. Sha®ur Rahman), ngcwei98@yahoo.com ipane et al., 1992; Senadeera, Bhandari, Young, & Wi-
(C.W. Ng). jesinghe, 1998; McMinn & Magee, 1998; Rahman,
0260-8774/01/$ - see front matter Ó 2001 Elsevier Science Ltd. All rights reserved.
PII: S 0 2 6 0 - 8 7 7 4 ( 0 1 ) 0 0 0 6 3 - 2
240 M.A. Hussain et al. / Journal of Food Engineering 51 (2002) 239±248

Notation TS transfer function


Xw moisture content (mass fraction, wet basis)
F product category Xw0 initial moisture content before drying (mass fraction, wet
ke e€ective thermal conductivity (W/m K) basis)
ke0 e€ective thermal conductivity before drying e experimental porosity of dried sample at di€erent moisture
(W/m K) contents
MSE mean square error em porosity of dried sample from model at di€erent moisture
MPE absolute mean percent error contents
N number of optimum node e0 initial porosity before drying
n number of data point er porosity predicted from polynomial regression
T temperature of drying (°C) d residual …e er †

Perera, Chen, Driscoll, & Potluri, 1996; Rapusas & shrinkage and collapse during drying. However the glass
Driscoll, 1995), exponential forms (Mattea, Urbicain, transition theory does not hold true for all products.
& Rotstein, 1986), and power-law equation (Pietsch & Other concepts, such as surface tension, structure, en-
Calvelo, 1984; Wildegger-Gaissmaier & Agarwal, 1990). vironment pressure, and mechanisms of moisture
The drawbacks of these correlations are: (i) this does not transport also play important roles in explaining the
provide the physics of the process, (ii) the parameters do formation of pores (Rahman, 2001). Recently Rahman
not have any physical meaning, and (iii) this is only (2001) proposed a hypothesis that capillary force is the
limited to the speci®c products and experimental con- main force responsible for collapse and that counter-
ditions used (Rahman, 2001). Literature sources are balancing this force causes the formation of pores and
limited to the empirical correlations due to the com- reduced shrinkage. As the process of pore formation
plexity of the pore formation process. No attempt has mechanism is very complex, arti®cial neural network
been made to develop a generic model as identi®ed by can be utilized as a powerful tool for modeling. Their
the authors. The best way of developing a model is to most important attribute is that they are capable of
use fundamental knowledge if possible. Our funda- learning or adapting to various nonlinear and complex
mental understanding of the underlying scienti®c prin- data set such as pore formation. The ®rst arti®cial
ciples that govern the formation of pores is minimal. In neuron model was presented by MuCulloch and Pitts
a ®rst attempt, Rahman (2000) identi®ed the four ge- (1943) and later it has been extended by others (Ru-
neric trends of pore formation in foods during drying. melhart & McClelland, 1986).
The formation of pores in foods during drying can be Arti®cial neural networks are emerging as a promis-
grouped into two generic types: one with an inversion ing tool in process identi®cation and controls owing to
point and another without an inversion point (Figs. 1 their ability to model process with accuracy (Hussain &
and 2) (Rahman, 2000). These generic trends are pro- Rahman, 1999; Hussain & Kershenbaum, 2000). They
posed based on the experimental evidence published in also o€er a cost-e€ective method of developing useful
the literature. Fig. 1(A) shows that initially pores are relationships between variables, when the experimental
collapsed and reached at a critical value, and further data of these variables are available. Neural networks
decrease of water content during drying causes the for- are universal function approximators that typically
mation of pores until completely dried. Opposite con- work much better in practical applications than more
ditions exist in Fig. 1(B). Fig. 2 shows that pores are traditional (polynomial) function approximations
increased or decreased as a function of water content method. A typical `neuron' takes in a set of inputs, sums
(Rahman, 2000). The glass transition theory is one of them together, takes some function of them, and passes
the proposed concepts to explain the process of the output through a weighted connection to another

Fig. 1. Change of porosity as a function of water content (with an Fig. 2. Change of porosity as a function of water content (no inversion
inversion point) (Rahman, 2000). point) (Rahman, 2000).
M.A. Hussain et al. / Journal of Food Engineering 51 (2002) 239±248 241

neuron. The neuron is thus just a predictor variable, or a technologies. Neural network is essentially an inter-
function of a nonlinear combination of predictor vari- connected assembly of simple processing elements, units
ables. The connection weights serve as adjustable pa- or nodes, whose functionality is loosely based on the
rameters, which are set by a `training method', that is, human's neuron. The processing ability of the network
they are estimated from part of the data. Numbers of is stored in the inter-unit connection strengths, or
applications have been done in the area of food prop- weights. The values of the weights are obtained by a
erties prediction. Examples are in the prediction of process of training or learning from a set of training
fruit ripening (Morimoto, Purwanto, Suzuki, Aono, & pattern (input±output relations). These weights are
Hashimoto, 1997), inactivation kinetics (Geeraerd, adapted by one or the other learning rule and represent
Versyck, & Van Impe, 1998), viscosity (Bouchard & the long-term memory of the neural network.
Grandjean, 1995), smell by electronic nose (Payne, Numerous types of the arti®cial neural network exist
1998), food quality (Ni & Gunasekaran, 1998), loaf such as multi-layer perceptron (MLP), radial basis
volume during baking (Horimoto, Durance, Nakai, & function (RBF) networks and recurrent neural networks
Lukow, 1995), sensory perception (Bomio, 1998), clas- (RNN), but each type consists of the same basic fea-
si®cation of nut (Casasent, Sipe, Schatzki, Keagy, & tures: nodes, layers and connection. The smallest ele-
Lee, 1998), and thermal conductivity (Hussain & Rah- ment of a network is the node. Every node receives
man, 1999). signal from connection, or link. The signal is then
In science and engineering always there is a tendency summed together before being applied to transfer
to generalize a process or prediction. In the literature function to produce the output. The output signals are
there are a number of attempts that has been made to then propagated to other nodes until it reaches the
develop generic models in predicting food properties. output of the network. The type of network used in this
For example, Rahman (1992) plotted thermal conduc- work is the multi-layered networks, which are described
tivity as a function of moisture for di€erent foods and in detail next.
found varied curvatures. However, when ke =‰ke0 …1 e†Š is
plotted against Xw =Xw0 , all points for di€erent foods
2.2. Multi-layer perceptron networks
condensed into one line and the correlation can be
considered generalized model for the prediction of
Multi-layer perceptron networks are one of the most
thermal conductivity as a function of moisture and po-
popular and successful neural network architectures,
rosity. Later Rahman, Chen, and Perera (1997) devel-
which are suited to a wide range of applications such as
oped a generic model to predict the thermal conductivity
prediction and process modeling. An MPL network
of fruit and vegetable as a function of temperature,
comprises a number of identical units organized in lay-
moisture content, and porosity. Rahman (1994) also
ers, with those on one layer connected to those on the
developed a generic prediction model to estimate the
next layer so that the outputs of one layer are feed-
freezing point of meat.
forward as inputs to the next layer. Note that the output
In this paper, neural network is utilized in two ways
of the nodes is only to pass nodes in one direction and
to model the pores formation during air drying. First, it
not reversed. Fig. 3 illustrates the topology of an ex-
is used to predict porosity after drying of various fruits
tremely simple, fully connected three-layer MPL net-
and vegetables such as potato, starch, apple, lentil,
work. MPL neural networks are normally trained using
carrot, pear, garlic, calamari, celery, onion and squid
a supervised training algorithm, which are mentioned
from various combinations of inputs such as the initial
later.
porosity, temperature, water content and identifying the
Di€erent types of activation functions can be utilized
types of food. In the other approach, the neural network
for the network, however the common ones, which are
in hybrid fashion with a polynomial regression model
was utilized to predict the porosity with the same com-
bination of inputs as used earlier.

2. Model development

2.1. Arti®cial neural networks

2.1.1. Overview
Neural networks have the ability to resemble (to
certain extent) human characteristic of problem solving
that are dicult to simulate using the logical, analytical Fig. 3. Fully connected three-layered MLP network for standard
techniques of expert system and standard software neural network.
242 M.A. Hussain et al. / Journal of Food Engineering 51 (2002) 239±248

sucient for most applications, are the sigmoidal and The partial known model (the regression model) pro-
hyperbolic tangent functions. The hidden layers as well vides a better starting point than standard neural net-
as the number of neurons in each layer may vary from work and at the same time, the neural network allows
one to any ®nite number depending on the user speci- for both structural and parametric uncertainty. The
®cation. It is this hidden layer structure which essentially general structure of the hybrid neural network and its
de®nes the topology of the feedforward arti®cial neural training scheme can be seen in Figs. 4 and 5, respec-
network (FANN). The topology of the network, to- tively.
gether with the neuron processing function and massive
interconnection, imparts to the FANN its powerful, 2.4. Neural network training
robust signal processing and pattern recognition capa-
bilities. Although some preliminary results are available Normally, the backpropagation method is the ®rst-
for predicting the optimal number of nonlinear hidden order gradient method used to train neural networks to
neurons, no general and mature techniques exist, which correlate between input and output variables. It has,
determine the most appropriate topology for any spe- however few drawbacks. First the rate of convergence is
ci®c application. The structure and topology are at extremely slow and the improvement per iteration falls
present still determined from past experience and by sharply with increasing iteration. Secondly, without
trial-and-error method. modi®cation such as the addition of the momentum
It is generally accepted that the performance of a term, it can often lead to local minima in its optimiza-
well-designed MLP neural network is comparable with, tion process. Other second-order-based optimization
but generally no better than obtained using good clas- techniques such as the conjugate gradient, Levenberg±
sical statistical techniques. MLP networks score over Marquardt method and Newton's method have been
classical techniques in their much reduced development found to have better convergence properties, although
time, their ability to adapt to changing situations, and at the expense of computational complexity (Hagan &
their ability to make use of related information. How- Menhaj, 1994). Many di€erent ways of improving the
ever, for problems, which have already been much learning rate can also be found in the literature (Ed-
studied by statisticians, a dramatic increase in perfor- wards & Goh, 1995). Other variants of this algorithm
mance is unlikely to be achieved by using an MLP. A
combination of both methods in hybrid fashion would
be highly advantageous in itself and is one reason for its
use in our studies.

2.3. Hybrid neural network

Neural network modeling is essentially black box in


nature. No prior knowledge about the process is re-
quired and any existing prior knowledge is ignored. The
ability of neural network to learn non-parametric or
structure-free approximation is its strength, but this is
also its weakness. Typical neural networks have many
Fig. 4. Fully connected three-layered MLP network for hybrid neural
internal parameters (the weight and biases) and these network.
could lead to over-®tting of noise as well as the under-
lying function. This results in poor generalization. These
weaknesses of neural network have encouraged recent
research to focus on combining neural network with
other component, such as ®rst principle model, fuzzy
logic and Kalman ®ltering. In this paper, a hybrid neural
network was developed, which combined a regression
polynomial-based model from given data with a non-
linear neural network (Dimitris & Lyle, 1992; Michaet &
Mark, 1994; Thompson & Kremer, 1992).
The neural network in the hybrid architecture is re-
sponsible in learning the di€erence between the default
model and the target data. Although the neural network
is a semi-parametric estimator, it has the capability of
approximating this di€erence. So, the hybrid model has
internal structures, each part performing a di€erent task. Fig. 5. Training of hybrid neural network for porosity prediction.
M.A. Hussain et al. / Journal of Food Engineering 51 (2002) 239±248 243

and di€erent methods of weight training such as works, in part, because it is di€erentiable. Tansig utilizes
chemotaxis method are also available (Morris, Monta- a hyperbolic tangent transfer function with values from
gue, & Willis, 1994). Techniques to improve the e- )1 to +1. Both the Tansig and logsig activation func-
ciency of networks by weight pruning have also been tions are utilized in the hidden layers only.
developed. In this work, Levenberg±Marquardt method
is chosen to improve the speed of convergence, during
training of the network. 3. Results and discussion

2.5. Case study: porosity prediction 3.1. Di€erent inputs and modeling methods

2.5.1. Neural network modeling strategies There are four variables considered as the inputs for
In building the multi-layered feed forward neural the modeling process. These variables are temperature
network model, the crucial step is to derive the appro- (T), water content …Xw †, initial porosity …e0 † and the type
priate con®guration so that by supplying it with the of product (F). Two approaches were used in this study.
relevant inputs, it can give an unambiguous and precise Standard Neural network (SNN) and hybrid neural
prediction of the porosity of the foodstu€ concerned. network (HNN) were used in the modeling of the pro-
There was a total of 286 data collected for this modeling cess. In this work, ®rst both temperature and water
process. The data were separated into two sets with 143 content were considered as inputs, and then initial po-
in each of them. During the network training, the two rosity was added as second combination and ®nally a
sets of data were switched between one and the other to combination of all four variables mentioned above was
avoid over-learning. The data for temperature, water utilized, as detailed in the various parts below. The
content and initial porosity were scaled to values be- porosity data during air drying of apple, potato, starch,
tween 0.1 and 0.9 to facilitate the training of the net- lentil, carrot, pear, sweet potato, garlic, calamari, squid,
work. Training was done by switching between these celery, onion were used in our studies. The porosity used
two data sets until the mean square error (MSE) here is the apparent porosity of the material. More de-
achieved for both sets in training was less than a mini- tailed terminology of porosity is given by Rahman
mum value that was set di€erently for the di€erent (1995). The air drying temperature, moisture content,
number of inputs. The number of training cycles needed initial porosity were varied from 5±70°C, 0±0.94, and
to achieve this error was normally about 50 epochs. 0±0.9, respectively.
Note that the MSE value is de®ned as the average of the
square of the di€erences between the actual output (i.e., 3.1.1. Part I ± temperature and water content as the two
experimental values) and the neural network output inputs
(i.e., predicted). This is the normal error criterion used In this part, the temperature and water content,
during training (optimising) of the neural network to ®t which are considered as two important variables, were
the data given. chosen as the inputs for the modeling process. Initially a
The above method of training the neural network polynomial-based regression model that relates the po-
using two databases ensures that it could give a good rosity with T and Xw was developed which is necessary
generalized prediction of the porosity from knowing any for the hybrid model. From least-squares based regres-
input data supplied to it and not based on the training sion method, the best polynomial- based model is given
database only. The optimum con®guration of the neural as below
network, i.e. the number of hidden nodes (with the input
and output nodes already predetermined by the dat- er ˆ 0:5Xw2 0:8Xw 0:0002T 2 ‡ 0:02T 0:15: …1†
abase size) was decided after trial and error based on the The predicted porosity values from the regression model
minimum MSE achieved during the training process. (Eq. (1)) leave some residual values. It is clearly shown
The same approach in the modelling process is followed in Fig. 6. The residuals from the actual porosity values
when dealing with the hybrid method of combining are given by
neural network with the regression model.
dˆe er : …2†
Three types of transfer functions for training the
network were used in this study. The activation func- The estimation of error enables us to use neural network
tions are the purelin, tansig and logsig. Purelin is a linear to map the presence of these residuals in a hybrid ge-
transfer function. Neurons of this type are used as linear neric model strategy, as has been discussed previously.
approximators. It is usually used in the output layer. However modeling the porosity data by only using
Logsig is one of the sigmoid transfer functions that takes standard neural networks is also done as comparison
input value between plus and minus in®nity and then studies. During modeling with the neural networks, the
provides the output into the range 0±1. This transfer inputs for the hybrid neural network were the same as
function is commonly used in backpropagation net- the standard neural network i.e. temperature (T) and
244 M.A. Hussain et al. / Journal of Food Engineering 51 (2002) 239±248

Fig. 6. Regression model with two inputs (T and Xw ). Fig. 8. Result for hybrid neural network with two inputs (T and Xw ).

water content (Xw ). The supervised training method used the polynomial-based model has been modi®ed, the re-
requires target values. The target values for the two siduals of the ®nal porosity have to be recalculated using
types of network i.e. standard and hybrid are di€erent, Eq. (2). In this part, the inputs for training the model
as can be seen from Figs. 3±5, respectively. For the were T ; Xw and e0 while the target value for the standard
standard neural network, the target value is the experi- neural network is the experimental porosity (e) and the
mental porosity (e), but for the neural network in the target value for hybrid neural network is the porosity
hybrid scheme, the target value is the residual porosity residuals (d). The method of training these networks is
(d). The method of training these networks in terms of the same as in Part I. The results of their modeling can
switching from one data set to another was the same. be seen in Figs. 10 and 11, respectively, and the dis-
The results of these modelling using the SNN and HNN cussion will be seen in the next section.
are shown in Figs. 7 and 8, respectively, and are dis-
cussed in detail in the next section. 3.1.3. Part III ± temperature, water content, initial
porosity and type of product as inputs
The various data collected from the individual ex-
3.1.2. Part II ± temperature, water content and initial
periments originate from di€erent types of products. To
porosity as the three inputs
investigate whether the type of product can in¯uence the
To improve the network training, the initial porosity
value of porosity, this input was added together with the
was added as another extra input in the model. The
three other inputs as in the previous trial. As the type of
polynomial based model obtained in this case is
product is not in numerical form, the products were
er ˆ 0:5Xw2 0:8Xw 0:002T 2 ‡ 0:02T 0:15…1 e0 †: assigned with numerical values. A number was given to
each type of product for example: apple, carrot and pear
…3†
(sugar-based products) are represented by number `1',
The values of porosity from Eq. (3) were once again potato, starch and onion (starch-based product) by
compared to the actual value and shown in Fig. 9. As number `2' and lentil, garlic, calamari and celery (other

Fig. 7. Result for standard neural network with two inputs (T and Xw ). Fig. 9. Regression model with three inputs (T ; Xw and e0 ).
M.A. Hussain et al. / Journal of Food Engineering 51 (2002) 239±248 245

Fig. 10. Result for standard neural network with three inputs (T ; Xw Fig. 12. Regression model with four inputs (T ; Xw ; e0 and F).
and e0 ).

Fig. 13. Result for standard neural network with four inputs
Fig. 11. Result for hybrid neural network with three inputs (T ; Xw and (T ; Xw ; e0 and F).
e0 ).

products) by number `3'. To relate the type of product


into the model of the system, the following regression
equation is obtained:
er ˆ 0:5Xw2 0:8Xw 0:002T 2 ‡ 0:02T 0:05…1 e0 †F ;
…4†
where F is a numerical value given to the type of product
used. The residuals resulted are shown in Fig. 12. A
small improvement was found as compared to the ®rst
two regression models. By using this regression model,
value of the residual porosity (d) is recalculated again
from Eq. (2). Then the standard neural network and
hybrid neural network were trained by the same way
mention in Parts I and II but with di€erent inputs to the Fig. 14. Result for neural network with four inputs (T ; Xw ; e0 and F).
network i.e. T ; Xw ; e0 and F. The results of these
modelling using the SNN and HNN can be seen in Figs.
13 and 14, respectively. Figs. 6±14 visually indicate the It was found that di€erent combinations of the in-
variation of experimental data set and predicted values puts, number of nodes and the activation functions gave
from the models, thus provide visual presentation of di€erent results to the modelling process using the
accuracy. These ®gures also show the complexity of neural network methodology. The value of MSE and
experimental data set of pore formation during drying. absolute mean percentage error (MPE) were used to
246 M.A. Hussain et al. / Journal of Food Engineering 51 (2002) 239±248

Table 1
Prediction accuracy of the models developed from the data set
Modela Nb TSc MSEd MPEe; f
Hidden layer Output layer
2
Regression model I (2) ± ± ± 3:19  10 33.57 (4.11)
2
Standard neural network (2) 12 Tansig Logsig 1:00  10 75.70 (2.55)
3
Hybrid neural network (2) 12 Logsig Tansig 9:80  10 56.50 (2.01)
2
Regression model II (3) ± ± ± 2:64  10 23.69 (4.13)
4
Standard neural network (3) 15 Logsig Tansig 7:44  10 6.52 (0.67)
4
Hybrid neural network (3) 15 Logsin Logsig 4:07  10 0.98 (0.55)
2
Regression model III (4) ± ± ± 2:34  10 14.15 (4.21)
4
Standard neural network (4) 13 Logsig Logsig 4:52  10 6.42 (0.34)
4
Hybrid neural network (4) 13 Logsig Tansig 2:68  10 0.58 (0.39)
a
Number in parentheses indicates the number of input.
b
Number of optimum nodes.
c
Transfer function.
P
d
Mean square errorˆ …1=n† ‰…e em †2 Š.
e P
Absolute mean percent errorˆ …1=n† ‰100  …je em j†Š=e; em ˆ porosity of dried sample at di€erent moisture contents from model.
f
Values in parentheses indicate the standard deviation.

compare the capability of each type of network as well In the case of four inputs with the addition of product
as for the regression model. The MSE and MPE were type the prediction accuracy for the neural network
calculated as: hybrid model improved further from 0.98% to 0.58%
(Table 1). This was also far better than utilizing the
1X n
standard neural network model or the regression model.
MSE ˆ …e em †2 ;
n iˆ1 As discussed earlier the classes of product types con-
sidered are sugar-, starch-based, and other product
1X n
j…e em †j category. The prediction from the neural network
MPE ˆ  100: models with four inputs is far better than the two inputs
n iˆ1 e
and slightly better than the three inputs. The additional
Table 1 shows the MSE and MPE for the training data inputs (the type of product) improve the performance of
with all protocols as well as the standard deviation for neural network models signi®cantly. The further ad-
the mean error in the brackets. The number of nodes vantages of using four inputs were that it needed fewer
required and the transfer function used were also in- nodes than the three inputs case and gave more accurate
cluded in Table 1. In case of two inputs (temperature simulation results. The grouping of the product type is
and moisture content) the MPE errors for standard not an easy task, but here an innovative attempt was
neural network and hybrid neural network were 75.7% made in the modeling based on grouping the products.
and 56.5%, respectively. The hybrid neural network A well-de®ned class of products could be done when
models are far better than the standard neural network more data are available in the literature. The purpose of
models (Table 1). However, the overall result was still classi®cation was only to show that including product
poor for either SNN or HNN and not any better than type (i.e., e€ect of other components) could also help in
the regression model. The outputs from the network developing the generic model and improving the accu-
could not map to the actual values of the porosity; thus racy in predicting the porosity of these food materials.
using only the two inputs, temperature and water con-
tent was not enough to create a generic model for esti-
mating the porosity of these food materials. 4. Conclusion
In the case of three inputs (temperature, moisture
content, and initial porosity) the prediction accuracy In this work, three di€erent cases involving di€erent
increased from 75.7% to 6.5% in case of the standard combinations of inputs to estimate the porosity by
neural network models while utilizing the hybrid neural neural network modeling were studied. The ®rst involves
network models increased the accuracy further from two inputs, which are the temperature and water con-
56.5 % to 0.98 %. Both these results were far better than tent. From the results obtained, it was found that these
the results of the regression model only, i.e. 23.7% but two inputs for the models are not enough to produce
with the hybrid neural network model giving the best satisfactory prediction accuracy. When one more input
results. This indicated that initial porosity is one of the (initial porosity, e0 ) was added to the network, the pre-
important parameters to be considered for developing a diction error was improved dramatically from 56.5% to
generic model. 0.98% for the MPE result of hybrid neural network.
M.A. Hussain et al. / Journal of Food Engineering 51 (2002) 239±248 247

When four inputs as temperature, water content, initial Hutchinson, R. J., Siodlak, G. D. E., & Smith, A. C. (1987). In¯uence
porosity, and the type of product were used in the net- of processing variables on the mechanical properties of extruded
maize. Journal of Materials Science, 22, 3956±3962.
work modeling, the best performance was achieved with Madamba, P. S., Driscoll, R. H., & Buckle, K. A. (1994). Bulk density
prediction accuracy of 0.58%. It can be concluded that porosity and resistance to air¯ow of garlic slices. Drying Technol-
temperature, water content, initial porosity and the type ogy, 12(4), 937±954.
of product are important parameters to be considered in Mattea, M., Urbicain, M. J., & Rotstein, E. (1986). Prediction of
obtaining the best generic model for the estimation of thermal conductivity of vegetables foods by the e€ective medium
theory. Journal of Food Science, 51(1), 114±115, and 134.
porosity. The hybrid models that consist of the poly- McCulloch, W.S. & Pitts, W. (1943). A logical calculus of the ideas
nomial-based model plus the neural network model are immanent in nervous activity. In S. W. McCulloch (Ed.), Embod-
also more accurate than using the standard neural net- iments of mind (pp. 19±39).
work or the regression model alone in predicting the McMinn, W. A. M. & Magee, T. R. A. (1998). Density, shrinkage and
porosity of various foods during air drying. The hybrid porosity of starch gel during convective drying. In Proceedings of
the 11th International Drying Symposium IDS '98, Halkidiki.
neural network is well supplemented by the polynomial- Michael, L. T., & Mark, A. K. (1994). Modeling chemical processes
based model in modeling, which helps to obtain an ac- using prior knowledge and neural networks. AIChE Journal, 40(8),
curate generic model. 1328±1340.
Morimoto, T., Purwanto, W., Suzuki, J., Aono, T., & Hashimoto, Y.
(1997). Identi®cation of heat treatment e€ect on fruit ripening and
its optimization. In A. Munack, & H. J. Tantau (Eds.), Mathe-
References matical and control applications in agriculture and horticulture (pp.
267±272). Oxford: Pergamon Press.
Bhatnagar, S., & Hanna, M. A. (1997). Modi®cation of microstructure Morris, A. J., Montague, G. A., & Willis, M. J. (1994). Arti®cial neural
of starch extruded with selected lipids. Starch, 49, 12±20. networks: studies in process modelling and controls. Trans IchemE
Bomio, M. (1998). Neural networks and the future of sensory Part A, 72, 3±19.
evaluation. Food Technology, 52(8), 62±63. Ni, H., & Gunasekaran, S. (1998). Food quality prediction with neural
Bouchard, C., & Grandjean, P. A. (1995). A neural network networks. Food Technology, 52(10), 60±65.
correlation for the variation of viscosity of sucrose agents solutions Payne, J. S. (1998). Electronic nose technology: an overview of current
with temperature and concentration. Food Science and Technology, technology and commercial availability. Food Science and Tech-
28, 157±159. nology Today, 12(4), 196±200.
Casasent, D. A., Sipe, M. A., Schatzki, T. F., Keagy, P. M., & Lee, L. Palipane, K. B., Driscoll, R. H., & Srzednicki, G. (1992). Density,
C. (1998). Neural net classi®cation of X-ray pistachio nut data. porosity and composition of macadamia in-shell nuts. Food
Food Science and Technology, 31, 122±128. Australia, 44(6), 276±280.
Dimitris, C. P., & Lyle, H. U. (1992). A hybrid neural network ± ®rst Perez, M. G. R., & Calvelo, A. (1984). Modeling the thermal
principle approach to process modeling. AIChE Journal, 38(10), conductivity of cooked meat. Journal of Food Science, 49, 152±156.
1499±1510. Pietsch, W. (1999). Readily engineering agglomerates with special
Edwards, N. J., & Goh, C. J. (1995). Direct training method for a properties from micro-and nanosized particles. Chemical Engineer-
continuous-time nonlinear optimal feedback controller. Journal of ing Progress, 95(8), 67±81.
Optimisation Theory and Applications, 84(3), 509±528. Rahman, M. S. (1992). Thermal conductivity of four food materials as
Geeraerd, A. H., Versyck, K. J., & Van Impe, J. F. (1998). Optimal a single function of porosity and water content. Journal of Food
experimental design for parameter estimation of microbial inactiva- Engineering, 15, 261±268.
tion kinetics. Technical Report BioTeC 1998-06. Department of Rahman, M. S. (1994). The accuracy of prediction of the freezing point
Food and Microbial Technology, Kathalieke University, Leuven. of meat from general models. Journal of Food Engineering, 21, 127±
Gogoi, B. K., Alavi, S. H., & Rizvi, S. S. H. (2000). Mechanical 136.
properties of protein-stabilized starch-based supercritical ¯uid Rahman, M. S. (1995). Food properties handbook. Boca Raton, FL:
extrudates. International Journal of Food Properties, 3(1), 37±58. CRC Press.
Hagan, M. T., & Menhaj, M. (1994). Training feedforward networks Rahman, M. S., Chen, X. D., & Perera, C. O. (1997). An improved
with the Marguardt algorithm. IEEE Transactions on Neural thermal conductivity prediction model for fruits and vegetables as a
Network, 5(6), 989±993. function of temperature, water content and porosity. Journal of
Horimoto, Y., Durance, T., Nakai, S., & Lukow, O. M. (1995). Neural Food Engineering, 31, 163±170.
networks vs principal component regression for prediction of Rahman, M. S. (2000). Mechanism of pore formation in foods during
wheat ¯our loaf volume in baking tests. Journal of Food Science, drying: present status. In Proceedings of the Eighth International
60(3), 429±433. Congress on Engineering and Food ICEF-8, Peubla.
Huang, C. T., & Clayton, J. T. (1990). Relationships between Rahman, M. S. (2001). Towards prediction of porosity in foods during
mechanical properties and microstructure of porous foods: Part drying: a brief review. Drying Technology, 19(1), 3±15.
I. A review. In W. E. L. Spiess, & H. Schubert (Eds.), Engineering Rahman, M. S., Perera, C. O., Chen, X. D., Driscoll, R. H., & Potluri,
and food, vol. 1. Physical properties and process control (pp. 352± P. L. (1996). Density, shrinkage and porosity of calamari mantle
360). London: Elsevier. meat during air drying in a cabinet dryer as a function of water
Hussain, M. A., & Rahman, M. S. (1999). Thermal conductivity content. Journal of Food Engineering, 30, 135±145.
prediction of fruits and vegetables using neural networks. Interna- Rapusas, R. S., & Driscoll, R. H. (1995). Thermophysical properties of
tional Journal of Food Properties, 2(2), 121±137. fresh and dried white onion slices. Journal of Food Engineering, 24,
Hussain, M. A., & Kershenbaum, L. S. (2000). Implementation of an 149±164.
inverse-model-based control strategy using neural networks on a Rumelhart, E. D., & McClelland, J. L. (1986). Parallel distributed
partially simulated exothermic reactor. Institution of Chemical processing: explorations in the microstructure of cognition (Chapter
Engineers Trans I Chem E, 78, 299±311. 8). Cambridge, MA: MIT Press.
248 M.A. Hussain et al. / Journal of Food Engineering 51 (2002) 239±248

Scanlon, M. G., Day, A. J., & Povey, M. J. W. (1998). Shear sti€ness Vincent, J. F. V. (1989). Relationship between density and sti€ness of
and density in potato parenchyma. International Journal of Food apple ¯esh. Journal of the Science of Food and Agriculture, 47, 443±
Science and Technology, 33, 461±464. 463.
Senadeera, W., Bhandari, B. R., Young, G., & Wijesinghe, B. (1998). Wildegger-Gaissmaier, A. E., Agarwal, P. K. (1990). In A. S.
Change of physical properties of green beans during drying and its Mujumdar, & M. A. Roques (Eds.), Shrinkage during drying and
in¯uence on ¯uidization. In Proceedings of the 11th International devolatillization of wet coal. New York: Hemisphere.
Drying Symposium IDS '98, Halkidiki.
Thompson, M. L., & Kremer, M. A. (1992). Modeling chemical
processes using prior knowledge and neural networks. AIChE
Journal, 38(1), 41±54.

You might also like