Advances in Bioscience and Bioengineering

2014; 2(2): 14-24

Published online July 30, 2014 (http://www.sciencepublishinggroup.com/j/abb)
doi: 10.11648/j.abb.20140202.12
ISSN: 2330-4154 (Print); ISSN: 2330-4162 (Online)

Thermal conductivity of food products using:

A correlation analysis based on artificial neural
networks (ANNs)
Ajasa, Abiodun Afis*, Adenowo, Adetokunbo Abayomi, Ogunlewe, Adeyinka Oluremi,
Folorunso, Comfort Oluseyi
Department of Electronic and Computer Engineering, Faculty of Engineering, Lagos State University, Epe, Lagos, Nigeria

Email address:
ajasaaf@yahoo.com (Ajasa A. A.), adetokyom@yahoo.com (Adenowo A. A.), yogunlewe@yahoo.com (Ogunlewe A. O.),
comfortobolajoko@yahoo.com (Folorunso C. O.)

To cite this article:

Ajasa, Abiodun Afis, Adenowo, Adetokunbo Abayomi, Ogunlewe, Adeyinka Oluremi, Folorunso, Comfort Oluseyi. Thermal
Conductivity of Food Products using: A Correlation Analysis Based on Artificial Neural Networks (ANNs). Advances in Bioscience and
Bioengineering. Vol. 2, No. 2, 2014, pp. 14-24. doi: 10.11648/j.abb.20140202.12

Abstract: This paper presents the correlation between the predicted and desired/targeted thermal conductivity of food
products as a function of moisture content, temperature and apparent density. The food products considered in this work are
the bakery products which include bread, bread dough, cake, and whole-wheat dough. Statistical data of results from
previous work in existing literatures were used in this work for a wide range of moisture contents, temperatures and
apparent densities resulting from baking conditions. The results of this work showed straight line curves when the predicted
values of thermal conductivity were plotted against the targeted values of thermal conductivity. This demonstrates
correlation between the predicted and targeted thermal conductivities when the points are joined together (best fit-points),
hence, a very good agreement between the predicted and the desired values of thermal conductivity. The two ANN models
that were finally selected, after several configurations had been considered and evaluated, are the optimal ANN model that
was found to be a network with two hidden layers and eight neurons and the simplest ANN model was equally found to be
a network with one hidden layer and ten neurons. The estimated errors between the predicted and desired (or targeted)
thermal conductivity values of the bakery products for both the optimal ANN and simplest ANN models are the MRE,
MAE and SE. Moreover, the results also showed that the optimal ANN model had an MRE of 0.04878%, an MAE of
0.0054W/mK and an SE of 0.0015W/mK while the simplest ANN model was estimated to have an MRE of 0.03388%, an
MAE of 0.0034W/mK and an SE of 0.0011W/mK. These errors are approximately equal to zero (i.e., 0 W/mK) and could,
therefore, be regarded as a good result for the prediction. Since the simplest ANN model had the least values of all three
errors (MRE, MAE and SE) when compared with other configurations, including the optimal ANN model, it is, however,
regarded as the best ANN model and is, thus, recommended.
Keywords: Thermo-Physical Properties of Biological Products, Thermal Conductivity of Bakery Products,
Back-Propagation, Artificial Neural Network, Mean Absolute Error, Mean Relative Error, Standard Error

charcoal, or microwaves are used for generating the

1. Introduction required heat.
In a typical baking process, bakery products undergo The thermal conductivity of a material can be defined as
physical, chemical and biochemical changes that the ability of a substance to conduct heat. It is a property of
cumulatively result in expansion of bulk volume, materials that expresses the heat flux, F (W/m2) that will
evaporation of water, formation of a porous structure, flow through the material if a certain temperature gradient
denaturation of protein, gelatinization of starch, formation ∆T (K/m) exists over the material. Thermal conductivity of
of crust and browning reactions respectively. During such materials can be defined as the heat flow per unit area per
processes, ovens powered by gas, electricity, firewood, unit time when the temperature decreases by one degree in
unit distance.
15 Ajasa, Abiodun Afis et al.: Thermal Conductivity of Food Products using: A Correlation Analysis Based
on Artificial Neural Networks (ANNs)

Accordingly, heat is transferred mainly by convection ANN modelling differently for the prediction of food
from the heating media, and by radiation from oven walls quality. Recently, Sablani and Shayya in 2001 applied ANN
to the product surface and then by conduction to the modelling for the prediction of heat penetration parameters
geometric centre. At the same time, moisture diffuses in Stumbo’s method of thermal process calculations [20].
outward to the product surface [1, 5]. Rahman’s model [4] (data considered only above 0°C)
The problem of interest in the design of bakery ovens is predicted thermal conductivity with mean relative errors of
concerned with the promotion of the required rate of heat 24.3 and 81.6%, respectively. This model was able to
transfer with the minimum possible surface area and predict thermal conductivity with a mean relative error of
temperature difference. And from the engineer’s point of 12.6% and a mean absolute error of 0.081 W/mK. The
view, it is usually sufficient to know the total quantities of model can be incorporated in heat transfer calculations
energy emitted and absorbed by the material at various during food processing where moisture, temperature and
temperatures [16]. As such, it is frequently necessary to apparent porosity dependent thermal conductivity values
establish the rate at which heat will be conducted through a are required.
solid if a known temperature difference exists across the Shyam et al. (2002) optimal ANN model was found to be
solid. For such purposes, and especially if the process a network with 6 neurons in each of the 2-hidden layers.
varies with time, sophisticated mathematical techniques are This optimal model was capable of predicting the thermal
required to establish this, the phenomenon being known as conductivity values of various bakery products (such as
transient-heat conduction. A knowledge of the product bread, bread dough, French bread, yellow cake, tortilla chip,
properties, including thermal conductivity as a function of whole wheat dough, baked chapatti and cup cake) for a
processing conditions is needed in order to predict the wide range of conditions with a mean relative error of 10%,
temperature and water distribution in the product during a mean absolute error of less than 0.02 W/m K and a
baking [5, 8]. The temperature and moisture distribution standard error of about 0.003 W/m K. The simplest ANN
within the porous product can be predicted using diffusion model, which had 1-hidden layer and 2 neurons, predicted
equations of heat and water. thermal conductivity values with a mean relative error of
less than 15% [3]. All these work were successfully carried
2. Literature Review out with satisfactory results obtained using ANN modelling.
In predicting thermal properties of a material at desired
Neural networks have been trained to perform complex conditions, several modelling approaches have been
functions in various fields including pattern recognition, proposed and none of them was found suitable for use over
identification, classification, speech, vision, and control a wide range of foods. According to Murakami and Okos
systems. Today, neural networks can be trained to solve (1989) the most promising approach is based on chemical
problems that are difficult for conventional computers or composition, temperature and physical characteristics [12].
human beings. Commonly, neural networks are adjusted (or More recently, Baik et al. in 2001 reviewed common and
trained) so that a particular input leads to a specific target new measurement techniques, prediction models and
output. Typically, many such input/target pairs are needed published data on thermo-physical properties of bakery
to train a neural network. products [2].
Linko and Zhu also stated that of all the various The series model of specific heat, density and thermal
modelling approaches of predicting the thermal diffusivity has been successfully applied to many food
conductivity of a wide range of foods, including bakery materials including porous materials such as baking
products, the neural network-based models have proven to products. However, for the prediction of thermal
be excellent. Amongst the major benefits of using ANN are conductivity of porous food, there is still some theoretical
excellent management of uncertainties, noisy data and non- argument for the use of the structural models [12].
linear relationships. Neural network modelling has Murakami and Okos (1989) evaluated nine different
generated increasing acceptance and is an interesting structural models with specific types of porous foods and
method in the estimation, prediction and control of found that parallel and perpendicular models showed 12–97%
bioprocesses [9]. and 18–61% standard errors respectively.
Review of past work showed that Ruan et al. applied Among the models, Keey’s model was found to be the
ANN modelling to predict the rheological properties of best prediction model for porous grains and powders. The
dough in 1995 [15]. Fang et al. also applied the ANN model produced standard errors of <28% for full fat dry
modelling to predict the physical properties of ground milk and <10% for other food materials. In addition, all
wheat in 1998 [18] while Hussain and Rahman, in 1999, structural models neglect interactions between components,
predicted the thermal conductivity of fruits and vegetables phase transition and distillation heat transfer, which may be
with the application of ANN modelling [6]. Similarly, ANN significant in the baking process [1]. Hence, most thermal
modelling was applied by Myhara et al. in 1998 for the conductivity models reported are usually empirical rather
prediction of isotherms of dates [11], Ni and Gunasekaran than theoretical.
in 1998 [7] and Xie and Xiong in 1999 [13] also applied
 Advances in Bioscience and Bioengineering 2014; 2(2): 14-24 16

3. Methodology have an n influence on the characteristics or technological

aspect of the dough.
In this section, the various materials and methods used Flour
are discussed here. Flour account for the main portion of the raw materials
3.1. Materials involved in baked product production and the thermal
conductivity of flour need to be considered while applying
Basically, baked products made by bakeries can be the neural network. Flour used is mainly those extracted
divided into three (3) different groups: from two basic cereals – wheat and rye. Flour from other
• Bread loaves (e.g. Wheat bread, rye bread, specialty grains which do not contain gluten-forming proteins is
bread) usually blended with wheat flour for the production of
• Small baked products (e.g. rolls, buns, croissants) bakery products. The quality of flour is basically dependent
• Fine baked goods (e.g. plaits, apple strudel, biscuits, on its intended use. The flour quality depends on the
salts pretzels) following factors:
Depending on the specific type of bakery product to be • Wheat variety
produced and associated production process, a wide variety • Growing conditions
of raw materials are available to the baker. In addition to • Grain storage
the basic Raw materials (flour, water, salt and yeast) • Flour production technique
various other ingredients can be used. The ingredients used • Flour storage
Table 1. Example of a requirement profile for various baking flours

Patent gluten % High-Ash whole-Grain

Flour Flour Flour Flour
Moisture % 13.0-15.0 13.0-15.0 13.0-15.0 11.0-13.0
Ash % DM 0.38-0.60 0.64-0.78 1.05-1.15 1.75-1.95
Protein % 12.0-14.0 13.5-15.0 14.0-15.5 13.5-15.0
Wet gluten % 28.0-33.0 31.0-35.0 32.0-36.0 29.0-33.0
Falling N umber sec 320-410 300-390 280-380 300-380
Sedimentation ml 38-45 38-43 25-30 NIL
Water Absorption % 60-64 61-65 65-70 66-71
Weakening FU 20-70 60-90 60-90 60-90
Dough energy cm2 90-130 80-110 55-85 60-90
Max. viscosity AU 500-1000 350-800 300-550 250-500

Average patent flour (first grade) is made up of the Thermal properties (thermal conductivity and diffusivity)
following: of gluten and glutenin were measured in the temperature
• Carbohydrates: 73.5%. This includes starch: 71%, range 60-175°typically used in extrusion processing.
soluble sugars 2.4% and cellulose 0.1%. Thermal conductivity and diffusivity of gluten decreased
• Proteins: 11.0 % :This includes gluten-forming with increasing temperature and increased with increasing
proteins 10% & water soluble proteins 1% moisture content. Thermal conductivity and diffusivity of
• Water:14% glutenin increased with temperature and moisture content.
• Fat: 1.0% Thermal conductivity of gluten was 0.06-0.35 W/m-C and
• Minerals: 0.5 % glutenin was 0.29-0.49 W/m-C for the temperature range
In terms of quantity, carbohydrates account for the 60-175°and moisture content range of 0-30%.
greatest portion in flour which incidentally forms the
greater part of bakery products. Starch essentially fulfills 3.2. Measurement of Thermal Conductivity
the following functions: (1) a source of nutrient for yeast Thermal conductivity of materials can be measured and
after enzymatic degradation, (2) absorption of free dough obtained through different materials. Depending on the
water during gelatinization, (3) contribution to crust, crumb measurement time these methods can be classified as either;
and coloration formation. • Steady State or
The content of soluble sugar substances in wheat flour • Transient or Unsteady
amounts to approximately 1.5 – 3%. The main soluble These two methods are explained below:
sugar substances are glucose, maltose and dextrin. They are A. Steady State Method
dissolved during dough production in the available dough In the steady state methods, the measurement time can be
liquid. Glucose and Maltose are available as yeast food much longer for poor thermally conductive samples than in
while Dextrin cannot be fermented by yeasts the transient methods. During testing, moisture migration
17 Ajasa, Abiodun Afis et al.: Thermal Conductivity of Food Products using: A Correlation Analysis Based
on Artificial Neural Networks (ANNs)

and property changes can occur due to long exposure to the capped column test device was operated horizontally.
high temperatures. Thus, this method may not be suitable Temperature distribution along the height of the test
for measurement of the thermal conductivity of bakery sample water stream was measured. The moisture content
products. The steady state methods include: distribution of test sample was determined by measuring
• Guarded hot plate moisture content of the sample cut into several section of
• Differential scanning calorimeter (DSC) attachment equal height. The k value can be determined by applying a
method heat and mass transfer equation to the experimental
• Capped column test device temperature and moisture gradient data. The measurement
(i) Guarded hot plate Method is fast (within several minutes), and there is no worry about
This method is a representative steady state method. A experimental deviation because of the moisture gradient,
time- independent heat flow is generated through the because the measurement device and data analysis are
sample between a heat source and a heat sink. This method designed to evaluate thermal and mass transfer properties
is mathematically simple to process, and it is easy to simultaneously.
control experimental conditions. The k values can be B. Transient Method
calculated as a mean value measures over the temperature The transient techniques are most suitable for bakery
interval used in the experiment using Fourier’s heat products because testing is very fast and it yields more
conduction equation. Experimental uncertainty in the accurate results. In the transient method, the sample is
measured thermal conductivity of biscuit dough was ≤ 7.4% subjected to a time-dependent heat flow and the
using a similar method, single-plate method [22]. This temperature is measured at one or more points within the
evaluation was based on combination of instrumentation sample or at its surface. The transient technique includes:
errors, geometrical uncertainties, and deviations from the • Line heat source probe method
assumed one-dimensional nature of heat flow. • Temperature history
(ii) DSC Attachment Method • Transient hot strip method (THS)
In order to alleviate the disadvantage of conventional (i) Line Heat Source Probe Method
steady state method, an attachment to a DSC was This method is recommended for most food applications,
constructed by Buhri and Singh [23]. The main innovations including bakery products. The method is simple and fast
are relatively rapid measurement (10 to 15 min), small (measuring times are between 3 and 600s) and requires
sample size, and no risk of drying of the samples. The relatively small samples, but it does require a fairly
thermocouple probe was inserted into the sample, and the sophisticated data acquisition system. The probe (0.66mm)
DSC heating pan temperature was kept at 40oC. After 5 consists of a constantan heater wire and chromel-
minutes the sample initial temperature was recorded. The constantan thermocouple wire.
pan temperature was then immediately increased by 10oC. A line heat source probe is embedded in the sample
After 10 to 15 minutes, a new steady state existed, and the (regarded as an infinite body), which is initially at a
final sample temperature was recorded. The k was obtained uniform temperature, resulting in a cylindrical temperature
using the following equation, which is based on Fourier’s distribution. Heating and temperature monitoring occur
heat conduction equation: simultaneously. The rate of temperature rise of the heater is
∆
directly related to the sample’s conductivity. The slope of
= (1) the linear portion of each data set was used to determine
(∆ ∆ )
effective thermal conductivity by equation (2)
where k = thermal conductivity, L = sample length,
∆Q = difference of energy required to maintain pan
= (2)
temperature, ( )
A = sample area perpendicular to heat flow,
where k = thermal conductivity of the sample,
∆t2 = final temperature difference between DSC heating
τ1 = initial time when probe heater was energized,
pan and sample
τ2 = final time since probe heater was energized,
∆t1 = Initial temperature difference between DSC heating
t1 = temperature of probe thermocouple at t1,
pan and sample
t2 = temperature of probe thermocouple at t2,
(iii) Capped column test device
Q = heat flux generated by probe heater.
The use of a capped column test device was introduced
In order to obtain correct results and avoid edge effects
by Zhou et al [24]. Experiments with the capped column
in a line heat source probe method, the following
test device were conducted with constant heat flux and no
constraints [25] should be considered:
net water loss. Constant heat flux was provided by
circulating hot cold water at constant temperatures at the Probe size: < 2/5 √( ) (3)
two ends of the cylindrical test sample (diameter 3cm,
height 5cm, and 2.5cm). The test sample was enclosed in where rp = probe radius, = thermal diffusivity of the
polystyrene foam to minimize heat loss to the surroundings. sample, ! = time, test duration
To eliminate the gravity-induced moisture migration effect,
 Advances in Bioscience and Bioengineering 2014; 2(2): 14-24 18

Size of sample: " > 2.6√ ! (4)

heat flux
(free convection)
where " = sample radius heat & moisture
Other experimental errors can be produced by inserting a
bread
probe into unstable structure because this may rupture the dough
structure close to the probe, giving rise to false values.
Thus, the use of a linear movement probe header is strongly
bread/
recommended for thermal conductivity measurement using baking pan
this method.
(ii) Temperature History
heat source
This method is suitable for thermal conductivity of
Fig. 2. Conceptual representation of Oven dynamics during typical Baking
porous, fine particles grains or beans. By measuring the
process
temperature profile in a product subject to a known energy
flow, k and can be calculated (k = thermal conductivity From Fourier’s law of heat conduction, the rate at which
and = thermal diffusivity). This method requires rather such heat is conducted through a body per unit cross-
long measuring times, from 10 to 200 minutes. For bakery sectional area is said to be proportional to the negative of
products, this method could be applied to bread dough and the temperature gradient existing in the body [16]. In other
tortilla. words,
(iii) Transient Hot Strip Method (THS)
A thin foil of platinum or nickel serves both a heat Q ∝ − ∇T (5)
source and a temperature sensor. A constant current is
supplied to the metal foil and the voltage increase is The proportionality factor is called the thermal
monitored over a short period; typically 5 seconds. The conductivity of the material. By definition, it is the ability
voltage change is due to the temperature increase, normally of the material to conduct heat and thus, a measure of the
less than 1℃, which causes an increase in the electrical rate at which heat flows through a material between points
resistance of the metal strip. The thermal properties of the at different temperatures, measured in watts per meter per
surrounding material determine the temperature change in degree.
the material foil. The THS method has been used mostly for As a property, the thermal conductivity expresses the
building materials and liquids [26], but it was also applied heat flux, Q (W/m2) that will flow through the material if a
measuring the thermal conductivity of potatoes, meat and certain temperature gradient, ∆T (K/m) exists across the
dough [27]. The time needed for one measurement often material. That is, it is the heat flow per unit area per unit
has to equilibrate before the test can start. time when the temperature decreases by one degree in unit
distance. Thus,
3.3. Methods
∆T
Q = − kA (6)
(i) Mathematical Modelling for Thermal Conductivity of ∆x
Bakery Products
or
Although the exact mechanism of heat conduction in
solids is not entirely understood, it is believed, however, to TH − TL
Q = − kA (7)
be partially due to the motion of free electrons in the solid l
matter, which transport energy if a temperature difference is
applied (Refer to Fig.1) and the conceptual representation where Q = heat flux, k = thermal conductivity, A = cross-
of Oven dynamics during typical baking process as sectional area, TH = temperature at hot end, TL =
depicted in figure 2 below. temperature at cold end, and l = thickness of material
respectively, and the negative sign indicates that the heat
heat & moisture flow is positive in the direction of temperature fall.
(ii) Training the Ann Model Using Back Propagation
TL
(low) Algorithm
apparent temperature Neural networks have been trained to perform complex
thickness gradient, d T functions in various fields including pattern recognition,
l
(high) identification, classification, speech, vision, and control
TH
systems. Today, neural networks can be trained to solve
problems that are difficult for conventional computers or
human beings. Commonly, neural networks are adjusted (or
heat from source trained), so that a particular input leads to a specific target
Fig. 1. Conceptualisation of thermal conductivity of a bakery product output. The schematic of such a situation is shown in Fig. 3
where the network is adjusted based on a comparison of the
output and the target until the output matches the target.
19 Ajasa, Abiodun Afis et al.: Thermal Conductivity of Food Products using: A Correlation Analysis Based
on Artificial Neural Networks (ANNs)

Typically, many such input/target pairs are needed to train a compared. The mathematical representation of these error
neural network. parameters are defined in equations 8 to 10 as follow [10, 14]:
The back-propagation algorithm was utilized in model
1
∑ ( K D − K P )
n
training. A hyperbolic-tangent transfer function was also MAE = (8)
n  i =1 
used in all cases. Properly trained back-propagation networks
(Fig. 4) tend to give reasonable answers when presented with 1
∑ ( KD − KP ) / KD 
n
MRE = (9)
unfamiliar inputs that have never been seen earlier. Typically, n  i =1 
a new input leads to an output similar to the correct output
for input vectors used in training that are similar to those  2 
∑
n
 (K D − K P ) 
being presented. This generalization properly makes it i =1
SE =  

(10)
possible to train a network on a representative set of n −1
 
input/target pairs and get good results without training the  
network on all possible input/output pairs.
The back-propagation algorithm uses the supervised where n is the number of data points, KD and KP are the
training technique where the network weights and biases desired and predicted values of thermal conductivity
are initialized randomly at the beginning of the training respectively.
phase. For a given set of inputs to the network, the response The optimum configuration of the network was chosen
to each neuron in the output layer is calculated and by selecting the lower value from the different
compared with the corresponding desired output response. configuration of the network. It was evidently based upon
The errors associated with desired output response are minimizing the difference between the neural network
adjusted in the way that reduces these errors in each neuron predicted values and the desired outputs. The datasets of 52
from the output to the input layer. cases obtained from other literature [1] were divided into
In order to avoid the potential problem of over-training two sets. The first set consisted of 36 (∼70%) cases for
or memorization while employing the back-propagation training/testing and 16 (∼30%) cases for validation
algorithm, the option of saving the best result is adopted (simulation), chosen randomly from the set of 52 cases.
during the selected number of training cycles of 2,000.
4. Simulation and Results
Neural network
Input including connections Output Compare Target 4.1. Simulation
(weights) between

Computer simulation of ANN was employed for the

purpose of this work using the MATLAB version 7.0.4.365
Adjust
weights
(R14) Service Pack 2 commercial software package with
embedded neural network add-in toolbox.
Fig. 3. Adjusting (or training) of a Neural network Several ANN models were simulated (or trained) using
the thermo-physical properties datasets of Table 2. The
(iii) Optimal Configuration of Ann Model for Bakery feed-forward network structure with input, output and
Products hidden layers were also used and the generalized network
Upon using the mean relative error (MRE), mean absolute structures are as shown in Fig. 5 and Fig. 6 respectively.
error (MAE) and standard error (SE) as standard criteria, the
performances of the various ANN configurations were
Table 2. Thermal conductivity of dough and bakery products

Apparent Thermal Conductivity

Product Temp. (oC) Moisture content (%) Technique
Density (kg/m3) (W/mK)

Bread 22 28 - 36 190 – 210 0.091 - 0.104 Line heat source [19]

Bread dough -43.5 - 21 43.5 – 46.1 1100 0.500 - 0.920 Transient hot strip [27]
French bread 22 42.0 161.4 0.0989 Linear heat source [17]
Yellow cake batter 22 35.5 – 41.5 300 – 694 0.121 – 0.223 Linear heat source [17]
Cup cake batter 20 - 104 25 - 37 272 – 815 0.068 – 0.290 Linear heat source [1, 2]
Tortilla chip 25 1.4 – 35.6 520 – 880 0.09 – 0.23 Linear heat source [4]
Whole wheat dough 35 – 62.5 42.4 – 46.0 1035 – 1093 0.221 - 0341 Modified guarded hot plate [15]
Baked chapatti 35 – 62.5 38 – 48.7 1050 – 1200 0.142 - 0343
 Advances in Bioscience and Bioengineering 2014; 2(2): 14-24 20

Table 3. Error parameters in the prediction of thermal conductivity with

different neural network configurations

No. of
No. of
neurons in MAE SE
hidden MRE (%)
each hidden (W/mK) (W/mK)
layers
layer
1 2 0.17 0.0199 0.0046
1 4 4.2 0.2212 0.0970
1 6 0.035077 0.0038 0.0011
Fig. 4. A back propagation network 1 8 0.18 0.0191 0.0050
1 10 0.03388 0.0034 0.0011
1 12 0.26 0.0297 0.0073
1 14 0.090897 0.0149 0.0041
1 16 0.12 0.0115 0.0038
2 2 0.24 0.0298 0.0075
2 4 0.41 0.0641 0.0125
2 6 0.37 0.0606 0.0121
2 8 0.048776 0.0054 0.0015
2 10 0.29 0.0441 0.0106
2 12 0.22 0.0382 0.0095
2 14 0.17 0.0166 0.0047
Fig. 5. Generalized multilayer neural network
2 16 0.22 0.0323 0.0072

Table 4. Error prediction for 1 hidden layer with their corresponding

neurons

No. of neurons in MRE (% MAE (W/mK SE (W/mK

each hidden layers x10– 2) x10– 3) x10– 4)
2 17 19.9 46
Fig. 6. Generalized multilayer neural network obtained with MATLAB
4 420 221.2 970
For the purpose of this work, the input
put layer consisted of 6 3.5 3.8 11
3 neurons which corresponded to a product’s moisture 8 18 19.1 50
content, temperature and apparent density respectively, 10 3.4 3.4 11
while the output layer had 1 neuron representing the 12 26 29.7 73
thermal conductivity. The number of hidden layers and 14 9.1 14.9 41
neurons within each hidden layer can be varied based on 16 12 11.5 38
the complexity of the problem and dataset. Moreover, the
number of hidden layers was varied from 1 to 2. The Table 5. Error prediction for 2 hidden layers with their corresponding
neurons within each of these layers were varied from 2 to neurons
16 with increments of 2. This resulted in a total of 16 No. of neurons in MRE (% MAE (W/mK SE (W/mK
networks. each hidden layers x10– 2) x10– 3) x10– 4)
2 24 29.80 75
4.2. Results of Simulation 4 41 64.10 125
6 37 60.60 121
Once a given ANN configuration was trained using the
8 4.88 5.40 15
input data, its performance was evaluated using the same
10 29 44.10 106
dataset. The analysis was repeated several times. The ANN
12 22 38.20 95
configuration (out of 16) that minimized the three error 14 17 16.60 47
measures: MRE, MAR and SE, was selected as the 16 22 32.30 72
optimum.
The error measures associated with different ANN 4.3. Discussion of Results
configurations for prediction of thermal conductivity are
presented in Table 3. The optimal ANN configuration Tables 3 and 4 show the error parameters in the
included 2-hidden layers with 8 neurons in each layer. The prediction of thermal conductivity for 1- and 2-hidden
MAE, MRE and SE for this optimal configuration were layers with 2 to 16 neurons in each respective hidden layer.
0.0054W/mK, 4.8776 X 10-4 W/mK (4.8776 X 10-2 %) and The corresponding charts of Tables 4 and 5 are shown in
0.0015 W/mK respectively (shown highlighted below). Figs. 7and 8 respectively. For each of the corresponding
diagrams, there was a considerable and first-rate agreement
21 Ajasa, Abiodun Afis et al.: Thermal Conductivity of Food Products using: A Correlation Analysis Based
on Artificial Neural Networks (ANNs)

between the predicted and desired/targeted values of

thermal conductivities for different parameters of MAE,
MRE and SE.
It can be concluded that the predicted thermal
conductivity is good, efficient and credible prediction for
thermal conductivity of bakery products.

1200

1000

800
MRE (%) X 10– 2
600 MAE (W/mK) X 10– 3
SE (W/mK) X 10– 4
400

200 Fig. 9. 2-hidden layers with 8 neurons

0
1 2 3 4 5 6 7 8

Fig. 7. Corresponding chart showing the error values for 1-hidden layer

140

120

100 MRE (%) X 10– 2

80
MAE (W/mK) X
10– 3
60
SE (W/mK) X 10–
40 4

20 Fig. 10. 1-hidden layers with 10 neurons

0
4.4.2. Validating Dataset
1 2 3 4 5 6 7 8
The performance of the optimal neural network was
validated using a second dataset consisting of 16 cases not
Fig. 8. Corresponding chart showing the error values for 2-hidden layers previously used in the training. Thermal conductivity
values with an MRE of 0.0217 W/mK, MAE of 0.1544
4.4. Training, Validation and Prediction W/mK and SE of 0.0156 W/mK for 2-hidden layers with 8
4.4.1. Training Dataset neurons were predicted. The graph between the predicted
To reveal the credibility of prediction (with the training and the desired values of thermal conductivity for this
data set) from the optimal ANN, predicted values of model is shown in Fig. 11. There seemed to be no
thermal conductivity are plotted against the desired/targeted correlation between the predicted and the targeted thermal
values of thermal conductivity as shown in Fig. 9. Similarly, conductivities because of the few number of cases (16)
that of the simplest ANN model is shown in Fig. 10. These used for the validating dataset, as compared to that used for
two figures showed straight line curves, demonstrating the training set (36).
correlation between the predicted and targeted thermal This data set was also used with the simplest ANN
conductivities when the points are joined together (best fit- configuration of one hidden layer consisting of ten neurons.
points). The results demonstrate a very good agreement The model predicted thermal conductivity with a 4.73% (or
between the predicted and the desired values of thermal 0.0473 W/mK) MRE, 25.6% (or 0.2559 W/mK) MAE and
conductivity. Considering the inherent variation in the input 11.6% (or 0.1163 W/mK) SE. The graph between the
dataset, the simplest ANN configuration can be considered. predicted and the targeted values of thermal conductivity
The simplest ANN model with 1-hidden layer and 10 for the simplest ANN model is shown in Fig. 12. Similar
neurons predicted thermal conductivity with a 0.34% reasons could also be given for the non-correlation between
(0.0034 W/mK) MAE, 3.386 X 10-4 W/mK MRE and the predicted and targeted thermal conductivities for the
0.0011 W/mK SE being the least of all values of the error validating dataset, that is, fewer numbers of cases (16) were
measures. used.
 Advances in Bioscience and Bioengineering 2014; 2(2): 14-24 22

the predicted thermal conductivity is a good, efficient and

credible prediction in comparison to the desired/targeted
value (Figs. 13 to 18) for both the optimal and simplest
ANN models.

Fig. 11. 2-hidden layers with 8 neurons

Fig. 13. Thermal conductivities vs. Temperature (2 hidden layers with 8

neurons)

Fig. 12. 1-hidden layer with 10 neurons

4.4.3. Predicted and Targeted Thermal Conductivities

The thermal conductivity of the bakery products was
modelled by simulation. As a result, both the predicted and
Fig. 14. Thermal conductivities vs. Moisture content (2 hidden layers with
targeted/desired thermal conductivity values are plotted 8 neurons)
separately against each of the three dependent variables as
a function of moisture content (%), temperature (oC) and
apparent density (kg/m3). The corresponding curves were
obtained for the optimal ANN configuration; 2-hidden
layers with 8 neurons (Figs. 13 to 15) and the simplest
ANN configuration; 1-hidden layer with 10 neurons (Figs.
16 to 18).
For each of the corresponding diagrams, black triangles
(π) represent the predicted thermal conductivity while
purple stars ( ) represent the targeted/desired thermal
conductivity. Figs. 13 to 18 showed a considerable and
first-rate agreement between the predicted and
desired/targeted values of thermal conductivities as the
points considered (shown as triangles and stars) intersected
at some locations on the curves and nearly intersected at
other locations. Fig. 15. Thermal conductivities vs. Apparent density (2 hidden layers with
Since most of the triangle-points of the predicted values 8 neurons)
of thermal conductivity intersected with most of the star-
points of the thermal conductivity, it can be concluded that
23 Ajasa, Abiodun Afis et al.: Thermal Conductivity of Food Products using: A Correlation Analysis Based
on Artificial Neural Networks (ANNs)

Fig. 16. Thermal conductivities vs. Temperature (1 hidden layer with 10

neurons) Fig. 19. Thermal conductivities vs. Apparent density (1 hidden layer with
4 neurons)

Fig. 17. Thermal conductivities vs. Moisture content (1 hidden layer with
10 neurons)

Fig. 20. Thermal conductivities vs. Moisture content (2 hidden layers with
12 neurons)

5. Conclusion
In this paper, an ANN model was developed for
calculating the thermal conductivity of a variety of bakery
products under a wide range of conditions of moisture
content, temperature and apparent density. The optimal
model consisted of 2-hidden layers with eight neurons in
each hidden layer, and was able to produce thermal
conductivity values with a MAE of 54 X 10-4W/mK, MRE
Fig. 18. Thermal conductivities vs. Apparent density (1 hidden layer with of 4.878 X10-4W/mK and a SE of 15 X 10-4W/mK (see
10 neurons)
Table 3). However, the simplest ANN model has 1-hidden
Finally, there is a great disagreement or deviation layer with 10 neurons. This also showed a good prediction
between the predicted values of thermal conductivity and with a MRE of about 3.388 X 10-4 W/mK, MAE of 34 X
the targeted values when plotted against density for an 10-4 W/mK and SE of 11 X10-4 W/mK (see Table 4).
ANN model that has 1 hidden layer with 4 neurons, as From these values, it can be deduced and concluded that
shown in Fig. 19. Similarly, when the predicted and the simplest ANN model (with 1-hidden layer and 10
targeted values of thermal conductivity are plotted against neurons), when compared with the optimal ANN model
moisture content for an ANN model that has 2 hidden (with 2-hidden layers and 8 neurons in each hidden layer)
layers with 12 neurons, there was no agreement, as shown has smaller mean relative error,smaller mean absolute error
in Fig. 20. These ANN models are said to be inefficient and lesser standard error. Therefore, this model performs
models because of the non-intersection of most of their better accordingly.
triangles and stars points.
 Advances in Bioscience and Bioengineering 2014; 2(2): 14-24 24

neural network models in modelling quality changes of dry

peas in long time cooking”; Journal of Food Engineering,
References 41, (3/4), Pgs. 151 – 162; 1999.

[1] Baik, O. D., Marcotte, M., Sablani, S. S., and Castaigne, F. [14] Bishop, M. C. (1994): “Neural networks and their
(2001): “Thermal and physical properties of Bakery applications”; Review in Scientific Instruments, 64(6), Pgs.
products”; Critical Review in Food Science and Nutrition. 1803 – 1831; 1994.
CRC Press LLC; 2001.
[15] Ruan, R., Almaer, S., & Zhang, J. (1995): “Prediction of
[2] Baik, O. D., Marcotte, M., Sablani, S. S., and Castaigne, F. dough rheological properties using neural networks”;
(2006): “Modelling the Thermal properties of a Cup cake Cereal Chemistry, 72(3), Pgs 308 – 311; 1995.
during Baking”; Journal of Food Science Volume 64, Issue
2, Pgs. 295 – 299; 2006. [16] Rogers, G. F. C and Mayhew, Y. R. (1989): “Engineering
Thermodynamics”; ELBS 3/e, Pgs. 469 – 502; 1989.
[3] Ajasa, A. A., Akinyemi, L. A., Shoewu, O. O. and Adenowo,
A. A. (2014): “Exploitation of Artificial Neural Networks to [17] Sweat, V. E. (1985): “Thermal properties of low and
predict the Thermal Conductivity of Food Products in intermediate moisture food”; ASHRAE Transaction, 91, Pgs.
Nigeria”; American Journal of Engineering Research 369 – 389; 1985.
(AJER); Volume 03,Issue 03, Pgs 22 – 29.
[18] Fang, Q., Bilby, G., Haque, E., Hanna, M. A., and Spillman,
[4] Shyam Sablani, S and Shafiur Rahman, M. (2003), “Using C. K. (1998):” Neural network modelling of physical
neural networks to predict thermal conductivity of food as a properties of ground wheat”; Cereal Chemistry, 75(2),Pgs
function of moisture content, temperature and apparent 251 – 253; 1998.
porosity”; Food Research International. Elsevier. Volume 36,
Issue 6, 2003, Pgs. 617 – 623; 2003. [19] Bakshi, S. A. and Yoon, J. (1984): “Thermophysical
properties of bread rolls during baking”; Lebensmittel
[5] Rask, C. (1989): “Thermal properties of dough and bakery Wissenschaft und Technologie, 17, Pgs. 90 – 93; 1984.
products: A review of published data”; Journal of Food
Engineering, 9, Pgs 167 – 193; 1989 [20] Sablani, S. S. and Shayya, W. H. (2001): “Computerization
of Stumbo’s method of thermal process calculations using
[6] Hussain, A. M. and Rahman, M. S. (1999): “Thermal neural networks”; Journal of Food Engineering, 47, Pgs.
conductivity prediction of fruits and vegetables using neural 233 – 240; 2001.
networks”; International Journal of Food Properties; 2(2);
1999. [21] Christina Rask (1989): “Thermal properties of dough and
bakery products”; A review of published data. Journal of
[7] Ni, H., and Gunasekaran, S. (1998): “Food quality Food Engineering. Volume 9, Issue 3, 1989, Pgs. 167 – 193;
prediction with neural networks”; Food Technology, 52(10), 1989.
Pgs 60 – 65; 1998.
[22] Kulacki, F. A. and Kennedy, S. C. (1978): “Measurement of
[8] Sablani, S. S., Marcotte, M., Baik, O. D., and Castaigne, F. the thermo-physical properties of common cookie dough”;
(1998): “Modelling of simultaneous heat and water Journal of Food Science; 43, Pgs. 380 – 384; 1978.
transport in the baking process”; Lebensmittel Wissenchaft
and Technologie, 31, Germany; Pgs 201 – 209; 1998. [23] Buhri, A. B. and Singh, R. P. (1993): “Measurement of food
thermal conductivity using differential scanning
[9] Linko, P., and Zhu. Y. H. (1991): “Neural network calorimetry”; Journal of Food Science; 58, Pgs. 1145 – 1147;
programming in bioprocess variable estimation and state 1993.
prediction”; Journal of Biotechnology, 21(3), Pgs 253–270;
1991. [24] Zhou, L., Puri, V. M. and Anantheswaran, R. C. (1994):
“Measurement of coefficients for simultaneous heat and
[10] Hornik, K., Stinchombe, M. and White, H.: (1989): mass transfer in food products”; Drying Technology; 12;
“Multilayer feed-forward network and universal Pgs. 607 – 627; 1994.
approximator. Neural Network”, 2, Pgs. 359 – 366; 1989.
[25] Vos, B. H. (2005):”Measurements of thermal conductivity
[11] Myhara, R. M., Sablani, S. S., Al-Alawi, S. M., and Taylor, by a non-steady state method”; Applied Science Research;
M. S. (1998):”Water sorption isotherms of dates: Modelling A5; Pgs. 425 – 438; 2005.
using GAB equation and artificial neural network
approaches”; Lebensmittel Wissenschaft und Technologies, [26] Gustafsson, S. E., Karawacki, E. and Khan, M. N. (2001):
31(7/8), Germany; Pgs 699 – 706; 1998. “Transient hot-strip method for simultaneously measuring
thermal conductivity and thermal diffusivity of solids and
[12] Murakami, E. G., and Okos, M. R. (1989): ‘‘Measurement fluids”; Journal of Applied Physics; 12; Pgs. 1411 – 1421;
and Prediction of thermal properties of foods’’. In R. P. 2001.
Singh, and A. G. Medina (Eds.), Food Properties Computer-
aided Engineering of Food Processing System; Pgs. 3 – 48; [27] Lind, I. (1998): “Thawing of minced meat and dough:
Norwell, MA; Kluwer Academic Publishers, 1989. thermal data and mathematical modeling”; Pro. Int. Symp.
On progress in food preservation process; Vol. 1, CERIA,
[13] Xie, G. and Xiong, R. (1999): “Use of hyperbolic and Brussels, Belgium; Pgs. 101 – 111; 1998.