chapter 2 neural network applications in heat and … 2 neural network applications in heat and mass...

CHAPTER 2

Neural network applications in heat and masstransfer operations in food processing

C.R. Chen1, H.S. Ramaswamy2 & M. Marcotte1

1Food Research and Development Centre, Agriculture and Agri-FoodCanada, Canada.2Food Science Department, McGill University, Canada.

Abstract

Artificial neural networks (ANNs) as a potential method for characterization, mod-elling and classification have been widely used in many application areas includingfood processing. This chapter focuses on applications of neural networks in the heatand mass transfer operations/techniques used in food processing. Neural networkprinciples, development of neural network models and major applications of neuralnetworks in heat and mass transfer areas will be covered along with some informa-tion on future trends. The first section introduces the concept of neural networks, itsdevelopment history and its potential application purposes in food processing. Thesecond section highlights the development of neural network models with some the-oretical background and demonstrates its advantages over conventional methods.Major research activities involving the use of ANN models with relation to heat andmass transfer applications in food processing such as thermal processing, baking,extrusion, drying and freezing will be addressed next. Finally, the future trends forapplication of neural networks in food processing will be briefly discussed.

1 Introduction

The concept of neural networks was based on the research in artificial intelligence,which was intended specifically to try mimic the fault tolerance and learning capac-ity of biological neural systems by modelling the low-level structure of the brain.Warren McCulloch and Mathematician Walter Pitts in 1943 were the first to openthe idea on how neurons might work and they modelled simple neural networks

www.witpress.com, ISSN 1755-8336 (on-line)

© 2007 WIT PressWIT Transactions on State of the Art in Science and Engineering, Vol 13,

doi:10.2495/978-1-85312-932-2/02

40 Heat Transfer in Food Processing

using electrical circuits (NeuralWare [1]). Donald Hebb in 1949 wrote the book TheOrganization of Behavior, which described the concept of learning rule essential tothe ways in which humans learn.As computers became more advanced in the 1950s,it was finally possible to simulate a hypothetical neural network. In 1959, BernardWidrow and Marcian Hoff of Stanford developed models called ‘ADALINE’ and‘MADALINE’ [1, 2]. In 1962, Widrow and Hoff developed a learning procedurethat examined the value of the binary variable before the weight adjustment (i.e. 0or 1), which was one of important fundamentals to the subsequent success of neuralnetworks [3]. However, these initial and later successes of the neural network didnot result in wide practical applications until the 1980s. During that time severalnew approaches such as bi-directional lines, hybrid network and multi-layer neuralnetworks were developed [1–5]. In addition to these advances in the algorithmsthemselves, rapid development of computer technologies, including both hardwareand software, became a prominent driving force for the use of neural networks as acomputational technique to be widely used not only in the area of computer sciencebut also in other science and engineering areas for prediction, classification andoptimization.

The application of neural networks in food processing operations has consider-ably lagged other prominent areas of application such as chemical engineering, agri-cultural engineering and petro-engineering. Nevertheless, there has been a steadygrowth in the use of ANN concepts even in the food area in recent years. Figure 1shows the reported ANN applications in food processing from 1994 to 2000 [6].

2530 34

40

6172 66

0

20

40

60

(a)

(b)

Num

ber

of p

aper

s

1994 1995 1996 1997 1998 1999 2000

164140

6447

0

40

80

120

160

Num

ber

of p

aper

s

Modelling Control Classification Optimization Computation

1

Figure 1: Applications ofANN in food processing areas: (a) papers published annu-ally and (b) papers on application purposes from 1994 to 2000.



Neural Network Applications 41

It can be seen from Fig. 1a that applications of ANN in the food processing areahave greatly increased since 1994, specifically during 1994–1998. Figure 1b showsthat ANN applications reported in food processing areas have also covered all theprincipal objectives including modelling, prediction, computation, classification,optimization and control.

This chapter focuses on the introduction of basic principles of neural networks,development of neural network models and their application advances in heat andmass transfer operations related to food processing.

2 Principles of a basic artificial neural network (ANN) model

The basic principle of neural networks is the simulation of human neural networkbehaviour and structure. A typical biological neuron as shown in Fig. 2 containsneuronal cell bodies (soma), dendrites and axons. Each neuron receives electro-chemical inputs from other neurons at the dendrites. If the sum of these electricalinputs is sufficiently powerful to activate the neuron, it transmits an electrochemicalsignal along the axon and passes this signal to the other neurons whose dendritesare attached at any of the axon terminals. These attached neurons may then fire.

It is important to note that a neuron fires only if the total signal received at the cellbody exceeds a certain level. The entire brain is composed of these interconnectedelectrochemical transmitting neurons. From a very large number of extremely sim-ple processing units (each performing a weighted sum of its inputs, and then firing a

Cell body

Axon

Synapses

Figure 2: The structure of a typical biological neuron.




binary signal if the total input exceeds a certain level), the brain manages to performextremely complex tasks.

However, it is noted that ANNs only represent extremely simplified formal mod-els of biological neurons and their interconnections without making any attempt tomodel the biological system itself. Their importance lies in the fact that artificialnetworks are brain-inspired computational tools for solving complex problems.

2.1 Neural network architecture

Figure 3 shows a typical feed-forward neural network with multiple layers formedby an interconnection of nodes. This neural network has an input layer, two hiddenlayer(s) and one output layer. Each layer is essential to the operation of the network.A neural network can be viewed as a ‘black box’ into which a specific input to eachnode in the input layer is sent. The network processes this information throughthe interconnections between the nodes, but the entire processing step is hidden.Finally, the network gives an output from the nodes on the output layer.

The purpose of each layer is

• input layer – receives information from an external source, and passes thisinformation to the network for processing.

• hidden layers – receive information from the input layer and do all of the infor-mation processing. The entire processing is hidden from view. The number ofhidden layers can be 1–3, depending on the problem being investigated.

• output layer – receives processed information from the network, and transmitsthe results to an external receptor.

When the input layer receives information from an external source, it becomes‘activated’ and emits signals to its neighbours. The neighbours receive excitationsfrom the input layer, and in turn emit signals to their neighbours. Depending on the

Input

Output

Hidden

Y1

Ym

X1

X2

Xj

Learning rule

Targetvector

e

Adjust Wij

Figure 3: A typical multi-layer neural network with one hidden layer.




strength of the interconnections (i.e. the magnitude of the so-called weight factorthat adjusts the strength of the input signal), these signals can excite or inhibit thenodes. What results is a pattern of activation that eventually manifests itself inthe output layer. Finally, the values from the output layer will be compared with thedesired values. If the difference between the output and the desired values is largerthan the set error range, then the weight factors are adjusted through the repeatedtraining until the error is within the set error range or the number of learning runsexceeds a pre-set number.

2.2 Artificial neurons

Artificial neurons are simple processing units similar to biological neurons: theyreceive multiple inputs from other neurons but generate only one output. This outputmay be propagated to several other neurons.

Each neuron has two basic functions: gathering information from the other neu-rons in the preceding layer and sending the signals to the neurons in the next layers.The first artificial neuron model proposed in 1943 by McCulloch and Pitts (Fig. 4)is based on the simplified consideration of the biological model.

The elementary computing neuron functions as an arithmetic logic comput-ing element. The binary inputs of the neurons are x1, x2, . . . , xn. Zero representsabsence, and one represents existence. The weight of connection between the ithinput xi and the neuron is represented by wi. When wi > 1, the input is excitatory.When wi < 0, it is inhibitory.

The net summation of inputs weighted by the synaptic strength wi at connectioni is

net =n∑

i−1

wixi (1)

The net value is then mapped through an activation function f of neuron output.The activation function used in the model is a threshold function:

y = f (net) (2)

f (x) ={

1, x > θ

0, otherwise(3)

where θ is the threshold value.

∑ f Y..

x1

xn

x2

.

.

SumTransferfunction

Figure 4: Representation of an artificial neuron.




Neuron models used in current neural networks are constructed in a more generalway. The input and output signals are not limited to the binary data, and the activationfunction can be any continuous function other than the threshold function used inthe earlier model. The activation function is typically a monotonic non-decreasingnonlinear function. Some of the often used activation functions are (where α denotesthe parameter and θ denotes the threshold value):

Sigmoid function: f (x) = 1

1 + e−αx(4)

Hyperbolic function: f (x) = tanh (αx) = eαx − e−αx

eαx + e−αx(5)

Linear threshold: f (x) =

1 x ≥ θ

x/θ 0 ≺ x ≺ θ

0 x ≤ θ

(6)

Gaussian function: f (x) = e−αx2(7)

2.3 Learning rules

There are two learning modes available for networks: supervised and unsupervised.In the supervised mode, training a neural network involves feeding the network a setof known input–output patterns, and adjusting the network weights until each inputproduces the appropriate output. Thus, training a neural network means adjustingthe weight factors until the output pattern (response) calculated from the given inputreflects the desired relationship. In the unsupervised mode, the neural networks arepresented only with a series of input patterns without any information about theirdesired outputs. Thus, there is no defined criterion to adjust the weights basedon the specific or target outputs. During training, the network attempts to groupinput patterns that are similar to each other and adapts according to a particularorganization scheme.

The primary training method for supervised learning is error correction learning,which uses the back propagation of error to adjust network’s weights and thresholdsso as to minimize the error in its prediction on the training set, also known as thedelta rule. It is based on the gradient descent method to minimize the squaresof differences between the actual and desired outputs by adjusting values of theconnecting weights. Mathematically, the difference between the actual and desiredoutputs is given as

εi = di − ci (8)

where εi is the output error, di is the desired output and ci is the calculated outputfor the ith neuron on the output layer only. If there are n outputs in the output layer,the total square error on the output layer can be calculated as

E =n∑

i=1

ε2i =

n∑i=1

(di − ci)2 (9)




The target of network training is to minimize the total square error (value E)by adjusting the node connection coefficients or weights (W ). Generalized deltarule (or delta rule) is the most commonly used learning mechanism for multi-layerfeed-forward networks with non-linear node function such as a sigmoid. By use ofthis rule, the weight change can be calculated by the following equation:

wnew = wold + ηaδ + λ�wold (10)

where

δ = (d − c)f (net)(1 − f (net)) (11)

and

f (net) = 1

1 − exp (−net)(12)

where η is a linear proportionality constant for node j, called the learning rate(typically, 0 < η 1), and a is the input to the neuron and λ is a constant knownas momentum.

2.4 Advantages and limitations of neural networks

Compared to conventional modelling methods, neural networks have many advan-tages as below [7]:

• Learning ability Neural networks have learning ability similar to the humanbrain, meaning that they can construct the cause-and-effect relationships throughrepeated training without any prior knowledge of the system being investigated.Therefore, neural networks are suitable for use in cases with multiple variablesand complicated internal relationships, which are often difficult to describe bymathematical equations.

• Robust and fault tolerance The ANN is more tolerant of noisy and incompletedata because the information is distributed in the massive processing nodes andconnections. Minor damage to parameters in the network will not degrade theoverall performance significantly.

• High computational speed TheANN is an inherently parallel architecture. Theresult comes from the collective behavior of a large number of simple, parallelprocessing units. Therefore, once trained, neural networks can calculate resultsfrom a given input very quickly. This has made neural network models to havea greater potential than conventional modeling methods to be used in onlinecontrol systems.

However, there are also some limitations for applications of neural networkssuch as:

• Works as a black box Neural networks work as a black box, meaning that theycan only provide results but not give any reasonable interpretations between




input and output variables. In this aspect, neural networks are not suitable forprecise numerical computations or following the logical sequence of operations.

• Large amount of training data If little input–output data exists on a problemor process, the use of neural networks should not be considered since they relyheavily on such data. Consequently, neural networks are best suited for problemswith a large amount of historical data, or those that allow to train the neuralnetwork with a separate simulator. In addition, there may also be situationswhere there exist large databases, but all training data are similar, causing thesame problems as having small training data sets. Thus, a broad-based data setand experimental design are essential.

• No guarantee of optimal results Most training techniques are capable of‘tuning’ the network, but they do not guarantee that the network will operateproperly. The training may ‘bias’ the network, making it accurate in some oper-ating regions, but inaccurate in others.

3 Developing neural networks

Neural networks can be developed either based on the principles of neural net-works using a specific computer language or by use of one of commercial neuralnetworks software. Of course, developing neural network codes, which means toturn the theory of a particular network model into the design for a computer sim-ulation implementation, can be a challenging task for most applied scientists andengineers who do not have the programming and related knowledge of neural net-works. So the use of a commercial software has been the most popular method fordeveloping an ANN model. With the rapid development of computer software, sev-eral ANN softwares have been developed which can be used for developing ANNmodels for various specific purposes. Some examples include Neural-Ware Pro-fessional, Neural-Shell, Neuro-solution (Neuro-Dimension, Inc., Gainesville, FL),Matlab Neural network toolbox (MathWorks, Inc., Natick, MA), Statistica Neu-ral Networks (StatSoft, Inc., Tulsa, OK) and Neuro-Genetic Optimizer (BioCompSystems, Inc, Bloomington, MN).

Developing a neural network by using a commercial neural networks softwareconsists of following steps:

• Selection of inputs and outputs• Data collection• Optimization of configurations• Training or learning• Testing or generation.

The number of outputs can be determined by the problem being investigatedwhile the number of inputs should be all those that have a significant influence onthe outputs which can be decided by comparison of results from different experi-ments with different inputs. The size of data required for neural network trainingis dependent on the complexity of the underlying function which the network istrying to model and the variance of the additive noise. Normally, neural networks




can only process numeric data in a limited range but it is also possible for neuralnetworks to handle different types of data such as an unusual range, missing data ornon-numeric parameters through some methods. For example, numeric data can bescaled into an appropriate range for the network, missing values can be substitutedby using mean values and non-numeric data can be represented by a set of numericvalues.

For each specific problem, in order to develop a neural networks model withthe best performance, the configuration parameters of the neural network must bedetermined by trials and errors. These parameters include transfer functions, learn-ing rules, learning rate, momentum coefficient, number of hidden layers, numberof neurons in each hidden layer and learning runs.

In the training or learning step, a set of known input–output data is repeatedlypresented to train the network. During this repetition process, the weight factorsbetween nodes are adjusted until the specified input yields the desired output.Through these adjustments, the neural network ‘learns’ the correct input–outputresponse behaviour. In neural network development, this phase is typically thelongest and most time consuming, and it is critical to the success of the network.

After the training step, the recall and generalization step will be carried out. Inthe recall step, the network will be subjected to a wide array of input patterns usedin training, and adjustments are introduced to make the system more reliable androbust. During the generalization step, the network will be subjected to input pat-terns which it has not seen before, but whose outputs are known, and the system’sperformance will be monitored. The performance of neural networks can be eval-uated visually by graphs or in quantity by different statistical measures. Figure 5shows three kinds of graphs used for the performance of neural networks modelling.

In Fig. 5a, x and y axes represent desired and predicted results, respectively.The diagonal line called the ideal line can be easily used for evaluation of themodelling performance. If neural network predicted values are equal or close todesired results, then the points should be located on or close to the ideal line. InFig. 5b, x axis is one of the input variables while y axis is used for outputs includingdesired and predicted results. This graph can be used to determine the matchabilityof the neural networks model with the specific input variables. In Fig. 5c, x axis is

Experimental

AN

N p

redi

cted

Input

Out

put Experimental

ANN

Data number

Out

put v

aria

ble

(a) (c)(b)

Figure 5: Graphs used for the evaluation of neural networks modellingperformance.




just representing the number or order of data used for training or testing, while yaxis is used for outputs as Fig. 5b. Its emphasis is on comparison of the agreementbetween each pair of predicted and desired values.

Often used statistical parameters include the coefficient R2 of relationshipbetween predicted and experimental results, and average relative error (Er), whichare given as

R2 = 1 −∑n

i=1 (yi − ydi)2∑ni=1 (yi − ym)2 (13)

Er =∑n

i=1 |(yi − ydi)|/n

ymax − ymin(14)

where yi is predicted by the ANN model, ydi is the actual desired value (actualvalues), n is the number of data and ym is the average of actual values.

4 ANN applications in heat and mass transfer operations

Heat and mass transfer applications can be found in many food processing unitoperations such as cooking, drying, freezing, canning and baking. The process ofheat and mass transfer affects many processing results such as quality and safety ofproducts, process time and energy consumptions, therefore, investigations of phe-nomena related to heat and mass transfer in food processing operations are alwayspopular subjects for food scientists and engineers. To date, various cconventionalmethods such as analytical method, numerical method including finite differenceand finite element simulations and regression methods have been widely used forthe modelling of heat and mass transfer problems. The common demand for usingthese conventional methods are that the relationship between input and output vari-ables is clear and principal properties of materials or process system can be experi-mentally measured. Unlike other disciplines, food processing techniques deal withbio-materials which show much more complicated thermophysical properties anduncertainties during processing. This also results in more complex relationshipsbetween input and output variables. Thus, in many cases, it is difficult to use a stan-dard partial differential equation or model as generally used in other engineeringapplications to accurately describe the phenomena occurring in food processingoperations. With the rapid development of modern computer technologies, neuralnetwork as an artificial intelligence technology is emerging as a potential and alter-native tool to deal with such problems. These can have a significant advantage overconventional methods described earlier. A general status of scientific reports pub-lished in recent years in the area of heat and mass transfer operations is summarizedbelow.

4.1 Predictions

Neural networks have been frequently and successfully used for predictions ofheat and mass transfer variables in food processing. These applications includethe evaluation of heat transfer coefficients [8, 9], residence time distribution under




the aseptic processing conditions [10], thermal conductivity of fruit and vegetables[11], electrical conductivity [12], food extrusion [13–15], drying [16–21], freezing[22], retort thermal processing [23–28], deep-fat frying [29] and baking [30, 31].

4.2 Optimization

Neural networks have also been used for some optimization purposes combined withsome other search techniques which can be either conventional such as responsesurface method or artificial intelligence search technique such as genetic algorithms.Typical applications include heat treatment of fruit storage [32, 33], constant retorttemperature (CRT) thermal processing [26], and variable retort temperature (VRT)thermal processing [27, 28].

4.3 Control

Online use of neural networks is another important advantage of neural networkscompared with conventional modelling methods since they can compute resultsfrom a given input very quickly. Successful applications have been reported forfood extrusion [13, 14], corn drying [34], sugar refinery [35], food frying [36] andfermentation [37, 38].

4.4 Selected ANN examples in heat and mass transfer applications

4.4.1 Neural network modelling of end-over-end processing [9]Overall heat transfer coefficient (U) and fluid to particle heat transfer coefficient(hfp) are fundamental data needed to develop prediction models of transient temper-atures which are used for establishing and optimizing thermal processing schedulefor canned liquid/particle food system during agitation processing. Traditionally,dimensionless equations are used for development of experimental models of U andhfp by involving other influencing parameters through the use of multiple regres-sion analysis. However, selection of appropriate dimensionless groups requiresprior knowledge of the phenomena under investigation. Sablani and co-workers [9]developed an ANN model for U and hfp associated with liquid particle mixtures,in cans subjected to end-over-end rotation. Experimental data obtained for U andhfp under various test conditions (shown in Table 1) were used for both trainingand evaluation. Multi-layer neural networks with seven inputs and two outputs (fora single particle in a can), and six inputs and two output (for multiple particles ina can) were trained. The optimal network obtained by initial trials consisted of 2hidden layers, 10 neurons in each hidden layer and 50,000 learning runs. By usingthe trained ANN models with optimal configurations, the prediction performance(R2) of ANN models for both U and hfp were reported to be higher than 0.98,meaning that developed ANN models could be successfully used for predicting Uand hfp under the given experimental conditions. The comparison of ANN modelsand dimensionless regression model using same experimental data is summarizedin Table 2. Prediction errors using ANN were less than 3% and 5%, respectively, forU and hfp, which were about 50% better than those associated with dimensionless




Table 1: Range of system and product parameters used in the determination of heattransfer coefficients (U and hfp).

Parameter Experimental range

Retort temperature 110◦C, 120◦C and 130◦CRadius of rotation 0, 0.09, 0.19 and 0.27 mRotation speed 10, 15 and 20 rpmCan headspace 0.0064 and 0.01Test fluid Water and oilTest particle Polypropylene, nylonParticle concentration Single particle, 20%, 30% and 40% (v/v)Particle shape and size

Cube 0.01905 mCylinder 0.01905 × 0.01905 mSphere 0.01905, 0.02225 and 0.025 m

Can dimension 307 × 409 (8.73 × 11.6)

Table 2: Comparison of error parameters for neural network models and dimen-sionless correlation models.

Single particle Multiple particle

U hfp U hfp

Error parameters DC NN DC NN DC NN DC NN

MAE1a 17.1 5.11 31.3 17.2 25.1 9.85 75.4 48.1SDEb 25.4 4.76 43.3 16.0 32.0 11.0 63.4 40.7MREc (%) 5.00 2.46 16.9 5.82 5.70 2.57 8.26 4.52SREd (%) 3.76 2.51 11.9 7.00 4.65 1.96 7.12 3.90R2 0.99 0.99 0.83 0.98 0.98 0.99 0.96 0.98

aMean absolute error.bStandard deviation of error.cMean relative error.dStandard deviation of relative error.

number models, indicating that the predictive performance of the ANN was farsuperior than that of dimensionless correlations.

4.4.2 Thermal conductivity prediction of fruits and vegetables usingneural networks [16]

Thermal conductivity is an important thermal property used to estimate the rate ofconductive heat transfer in food processes such as sterilization, drying, cooking and




frying, but it is a difficult and time-consuming task to find proper functional formsfor developing a good model, which may be theoretical or empirical [16].

In this example, ANN models were used to predict the thermal conductivity ofvarious fruits and vegetables (apples, pears, corn starch, raisins and potatoes), andto estimate the error between the experimental values and the theoretical modeldeveloped. For prediction of thermal conductivity, three principal variables wereused: moisture content, temperature and porosity (three neurons in the input layer).The ANN model consisted of 1 hidden layer with 10 neurons and 1 output, thethermal conductivity. A data set of 276 experimental values was used to train theneural network model while another data set of 25 experimental results was fortesting the trained neural network model.

The performance of the ANN model was evaluated by root mean square (RMS)value. During training, the data size was adjusted until the RMS value was lessthan 0.05. This method of training the ANN ensured that it gave a reasonably goodgeneralized prediction of the thermal conductivity from any input data supplied tothe model and predicted not just from the training database. For prediction of mis-match between model and experimental values for thermal conductivity, a hybridmodel consisting of a neural network and an empirical model was used. The valuepredicted by an empirical model was considered to be one of neural network inputsso that the neural network model had four inputs including the other three factorsused in the previous ANN model. An optimum configuration of 4 inputs, 10 neu-rons in the hidden and 1 output node was thus obtained for this network with asigmoidal activation function. Based on the associated sum of square error (SSE),they reported the hybrid model to significantly improve the prediction accuracy.

4.4.3 Dynamic modelling of a food extrusion process [13]Dynamic modelling of food extrusion process is complex specifically during thestart-up period and difficult to predict using conventional mathematical modellingsince a large number of variable parameters need to be taken into account. Popescuet al. [13] used ANN for modelling the start-up of a food extrusion process. Thestructure of the developed ANN model is shown in Fig. 6. It consisted of 36 neuronsin the input layer, 6 neurons in the hidden layer and 4 neurons in the output layer.Input variables included six independent variables – the flour mass flow rate (F), thewater flow rate (W ), the screw speed (RPM), and temperatures of the three barrelzones closest to the die section (T3, T4 and T5) – and four dependent (feedback)variables from outputs. Each independent variable needed four neurons for currentvalue plus three delays, respectively, while each dependent variable needed threeneurons for three delays, respectively.

The output variables were the total torque (Tq), the pressure at the die (P), the dietemperature (Tdie) and overall acceptability of the product (M). The product wasgraded on a 10 point scale by the experienced supervisor, 10 representing the worstoverall acceptability of the product at the beginning of the start-up process and 1representing the best overall acceptability of the product that should be reached atsteady state. The modelling performance for both training and testing was reportedto indicate that there existed good agreements between experimental and predicted




MRPM T3 T4 TPF TqT5W

t-3, t-2, t-1,t t-3, t-2, t-1,t

Tq P Td M

Inputlayer

Hiddenlayer

Outputlayer

t-3, t-2, t-1 t-3, t-2, t-1……

Figure 6: Neural network model for food extrusion process.

values for all outputs. Although at the beginning, the relative errors were high, theywere reported to decrease and settle down to within 10%.

4.4.4 ANN modelling and simulation control of cake baking [39]Baking is crucial in determining the quality of cake. Sponginess, flavour, texture,volume and browning are primary indicators for cake quality. Therefore, on-linemeasurement is necessary to control the process for high product quality. It is hardto measure sponginess, flavour and texture on a real-time basis during the processbecause of the absence of proper sensors, but online measurement of volume andbrowning are possible using image processing techniques. In this study, three neuralnetwork models were developed for the prediction of volume and browning after30 s, volume and browning after 2 min, and bread temperature during the baking byuse of multiple layer networks with back propagation algorithm.All the models weretrained and tested by two independent data sets, respectively. Model I developed forpredicting cake volume and browning after 30 s consisted of five neurons in inputlayer, eight neurons in the first hidden layer, six neurons in the second hidden layerand two neurons in the output layer. The five inputs included volumet , volumet−1,brownt , brownt−1, and oven temperaturet , while two outputs were for volumet+1and brownt+1 (the subscript ‘t + 1’ means 30 s after current state and t − 1 means30 s before current state).The statistical results for the prediction performance of thismodel indicated very good agreements between predicted and experimental valuesfor both volume and browning. Model I could also be used to predict volume andbrowning after 1, 1.5 or 2 min, but the prediction error increased with the length of




prediction time. Therefore, a second model (Model II) capable of directly predictingbread volume and browning after 2 min was developed, which contained the sameinput as the previous model but different outputs and different number of neuronsin hidden layers. Outputs were volumet+4 and brownt+4, the first hidden layer had10 neurons, and the second hidden layer had 6 neurons. The performance of themodel was reported to improve compared with Model I.

Model III was developed for the cake temperature prediction based on volumeand browning information. The model consisted of seven inputs, six neurons in thefist hidden layer, four neurons in the second hidden layer and two outputs. Inputsincluded volumet , volumet−1, volumet−2, browningt , browningt−1, browningt−2and oven temperature, while outputs were inside cake temperaturet and surfacecake temperaturet . The statistical results for the performance of model predictionswere 0.95% for inside cake temperature error and 0.43% for surface bread tem-perature error, showing that developed network model was successful for the caketemperature predictions.

These models were then used with a fuzzy logic controller for the baking processand to simulate the baking oven control. OVNTP (oven temperature) was used foroutput fuzzy variable. BROWN (browning of cake), SIZE (volume of cake) andBRDTP (cake temperature) were used for inputs. Knowledge from experiments andan experienced operator were used to construct 11 rules for the fuzzy controller.From the simulation based on the neural networks models, it could be verified thatthe cost of heating the oven could be reduced by using the fuzzy logic controllerrather than a human operator, without any loss of bread quality.

4.4.5 Modelling and optimization of CRT thermal processing using coupledneural networks and genetic algorithms [25, 26]

Modelling and optimization of thermal processing are of considerable interest, butare still mostly based on conventional mathematical methods. Traditionally, solv-ing an optimization problem consists of two steps. First, the different objectivefunction models are developed using mathematical approaches such as regressionmethods, theoretical analysis models and differential equations; second, optimalconditions are sought using one of several search methods such as direct search,grid search and gold section method, for single variables, and alternating variablesearch, pattern search and Powell’s method, for multi-variables. Like traditionalmodelling methods, neural networks cannot provide direct answers for optimiza-tion problems. In order to be used for optimization purposes, neural network modelshave to be combined with one of search techniques. Genetic algorithms are a com-binatorial optimization technique, searching for an optimal value of a complexobjective function by simulation of the biological evolutionary process, based oncrossover and mutation as in genetics. It has been found that the combination ofGA and ANN models can become effective tools for optimization problems. Chenand Ramaswamy [26] were the first to report on application of GA and ANN for thethermal processing optimization. ANN models were developed for predicting pro-cess time (PT), average quality retention (Qv), surface cook value (Fs), equivalentunit energy consumption (En), temperature difference (g) and ratio of F value from




heating to total desired F value (ρ) under different processing conditions. Thesewere then coupled with GA to search for the optimal quality retention and the cor-responding retort temperature. The combined ANN-GA models were then used forinvestigating the effects of process variables on both optimal quality retention andretort temperature.

Processing conditions as inputs for ANN models were selected as follows: retorttemperature (RT = 110–140◦C), thermal diffusivity (α = 1.1–2.14 × 10−7 m2/s),volume of can (V = 1.64–6.55 × 10−4 m3), ratio of height to diameter of can(Rdh = 0.2–1.8), total desired lethality value (Fo = 5–10 min) at the can centre,and quality kinetic destruction parameters: decimal destruction time (Dq = 150–300 min) and their temperature dependence (zq = 15–40◦C). Six separate ANNmodels were developed for prediction of process time (PT), average qualityretention (Qv), surface cook value (Fs), equivalent energy consumption (En), finaltemperature difference (g) at the can centre, and lethality ratio, ρ (heating/totallethality), respectively. The data for training and testingANN models were obtainedfrom a finite difference computer simulation program. A second-order central com-posite design was used for constructing experimental data for trainingANN models,while an orthogonal experimental design comprising six factors and three levels wasused for the generalization of trainedANN models. The hybrid optimization method(shown in Fig. 7) linking GAs with ANNs were employed for searching the optimalquality retention and corresponding retort temperature, and for investigating the

Initialpopulation

Compute the fitness of all individuals using ANN

Selection Crossover Mutation

Iterationnumber ?

Ni

input output

Iterationnumber ?

Y

Y

N

N

Figure 7: The procedure of hybrid optimization method using ANNs and GAs.




effects of main processing parameters on optimal results. ANN-based predictionmodels successfully described the various outputs of CRT thermal processing (cor-relation coefficients: R2 > 0.98; relative errors: Er ≤ 3%). The coupled ANN-GAmodels, verified under several typical processing conditions, could be effectivelyused for optimization of CRT thermal processing. The main processing parametersand their interactions in the order of their importance with respect to the optimalquality retention and corresponding retort temperature were V > zq > Fo > Rdhand zq > Fo > Rdh > V , respectively.

4.4.6 Modelling and optimization of VRT thermal processing using coupledneural networks and genetic algorithms [27]

VRT thermal processing is an effective technique used for improving the qualityof canned foods and reducing the process time. The key for designing a VRT ther-mal process is to choose a reasonable (or optimal) VRT function for a given foodproduct to be packaged in a given container and to be thermally processed. Sincethe often used VRT functions such as sine, exponential, step line and step pulseall involve several parameters, searching out an optimal VRT function is a multi-variable optimization problem. Although many conventional methods have beenused for solving a variety of optimization problems, they have several limitationssuch as lack of robustness for providing optimal results and low calculation speedboth of which are important for practical applications.Artificial intelligent technolo-gies have opened new avenues to deal with such problems. Chen and Ramaswamy[27] explored the use of ANN and GA to develop various prediction models ofVRT processing and search for the optimal VRT function parameters, respectively.The research methodology was similar to the one in the last example, but the researchobjective was focused on (1) analysing the effects of VRT function parameterson the main responses variables: process time, average quality retention, surfacecook value, (2) determining the search ranges for optimization of VRT processing,(3) developing ANN prediction models for each main response variable relatedto VRT function parameters and (4) searching the optimal processing temperatureprofiles using the coupled ANN-GA models.

ANN models were implemented to develop prediction models for VRT outputvariables: average quality retention (Qv), process time (Pt) and surface cook value(Fs). Genetic algorithms (GA) were coupled with trained neural network models tomeet different optimization objectives: minimum Pt and Fs, under given constraints.The searching range of each independent variable was based on a sensitivity analysisof effects of function parameters on response variables. The best results for Qv, Ptand Fs under CRT processing conditions were used as constraints. Test resultsindicated that coupled ANN-GA models could be effectively used for describingthe relationships between the operating variables andVRT function parameters, andfor identifying optimal processing conditions. VRT processes reduced the processtime by more than 20% and surface cook value by about 7–10% as compared to thebest CRT process. By use of ANN models, the benefits on both process time andsurface quality by using VRT method were predicted as shown in Fig. 8a and b.In Fig. 8a for processing time Pt, surface cook value was used as a constraint




(b)

5

6

7

8

9

10

11

Processing time PT (min)

Surf

ace

cook

ing

valu

e Fs

(m

in)

VR

CR

(a)

5

6

7

8

9

10

11

6 7 8 9 10 6 7 8 9 10 11Surface cooking value Fs (min)

Proc

essi

ng ti

me

PT (

min

)

Exponential

CR

Figure 8: Comparison of VRT versus CRT processing on optimal results of pro-cessing time and surface cooking value.

condition, while in 8b for surface cook value, the processing time was used as theconstraint condition.

5 Future trends

The capability of ANN modelling is not a question anymore as its potential hasbeen confirmed time and again by a variety of applications. The limitation of thistechnique is that it cannot give a clear internal relationship between variables. Basedon literature reviews, research activities can be categorized into two kinds: theo-retical and applied. The former ones often focus on developing and understandingthe intrinsic relationships between variables while the latter ones concentrate onmethods potentially used for practical applications. From this point of view, neuralnetworks should be considered more a tool explored for the applied aspects ratherthan theoretical. For the application purposes, it is more important that developedANN models be used for different aspects such as optimization, online controland identification, which often combine ANN models with other techniques, forinstance, fuzzy logic, expert system, and genetic algorithms or other search tech-niques. Therefore, future application of neural networks should focus on developinghybrid methods by using neural networks with other techniques. Such methods willbe of better use for industrial purposes. It can be expected that these techniqueswill make a greater contribution in food production, processing and optimizationapplications to promote food quality enhancement and public health safety.

References

[1] NeuralWare. Neural computing: A technology handbook for professionalII/PLUS and NeuralWorks Explorer. NeuralWare, Pittsburgh, PA, USA,1993.

[2] NeuralWare. Reference guide: Software reference for professional II/PLUSand NeuralWorks Explorer. NeuralWare, Pittsburgh, PA, USA, 1993.




[3] Haykin, S., Neural networks: A comprehensive foundation. Macmillan Col-lege Publishing Company, New York, NY, USA, 1994.

[4] Hopfield, J.J., Neural networks and physical systems with emergent collec-tive computational abilities. Proceedings of the National Academy of Sci-ences, USA, 79, pp. 2554–2558, 1982.

[5] Neelakanta, P.S. & Groff, D.F.D., Neural network modeling: Statisticalmechanics and cybernetic perspectives. CRC Press, Boca Raton, FL, USA,1994.

[6] Chen, C.R., Application of computer simulation and artificial intelligencetechnologies for modeling and optimization of food thermal processing. Ph.Dthesis, McGill University, Montreal, Canada, 2001.

[7] Baughman, D.R. & Liu, Y.A., Neural networks in bioprocessing and chem-ical engineering. Academic Press, 1995.

[8] Sablani, S.S., Ramaswamy, H.S. & Prasher, S.O.,A neural network approachfor thermal processing applications. Journal of Food Processing and Preser-vation, 19(4), pp. 283–301, 1995.

[9] Sablani, S.S., Ramaswamy, H.S., Sreekanth, S. & Prasher, S.O., Neural net-work modeling of heat transfer to liquid particle mixtures in cans subjected toend-over-end processing. Food Research International, 30(2), pp. 105–116,1997.

[10] Chen, C.R. & Ramaswamy, H.S., Neural computing approach for modelingof residence time distribution (RTD) of carrot cubes in a vertical scrapedsurface heat exchanger (SSHE). Food Research International, 33(7), pp.549–556, 2000.

[11] Hussain, M.A. & Rahman, M.S., Thermal conductivity prediction of fruitsand vegetables using neural networks. International Journal of Food Prop-erties, 2(2), pp. 121–137, 1999.

[12] Therdthai, N. & Zhou, W., Artificial neural network modelling of the electri-cal conductivity property of recombined milk. Journal of Food Engineering,50(2), pp. 107–111, 2000.

[13] Popescu, O., Popescu, D.C., Wilder, J. & Karwe, M.V., A new approachto modeling and control of a food extrusion process using artificial neuralnetwork and an expert system. Journal of Food Process Engineering, 24(1),pp. 17–26, 1998.

[14] Eerikainen, T., Zhu, Y.H. & Linko, P., Neural networks in extrusion processidentification and control. Food Control, 5(2), pp. 111–119, 1994.

[15] Ganjyal, G. & Hanna, M., A review on residence time distribution (RTD)in food extruders and study on the potential of neural networks in RTDmodeling. Journal of Food Science, 67(6), pp. 1996–2002, 2002.

[16] Hussain, M.A., Shafiur-Rahman, M. & Ng, C.W., Prediction of pores forma-tion (porosity) in foods during drying: Generic models by the use of hybridneural network. Journal of Food Engineering, 51(3), pp. 239–248, 2002.

[17] Chen, C.R., Ramaswamy, H.S. &Alli., I., Neural network based optimizationof quality of osmo-convective dried blue berries. FOODSIM’2000, 1st Inter-national Conference on Simulation in Food and Bio Industries, pp. 33–35,2000.




[18] Sreekanth, S., Ramaswamy, H.S. & Sablani, S., Prediction of psychometricparameters using neural networks. Drying Technology, 16(3–5), pp. 825–837, 1998.

[19] Kaminski, W., Strumillo, P. & Tomczak, E., Genetic algorithms and artificialneural networks for description of thermal processes. Drying Technology,V14, pp. 2117–2133, 1996.

[20] Kaminski, W., Stawczyk, J. & Tomczak, E., Presentation of drying kineticsin a fluidized bed by means of radial basis functions. Drying Technology,15(6/8), pp. 1753–1762, 1997.

[21] Kaminski, W., Strumillo, P. & Tomczak, E., Neuro-computing approaches tomodeling of drying process dynamics. Drying Technology, 16(6), pp. 967–992, 1998.

[22] Mittal, G.S. & Zhang, J.X., Prediction of freezing time for food productsusing a neural network. Food Research International, 33(7), pp. 557–556,2000.

[23] Sablani, S.S. & Ramaswamy, H.S., Fluid to particle heat transfer coeffi-cients in cans during end-over-end processing. Lenbensm-Wissenschaft unt-Technologie, 28, pp. 56–61, 1995.

[24] Afaghi, M., Application of artificial neural network modeling in thermal pro-cess calculations of canned foods. M.Sc. thesis, Food Science Department,McGill University, Montreal, Canada, 1999.

[25] Chen, C.R. & Ramaswamy, H.S., Dynamic modeling of retort thermal pro-cessing using neural networks. Journal of Food Processing and Preservation,26(2), pp. 91–112, 2002.

[26] Chen, C.R. & Ramaswamy, H.S., Prediction and optimization of constantretort temperature (CRT) processing using neural network and genetic algo-rithms. Journal of Food Processing Engineering, 25(5), pp. 351–380, 2002.

[27] Chen, C.R. & Ramaswamy, H.S., Prediction and optimization of variableretort temperature (VRT) processing using neural network and genetic algo-rithms. Journal of Food Engineering, 53, pp. 209–220, 2002.

[28] Chen, C.R. & Ramaswamy, H.S., Multiple ramp-variable (MRV) retort tem-perature control for optimization of thermal processing. Transactions of IChem E Part C (FBP) (accepted), 2003.

[29] Mittal, G.S. & Zhang, J.X., Use of artificial neural network to predict tem-perature, moisture, and fat in slab-shaped foods with edible coatings duringdeep-fat frying. Journal of Food Science, 65(6), pp. 978–983, 2000.

[30] Fang, Q., Bilby, G., Haque, E., Hanna, M.A. & Spillman, C.K., Neural net-work modeling of physical properties of ground wheat. Cereal Chemistry,75(2), pp. 251–253, 1998.

[31] Sablani, S.S., Baik, O.D. & Marcotte, M., Neural network prediction ofthermal conductivity of bakery products. Journal of Food Engineering, 52,pp. 299–304, 2002.

[32] Morimoto, T., De Baerdemaeker, J. & Hashimoto,Y.,An intelligent approachfor optimal control of fruit storage process using neural networks and genetic




algorithms. Computers and Electronics in Agriculture, 18, pp. 205–224,1997.

[33] Moritmoto, T., Purwanto, W., Suzuki, J. & Hashimoto, Y., Optimization ofheat treatment for fruit during storage using neural networks and geneticalgorithms. Computers and Electronics in Agriculture, 19, pp. 87–101, 1997.

[34] Trelea, I.C., Trystram, G. & Courtois, F.T.I., Optimal constrained non-linearcontrol of batch processes: application to corn drying. Journal of Food Engi-neering, 31(4), pp. 403–421, 1997.

[35] Benne, M., Grondin-Perez, B., Chabriat, J.P. & Here, P., Artificial neuralnetworks for modeling and predictive control of an industrial evaporationprocess. Journal of Food Engineering, 46(4), pp. 227–234, 2000.

[36] Huang, Y., Whittaker, A.D. & Lacey, R.E., Internal model control for a con-tinuous snack food frying process using neural networks. Transactions ofthe ASAE, 41(5), pp. 1519–1525, 1998.

[37] Kosola, A. & Linko, P., Neural control of fed-batch baker’s yeast fermenta-tion. Development of Food Science, 36, pp. 321–328, 1994.

[38] Honda, H., Hanai, T., Katayama, A., Tohyama, H. & Kobayashi, T., Temper-ature control of Ginjo sake mashing process by automatic fuzzy modelingusing fuzzy neural networks. Journal of Fermentation and Bioengineering(Japan), 85(1), pp. 107–112, 1998.

[39] Marcotte, M. & Trigi, N., Unpublished data, Agriculture and Agri-FoodCanada, St. Hyacinthe, QC, 2003.



chapter 2 neural network applications in heat and … 2 neural network applications in heat and mass...

Documents