Journal of Food Process Engineering ISSN 1745-4530

Department of Food Processing, Uludag University Karacabey Vocational School, Bursa 16700, Turkey
Department of Biosystems Engineering, Uludag University Agricultural Faculty, Bursa 16059, Turkey
Department of Food Engineering, Uludag University Agricultural Faculty, Bursa 16059, Turkey

Corresponding author. ABSTRACT
TEL: 190 224 2942662/61653;
FAX: 190 224 6765562; In this research, the aim was to predict F value (lethality or sterilization value) of
EMAIL: canned peas by using a nonlinear auto-regressive artificial neural network model
with exogenous input (NARX-ANN). During the model testing, training,
Received for Publication March 15, 2016
validation and reliability steps were followed, respectively. It was found that the
Accepted for Publication June 28, 2016
model tested was a useful tool to predict the F value for the canned foods with
doi:10.1111/jfpe.12457 high reliability. Cross-validation rules were performed for training and testing of
the model. F value of the 5 kg canned peas could be predicted with a high degree
of accuracy (R2 5 0.9982, mean square error (MSE) 5 0.1088) using training the
data yielded from 0.5 kg canned peas despite huge mass differences between
cross-validated data sets. When the same data sets were trained and tested
inversely, a high degree of prediction accuracy (R2 5 0.9914, MSE 5 0.6262) was
also observed. The model is also significant in terms of reducing the operational
costs due to the fact that higher temperatures and longer process times lead to
increased energy costs.

In this research, it was found that nonlinear auto-regressive artificial neural
network model with exogenous input is a reliable model for the prediction of
lethality rate (F value) in canned food factories. It also provides the advantage of
estimating process time more accurately in the retort and thus, reducing
operational costs.

are used to designate the heat resistance of the target micro-
F value (Lethality) is a term or measurement for determining organisms. The D value is described as the time, at a desig-
thermal process efficiency in canned foods in order to pro- nated temperature, required to reduce a determined number
duce healthy and safe foods. It is also known as a sterilization of microorganisms by 90% or to result in a 1-log reduction.
value. Sterilization is also an unavoidable heating process in This is also called the decimal reduction time because expo-
canned foods with low acidity (pH > 4.5), mainly aimed to sure for this length of time decreases the number of counts
eradicate spore forming bacteria (especially Clostridium bot- by 90%, therefore relocating the decimal point in the num-
ulinum) and vegetative forming bacteria. ber of microorganisms remaining by one place to the left. To
The conditions of heat processing required to terminate determine the D values at certain temperatures, a Z value
microorganisms in canned foods can be determined through can be determined from the slope of the line that results
thermal death time (TDT) studies. In order to compare the from plotting the log of D values versus temperature. The Z
sterilization efficiencies of heat processes, a unit of lethality value, indicative of the change in the death rate based on
or lethality rate needs to be determined. D and Z values temperature, is the number of degrees between a 10-fold
should be defined at first to determine lethality. These values change (1 log cycle) in a microorganism’s resistance. As a

Journal of Food Process Engineering 00 (2016) 00–00 V
C 2016 Wiley Periodicals, Inc. 1

2000. Province in Turkey. F.6–4. Graded peas were boiled in hot water at 90–958C by rotary 2003. fault tolerance. heating. determining the reliability of the current model was perature data are generally used to evaluate process efficien. Chen and Ramaswamy. as the investigated food mathematical models. theoretical models can be applied for the process design.PREDICTION OF LETHALITY BY NARX-ANN MODELING M. The heat 858C. al F values (integrated sterility value/lethality) by using inner ered such as container size and shape. exogenous input (NARX-ANN)) for prediction of the lethal- On the other hand. In addition. were used to predict F value by the NARX-ANN has been developed for prediction by using various process method. can modeling of inner temperatures and F values of the cans size and the food type to be canned are the most significant using the NARX-ANN model to obtain simulated data.1% computational speed. 1995. heating conditions. Sliwinska et al. Chen and age diameter were graded by means of a cylindrical sieve. cy. Chen and Ramaswamy. during the pea harvest season in May ming technologies have led to the development of new 2014. peas with 8. was very rarely consumer preferences. artificial intelligence material. filled tin cans were then thermally exhausted in a tunnel at pared to a classical regression modeling approach.18C (250 cool-down. sterilization during retort cycle follows a nonlinear plot The filled cans were immediately seamed with a COMACO. Inc. 2 Journal of Food Process Engineering 00 (2016) 00–00 V C 2016 Wiley Periodicals. There. produced considering the can sizes. Therefore. From the literature survey. citric acid. mal sterilization parameters by the NARX version of ANN zations have been done and sterilization parameters have modeling was also applied for the first time. the peas were 2008. For this reason. it is not possible to heat canned foods ity in this study. An advanced prediction model for F val. GURBUZ function of D and Z values. etc. 2014). These controls and experi. cooled by dipping into cold water. The head space was between 5 and 10 mm after filling. F value (broken bell-shaped curve) and has three phases. which directly focused on lethality concept tional and sensorial losses (such as off-flavor) in terms of during sterilization of particulate foods.5 min. robustness and poten. tested in this study. we used a nonlinear modeling (non- F) for sterilization of low acid foods (Radrigan and Ca~ numir linear auto-regressive artificial neural network model with 2009). F value determinations and temperature-time optimi. ANN modeling has a significant potential for prediction on com. There are many parameters to be consid. 2000.5 mm aver- (Afaghi et al. drained and placed into It has been stated that prediction performance of ANN the empty tin cans. been checked periodically in can plants at least once in every The purpose of this study is (1) the determination of actu- production season. time and tem. 2000c. (Tat Canned Foods Co. (lethality) can be defined as an equivalent heating of 1 min sterilization at a target temperature for a certain time and at a reference temperature. On the other hand. temperatures at the coldest spot of the canned peas. Three different can sizes (0. Production plex nonlinear systems (Lacroix et al. 2006. higher temperatures and investigated (Chen and Ramaswamy 2003) according to longer process times lead to increased energy costs. 2002b. F value can be calculated. predicted F values were and analytical concepts (Chen and Ramaswamy 2003). Ghazanfari 2014) and has been used for numerous food applications For the canned pea production. KURTULMUS AND O. Kannan and Gourisankar Sandaka. The brine solution contained 2% NaCl and 0. thermophysical prop. Sablani et al. On the other hand. 2005) and has high 85–908C. To predict F value by NARX-ANN method. The predicted values ANN is a mathematical modeling approach which simu- were compared to actual ones by the selected ANN method lates functions of the biological neurons which have a great used. (3) factors which influence F value. MATERIALS AND METHODS ue prediction can reduce the costs and time involved in F This study was established in a private canned food factory value evaluation. After conditioning for the nonlinear systems based on regression. Ramaswamy. which is generally 121.5. capacity to adopt nonlinear mechanisms. optimization and validation (Dwivedi and Ramaswamy 2010). authors’ knowledge. (2) the erties of foods. blancher for 1. at first the It has been widely acknowledged effective alternative tool model was established with actual temperature and F values compared to conventional mathematical models. In this context. simulation of ther- fore. fore. 1. Brine was poured into the tin cans at modeling was more versatile (Park et al. parameters and called as an artificial neural network (ANN). ANN modeling.0 kg) were cho- computer-based prediction models instead of complex sen for modeling and canned peas. There- thermal sterilization of canned liquid foods.0–1. Therefore. The tial for online use (Chen and Ramaswamy 2003) when com.) in the Karacabey Town of Bursa Recent developments observed in the computer program. GULDAS. After blanching.9. 2002a.0–8. The pH of the brine solution was 4. it was found that in high temperatures and long times without avoiding nutri.0 and 5. differential the model with the actual values. comparing the outputs of the NARX-ANN model with the ments are time consuming and costly processes for canned actual values and (4) predicting F values with the simulated food plants (Guldas et al. But in practice. 2008). especially which were measured for each can size. during data obtained from the ANN model (NARX-ANN). .

Tr 5 reference tem- out over five runs with nine thermocouples. many machine learning methods. the heat the model with the actual data to predict the sterility values transfer inside of the can mainly occurs by convective heat. The position of the microorganism was C. along the vertical axis from the bottom of the can for cold The actual and integrated F values were determined accord- spot measurement of canned convection-heated foods. Measurements were recorded at 30 s intervals during programmed retort cycles of the ther. Rounded needle-type thermocouple probes dicular slope (Ramaswamy and Marcotte 2006). Roedovre. three baskets were used for the experimental the D value (decimal reduction time) curve passes through cycles. 830 6 5 and 4. This system was calibrated ANNs are a subset of machine learning methods. The cans mounted with thermocouple were carefully one logarithmic cycle). KURTULMUS AND O. type A/S. considering the pH value. But. including many computa- used for the sterilization cycles was Lubeca LW 3003 ROS tional elements called neurons or nodes. which represents the temperature range between which this study. ANNs imitate central nervous systems of animals 30 s. Inc. control system. Canned (lethality) at the coldest spot of the cans. In terms of Retort Cycle a supervised learning approach. It was a water-cascading. adjusted for each batch. intervals. GURBUZ PREDICTION OF LETHALITY BY NARX-ANN MODELING AIM 4R (Italy) can seamer. botulinum and the sterilization tem- baskets in every run was also changed. the set temperature of the of parameters or category variables which can be observed retort was 1228C and the inner pressure was about 2.18C. Then. temperature measurements were F0 54 10 z dt5 (1) obtained from the liquid phase. The container sizes were 0. we statistically ing. for tor. each time period of peas was measured to determine the integrated sterility value 30 s was equal to 1 time index for the evaluation. They are prior to the experimental setup to kill the most harmful used for solving many complex problems such as function microorganisms in canned peas. ing to the formula below before the NARX-ANN model was The needles of the thermocouples were placed into the used: empty cans along the vertical axis at first. In this study. An Ellab CTF 9008 perature for the target microorganism was 121. 3 . the inner temperature of the canned mal sterilizations process. ð TðtÞ2Tr ly into the cans. neurons which correspond to features. There is Journal of Food Process Engineering 00 (2016) 00–00 V C 2016 Wiley Periodicals. Ellab. it automatically began to cool. While the experiments were per- 1 kg (99 3 119 mm) and 5 kg (154 3 242 mm). Liquid foods containing loosely packed particulates. cycle. ANNs try to use hidden relationships between the features for modeling. GULDAS.01) was used to Artificial Neural Network Modeling record temperatures in the retort. Therefore. The input layer has (Germany) with 1. middle and bot. vide convective heating. The whole retort was reloaded for each batch/ consisting of tremendous biological neurons (Smith 1996). Denmark) digital recorder with a track sense data logging system (TS version 4. The measurements were carried where F0 5 lethality/sterilization value. reflect a semilogarithmic curve thermometer in an oil bath at 908C before being mounted after a relatively short lag and reveal a much more perpen- into the cans.250 6 5 g. of other cans. Z 5 108C (temperature sensitivity indica- capacity to accommodate four baskets in each cycle. ensuring all the batches were true replicates.0 kg sizes of cans were 420 6 5. The retort has a perature (121.9. Each output value is automatically controlled horizontal retort. run. T 5 current temperature.000 kg capacity. Therefore. 0 The dimension of the retort baskets was 70 3 70 3 81 cm for the sterilization cycle. we trained peas in brine are liquid-solid mixtures. and 4.6 and were placed into the retort at the top. Contrary to the set retort temperature. Copenhagen.1C). they learn by using a series During the sterilization cycle. Data were recorded during the whole cycle at approach. The retort ANNs are constructed by layers. Thus. spaced into the retort to ensure good circulation of steam The pH value of brine in the canned peas was between 4. the targeted tom to ensure accurate measurement. The net weights for 0.M. classification and regression.5 kg (73 3 110 mm). The cycle periods were 1 h : 41 unseen data based on this learning process. After the system attained approximation. Denmark) were calibrated against a quartz like canned peas in brine. (Ellab. and then the grad- 2t 3 ed peas and the brine solution (2% NaCl) were poured slow. (GKM 13009 CO52) were used to measure the temperatures F Value Determination at the vertical axis of the cans. Convection-heated foods exhibit a broken heating curve Data Measurement between the heating and cooling periods of the sterilization The thermocouples (copper-constant. With this min : 30 s. F.0 or actually known in advance and make predictions with bars. therefore.0 and four baskets and an automatic steam heating and cooling 5. 1. equipped with represented with a neuron in the output layer. at first. The retort was used in static mode to pro- respectively. formed.5. The thermocouple is generally assembled at a point 1/10 compared predicted values with actual and measured values. the normal operation at the factory was disrupted. To increase model reliability.

jðt22Þ. . neuron weights were updated by Nonlinear Auto-Regressive With Exogenous the Levenberg-Marquardt algorithm. a general NARX model is defined: in the hidden layer. randomly split into three subsets: training (70%). Training processes were an independent or exogenous input signal (uðt21Þ. (2). and itself (yðt21Þ. Maximum iteration Input Model value was limited to 1. namely. 2014. OKO ET AL. net- from the output of the network (Lertworasirikul and Tipsu. ILLUSTRATION OF NARX- ANN MODEL USED IN THIS STUDY (BEALE ET AL. (3) can be rewritten as follows: ANN model was determined by considering the highest R2 4 Journal of Food Process Engineering 00 (2016) 00–00 V C 2016 Wiley Periodicals. 2014). neuron connection weights are updated the data corresponding to each configuration (can size) was according to a learning rule and a structure of hidden rela. A hyperbolic tan- wan 2008. Feedback information is fed to 11e22x ANN with stated variables. Þuðt2nu Þ) (Beale et al. Thus. 1. .000. . In this study. uðt22Þ. . In An ANN can be built with various numbers of the neurons Eq. . yðt2ny Þ) and previous values of the error value was set to 0. GURBUZ FIG. To eliminate the drawback of dif- A typical ANN has no connection from output to input neu. . The NARX-ANN model uses two types of inputs. And it is regressed on previous values of y(t) MSE was used as a performance function of the models. . predicted. In this study. purelinðxÞ5x (5) Tapped delay lines (TDL) hold previous values of the input (Lertworasirikul and Tipsuwan 2008. uðt2nu ÞÞ of neurons in the hidden layer for all NARX-ANN models investigated. a total of 10 neurons were used in the where y(t) is the signal of dependent variables to be hidden layer as the default network parameter of Matlab. 2014. . were used for hidden layer and output layer. (4) and (5). F. over-fitting the network was to approximate the function f with the previous values of the prevented. Beale et al. normalization was also rons. data belonging to each can size Fig. . validation error uðt22. Eq. conducted on data from other can sizes. did not reduce for five training iterations. state 2 tan sigðxÞ5 21 (4) and exogenous variables. information flow. Oko et al. . During training iterations. In the beginning of training models. 2009). Fðt2ny Þ. Tðt22Þ. 2015). The best NARX- Thus. . yðt22Þ. . Both inputs are time series. stopped when the desired MSE was reached. GULDAS. (3) Neurons transmit information to the next interconnected Tðt21Þ. or when the The NARX-ANN model uses a feed-forward neural network desired goal was reached. Fðt22Þ. ferent ranges between the variables. A simple representation of work with updated weights and biases was stored for the the NARX-ANN structure used in this study is illustrated in experiments. yðt2ny Þ. and to estimate F values for canned peas in this study. validating tionships is established inside the network. 2015). . respectively. . validating and testing NARX-ANN models. . F values were state variable to predict was used as training data at least once and experiments were and temperature values T correspond to exogenous input. FðtÞ5f ðFðt21Þ. When a NARX-ANN model was trained.PREDICTION OF LETHALITY BY NARX-ANN MODELING M. 2015) hidden layer (or layers) between input and output layers. . . For each training iteration that provides this For training. dynamic systems such as food processes applied in the range between 21 and 1 before training and involved in timed periods require feedback information testing the models. . However. A type of recurrent ANN. the net- data which are held using TDL. This study focused on predicting F value by using an actual F value obtained from the inner tempera- yðtÞ5f ðyðt21Þ. ear auto-regressive with exogenous input model (NARX. KURTULMUS AND O. tures of the can sizes reciprocally. work weights were initialized randomly. Equations of the transfer functions used in hidden layer and ANN) was used to model the thermal sterilization process output layer are given in Eqs. . 1. Inc. (15%) and testing (15%) according to cross-validation rules.001. . so we used a fixed number (2) uðt21Þ. These types of ANNs are called gent sigmoid transfer function and a linear transfer function recurrent neural networks. nonlin. . Thus. . . . Tðt2nu ÞÞ neurons through transfer or activation functions (Omid et al. . Oko et al.

0 kg canned peas were illustrated in Fig. (2) sterili- zation stage in which cans are exposed to steam at a set tem- perature for a predefined time and (3) cool-down stage in which high pressure steam is removed gradually to reach normal atmospheric pressure by air entering the chamber. During the sterilization process. rows represent training configurations and columns represent prediction performances on corre- sponding test data. each time index repre- sented a time period of 30 s. In these three plots. PERATURE (8C) AND F VALUE VERSUS TIME INDEX DURING RETORT The best cross-validation performance (R2 5 0.9993.5 kg can size was RETORT CYCLE FOR 5 KG CANS (1 TIME INDEX 5 30 S AND TOTAL used for testing. (b) THREE-DIMENSIONAL VARIATION PLOT OF TEM- control points based on the retort parameters.5 KG CANS (1 TIME INDEX 5 30 S AND TOTAL NUMBER OF TIME INDICES 5 204). This is a typical slope for convection-heated canned foods. M denotes the number of time index m. CYCLE FOR 1 KG CANS (1 TIME INDEX 5 30 S AND TOTAL NUMBER MSE 5 0. Inc. 1. modeling performance of configurations (training and testing couples) with different can sizes were investigated to reveal which sterilization configuration mod- els F values were better in order to predict other sterilization processes having different can sizes than the one used for training. The MSE and R2 values were computed by using Eqs. KURTULMUS AND O. respectively. the findings obtained from 1 kg can size were Journal of Food Process Engineering 00 (2016) 00–00 V C 2016 Wiley Periodicals. In this study. (c) THREE-DIMENSIONAL VARIATION PLOT the 1 kg can size was used for training of the NARX-ANN OF TEMPERATURE (8C) AND F VALUE VERSUS TIME INDEX DURING model. This shows the accuracy of the mod- els used for predicting F values. ta is the mean value of all outputs. 2a–c. PM m51 ðzm 2tm Þ2 MSE5 (6) M PM ðzm 2tm Þ2 R2 512 Pm51 M 2 (7) m51 ðtm 2ta Þ In Eqs. (a) THREE-DIMENSIONAL VARIATION PLOT OF TEMPERATURE (2003) have obtained similar high correlation coefficient (8C) AND F VALUE VERSUS TIME INDEX DURING RETORT CYCLE FOR (R2 > 0. F. 5 .M. GURBUZ PREDICTION OF LETHALITY BY NARX-ANN MODELING and the lowest MSE values to predict F value.0492) was obtained when the data set belonging to OF TIME INDICES 5 204). 2. In Table 1. most R2 values were higher than 0. RESULTS AND DISCUSSION The retort cycle during the sterilization of canned peas in our research consisted of three main stages: (1) heating (come-up) stage in which steam enters the sterilization retort and air is removed from the sterilization chamber. It can be distinguished eas- ily with a perpendicular ascending slope compared to con- duction heated canned foods. GULDAS. During the sterilization with real time.0 and 5. Chen and Ramaswamy FIG. (6) and (7). tm is the ground- truth (actually measured/observed data from sterilization experiment) of the mth F value.5.99. and the data belonging to the 0. According to Table 1. changes monitored regarding cold spot temperatures and F values versus time for 0. zm is predicted output of the mth F value. as inferred NUMBER OF TIME INDICES 5 204) from Table 1. (6) and (7) (Lertwor- asirikul and Tipsuwan 2008).99) from their ANN model used to evaluate critical 0.

This situation proved that there line with the cold spot temperatures until the end of the was no over-fitting in the model used. 3. a cross-validation procedure was performed and data were randomly split into training.6262 0. 3. the predicted and actual values followed quite similar time series. F VALUE PREDICTION ERROR AUTOCORRELATION FIG.5845 0.9993 0. 6. The F value slope reflected a similar ascending can size as shown in Fig. For evaluation of the trained model in terms of over-fitting. CROSS-VALIDATION PERFORMANCES OF THE NARX-ANN MODEL TO PREDICT F VALUES Test performance 0.5 kg can size. . Therefore. THE TIME SERIES RESPONSE OF NARX-ANN MODEL FOR in Fig.1088) using training data yielded from 0.9982. 4. When these the same data sets were trained and tested inversely.5 kg can size despite the huge mass difference between cross-validated data sets. GULDAS.9914 0. Despite successful results given in Table 1.0492 0.PREDICTION OF LETHALITY BY NARX-ANN MODELING M. The F value slops ascend by the model.6262) was also obtained. GURBUZ TABLE 1. Another remarkable result obtained from this study.9982 0.9974 0. 6. a high prediction accuracy (R2 5 0.9972 0. According to the training record. As it can be understood from steeply until reaching the set temperature. In Fig.1761 quite similar to the data obtained from 0.9921 0. FIG.0 kg 0.5 kg 0. TRAINING RECORD OF NARX-ANN MODEL FOR 1 KG CANS DISTRIBUTION FOR 1 KG CAN 6 Journal of Food Process Engineering 00 (2016) 00–00 V C 2016 Wiley Periodicals. F. and then remain Fig. this finding confirmed that the NARX-ANN model was suitable for modeling and prediction of different can sizes during the sterilization process. MSE 5 0.9914. for training and testing of the models.0 kg unchanged. 3. it was found that the prediction of F value for 5 kg can size was able to predict with a high accuracy (R2 5 0. 2a–c. KURTULMUS AND O. 5. the blue line The relationship between F value versus time during the and red line represent actual F value and predicted F values retort cycle was given in Fig. In this study.1724 5.0 kg 0. MSE values of PREDICTING F VALUES FOR 1 KG validation and testing followed similar vectorial behavior during training iterations of the model belonging to 1. MSE 5 0. it is important to check whether there was an over-fitting of the data.5 kg 1 kg 5 kg 2 2 Training configuration R MSE R MSE R2 MSE 0. 6. heating period as given in Fig.1088 1. the training record was calculated and investigated as illustrated FIG. Thus. Inc. validation and testing. respectively.

Therm. RAMASWAMY. Analysis of critical 2005. S. Eng.18C) X variable Y dependent variable D decimal reduction time (min). 26. R. 57. MA. 53. M. as stated earlier. Dynamic model- foods. and RAMASWAMY. Comparison of NOMENCLATURE heat transfer rates during thermal processing under end-over- end and axial modes of rotation. 20–21. C. 2010. 351–379. time at 121. S. Ghent. RELATIONSHIP BETWEEN TARGET AND PREDICTED F VALUES ny lag of the model output DURING THE RETORT CYCLE FOR 1 KG CAN nu lag of the model input M time index we could assure reliability of the current models. also provides the advantage to estimate process Food Eng. ments. C. there is a tendency to pro. T. and RAMASWAMY. 2015). 1st International from Fig. Journal of Food Process Engineering 00 (2016) 00–00 V C 2016 Wiley Periodicals. S.M. H. Modeling and achieve enough lethality was the aim in a recent study (Singh optimization of variable retort temperature (VRT) thermal et al. F. pp. 4. GURBUZ PREDICTION OF LETHALITY BY NARX-ANN MODELING T time Tr reference temperature (121. HAGAN. error distribution were not observed for F value predictions. Therefore. N number of observations FRYER. R. R. and RAMASWAMY. The time series data of F values regarding ta mean value of all outputs retort cycle during the sterilization was modeled with the TDH thermal death time NARX-ANN model as illustrated in Fig. H. long the process time of heat treatment to ensure food safety CHEN. and DEMUTH. H. H. J. 2002c. Significant autocorrelations in the Daniel Thiel.. J.. F0 lethality/sterilization value 350–360. 2003. and RAMASWAMY. C. the reduction of process time to CHEN. S. 2008). pp. and RAMASWAMY. O. control points in deviant thermal processes using artificial neu- ral networks. Inc. H. and PRASHER. 25. our research. 6. in practice. Preserv. 2014. S. ing of thermal processing using artificial neural networks. Fryer and Robbins CHEN. R. H. B. Belgium. 5. For this reason. 18–19. 43. J. processing using coupled neural networks and genetic algo- essary to avoid using longer heating time during heat treat. H. Food Eng. M. periods of longer heating inversely affects optimization of constant retort temperature (CRT) thermal the nutritional value of the canned food. M. Eng. During the train- ing iterations of the models. rithms. 25. P. LWT Food Sci. The retort zm predicted output of the mth F value cycle data obtained from the sterilization process was 1 h : tm ground-truth (actually measured/observed data) 41 min : 30 s. 2002b. it is nec. Food Proc. Edited by in confidence bounds. In this respect. this study were reliable and their generalization ability is Natick. As inferred cess calculation methods. BEALE. Modeling and in can plants. 2002a. Neural It can be concluded that the prediction models proposed in Network Toolbox User’s Guide. C. T current temperature 2499–2510. DWIVEDI. Heat transfer in food process- N number of terms ing: Ensuring product quality and safety. 225–235. 2000. Appl. P. M. the NARX-ANN model tested in processing using neural networks and genetic algorithms.. H. most of the autocorrelation errors were placed Conference on Simulation in Food and Bio Industries.18C to destroy 90% of the cells in a population Z 108C (temperature sensitivity indicator which repre- sents the temperature range between which the D value (decimal reduction time) curve passes through one logarithmic cycle) MSE mean squared error y(t) signal of dependent variables u(t) independent or exogenous input signal F function TDL tapped delay line FIG. 7 . But. 91–111. S. GULDAS. 2005. Figure 5 shows the autocorrelation errors of Artificial neural network models as alternatives to thermal pro- the best prediction model obtained in this study. J. KURTULMUS AND O. S. Food Proc. In FoodSIM’2001. adoptable for other heat sterilization processes of canned CHEN. 209–220. and ROBBINS. SCS Publications. Miri et al. The MathWorks. Inc. Technol. stopping rules were designated REFERENCES to assure obtaining models having high generalization ability AFAGHI. time efficiently (Lewis and Heppell 2000. On the other hand.

. S. Iran J. Food Eng. M. prediction. and RAMASWAMY.. OMID.. reciprocation agitation thermal processing. T... C. 2006. 2000. Y. Food analysis Food Eng. J. 87. 117–129. MD. LEWIS. RAHMAN. H. process conditions in batch thermal sterilization of food. Trans. 85. J.. J. (Principles and Applications). F. Appl. 2015.. Computer simulation of retort. 4. J. WADE. on Computers in Agriculture Conference Proceedings. 1996. M. TSOUKALAS. and crackers from drying process with artificial neural network. International Thompson Computer Press. C. P.. 153. GURBUZ GHAZANFARI. Inc. 1995. 1423–1466. J. Handbook of Food and Bioprocess Food Eng. 62. Pre. H. P. 215–420. GULDAS. R.. Taylor and Francis. HWANG. approach for predicting drum pressure and level in coal-fired pp. A statistical PARK. S. 2008. Fuel 151. J. E. A. Continuous Thermal Process. RUSTEM. 2008. WANG. 84. F. 139–145. Neural Networks for Statistical Modeling. Comparison approach to predict the sterilization value for canned olives. 1st ed.. S. J. N. Syst. Food Processing ASAE 42. Eng. M.. CRC Press. RAMASWAMY. GULDAS. E.H. thermal processing for canned food. In 7th World Congress LACROIX. 2008. and FRYER. Boston. Anim. G. cation of a static steam retort for evaluating heat transfer under Aspen Publishers. NY. S. 411–416. 59–81. 8 Journal of Food Process Engineering 00 (2016) 00–00 V C 2016 Wiley Periodicals. L. varying soil and land management conditions. MAHMOUDI. of adaptive techniques to predict crop yield response under Food Proc. Heat transfer analysis of canned food sterilization in a still RADRIGAN. S. 65–74.. A. J. 2005. S. W. 36. A. M.. S. Modeling Techniques. DYMERSKI. FL. A. LERTWORASIRIKUL. K. 11528–11535. S.. Gaithersburg. N. B.. K. S.. KURTULMUS AND O. MUJUMDAR. 213–228. and HAYES. Agric. T.S. Appl. A. NV.P. 2008. P. GONENC. M. BAKALIS. R. Global optimization of SLIWINSKA. MIRI. Agric.. Inc. pp. WISNIEWSKA. and MARCOTTE. 2014. M. and VLEK. 2009. OKO. and OMID. A. MA. Expert Syst. KANNAN. P. M. 2009. O. 320–465. and CANUMIR. SINGH. 485–494. 2015. Reno. diction of cow performance with a connectionist model. M. KOK. DATTA. 88. J. Food Eng. and GOURISANKAR SANDAKA. 2006. Moisture Boca Raton. 63–72. and HEPPELL. and TIPSUWAN. 299–316.. . and WARDENCKI. M. J.. M.PREDICTION OF LETHALITY BY NARX-ANN MODELING M. H. An intelli- ficial neural network for broiler chicken growth performance gent system for sorting pistachio nut varieties. 624 p. Neural network SMITH. 1573–1579. Food Chem. 2014. subcritical power plant. NAMIESNIK. using artificial senses. and ZHANG. A. and GURBUZ. content and water activity prediction of semi-finished cassava SABLANI. pp. 31. S. ~ J. J. Sci. PISTIKOPOULOS. SINGH. New York. Application of linear regression and arti. R. Modifi- ing of Foods: Pasteurization and UHT Sterilization.