Applying Neural Network to Dynamic Modeling of ... · consisting of an input layer, a hidden layer and an output for the biosurfactant production from soybean oil layer was selected

BUT

Corresponding Author: Tayebe Bagheri Lotfabad, National Institute of Genetic Engineering and Biotechnology (NIGEB), Shahrak-e Pajoohesh, km 15, Tehran-karaj Highway, Tehran, Iran,P.O.Box:14965/161. Tel: +98 21 44580321, Fax: +98 21 44580399, E-mail: [email protected]

1Department of Chemical and Petroleum Engineering,Sharif University of Technology, Azadi Ave.Tehran, Iran

National Institute of Genetic Engineering and Biotechnology (NIGEB), Tehran, Iran2

(Received: October 20, 2012; Accepted in Revised Form: February 7, 2013)Abstract: Biosurfactants are surface active compounds produced by various microorganisms. Production ofbiosurfactants via fermentation of immiscible wastes has the dual benefit of creating economic opportunitiesfor manufacturers, while improving environmental health. A predictor system, recommended in such processes,must be scaled-up. Hence, four neural networks were developed for the dynamic modeling of the biosurfactantproduction kinetics, in presence of soybean oil or refinery wastes including acid oil, deodorizer distillate andsoap stock. Each proposed feed forward neural network consists of three layers which are not fully connected.The input and output data for the training and validation of the neural network models were gathered frombatch fermentation experiments. The proposed neural network models were evaluated by three statistical criteria(R , RMSE and SE). The typical regression analysis showed high correlation coefficients greater than 0.971,2

demonstrating that the neural network is an excellent estimator for prediction of biosurfactant production kineticdata in a two phase liquid-liquid batch fermentation system. In addition, sensitivity analysis indicates thatresidual oil has the significant effect (i.e. 49%) on the biosurfactant in the process.

Key words: Batch fermentation Biosurfactant Dynamic modeling Neural network

INTRODUCTION modeling is solving fundamental principles in biochemical

Refinery vegetable oil wastes, the inevitable In many processes, due to the strong interactionby-products of vegetable oil production, have become between process variables and the complexity of theinexpensive and abundant sources of carbon. The dynamic system, process modeling becomes a complexworldwide surplus of these wastes has created task [6]. A black-box approach based on input and outputenvironmental problems and the economical data of the process and the relationship between them isadvantages of the vegetable oil industry have been another method for process modeling. A powerful toolgreatly affected. Given the highly reduced form of associated with this approach is the neural network thatcarbon in vegetable oils and by-products, their requires no prior knowledge of the fermentation processconversion to biosurfactants could result in yields higher [7]. Hur et al. [8] developed an artificial neural network tothan those obtained with the use of common sugars [1-3]. estimate the growth pattern of microorganisms during aRealizing this potential, however, would require the fermentation process. In another study, neural networkmodeling of a fermentation process suitable for the models were used to predict the concentration profile ofconversion of oily compounds to biosurfactants. In fact, a hydrochloric acid recovery process [9]. Koprinkova anda perfect model is usually recommended for many his colleagues [10] proposed the neural network modelingprocesses, such as optimization, control, fault detection approach for cultivation of a microorganism usingand diagnosis. One of the methods used for process chemostat is desired.

engineering [4, 5].

Applying Neural Network to Dynamic Modeling of Biosurfactant Production Using Soybean Oil Refinery Wastes

1 2 1Shokoufe Tayyebi, Tayebe Bagheri Lotfabad and Reza Roostaazad

Iranica Journal of Energy & Environment 4 (2): 161-170, 2013ISSN 2079-2115 IJEE an Official Peer Reviewed Journal of Babol Noshirvani University of TechnologyDOI: 10.5829/idosi.ijee.2013.04.02.14

161

( ) ( )( )

2 2expexp, , exp,2 1 1

2expexp,1

N Ni cal i ii i

Nii

V V V VR

V V= =

=

− − −=

−

∑ ∑∑

( )1 22

, exp,11 N

cal i iiRMSE V V

N = = − ∑

( )2, exp,11

Ncal i ii

V VSE

N=

−=

−

∑

To our best of knowledge, a comprehensive modelhas not yet been presented for estimating thebiosurfactant production kinetics in the presence ofdifferent immiscible carbon sources. Here, a dynamicneural network model was proposed to describe thebiosurfactant production process in a two-phase liquid-liquid medium containing soybean oil or refinery wastesincluding acid oil, deodorizer distillate and soap stock.

MATERIALS AND METHODS

Fermentation Kinetic Study and Analytical Methods:The microorganism, Pseudomonas aeruginosa MR01(EU795302) previously isolated from crude oil waspre-cultured overnight and cultivated in a 250 ml shakeflask [11]. SOM medium was used as the growth medium,which included 3 g/l NaNO , 0.25 g/l KH PO , 0.25 g/l

3 2 4

MgSO •7H O, 1 g/l yeast extract and one of following4 2

carbon sources at concentration of 6- 8% (v/v): soybeanoil, acid oil, deodorizer distillate or soap stock from asoybean oil refinery process. After pH adjustment to 7,flasks were incubated at 30°C, with shaking at 200 rpm [3].Chemicals were purchased from Sigma-Aldrich (St. Louis,MO, USA). Soybean oil and its by-products wereobtained from the Behshahr Industrial Company (BIC),one of the vegetable oil refineries in Iran.

To model P. aeruginosa MR01 in different broths,each containing one of the above-mentioned carbonsources, broth samples (50 ml) were withdrawn atspecified times. The number of viable bacterial cells per mlof each sample was reported as colony-forming units(CFU). Sample broths were centrifuged (4,618 ×g, 4 °C,30 min) and pellets were washed and dried to determinethe dry cell weight (DCW). Subsequently, the supernatantwas used to analyze oil consumption using the n-hexaneextraction method [12]. Briefly, a 50 ml sample of cell-freeculture was mixed vigorously with n-hexane 1:1 (v/v).The mixture was then centrifuged (7,441×g, 4°C, 20 min)to separate the phases. The organic phase was transferredto a new tube and evaporated to determine the residual oilconcentration. One half of the aqueous phase (25 ml) wasthen used for subsequent nitrate assessment, accordingto the method by Jagessar and Sooknundun [13], in whichthe ammonium hydroxide solution was replaced with 12 Npotassium hydroxide. The second half of the aqueousphase (25 ml) was used to quantify the rhamnolipidconcentration via the acid precipitation and solventextraction method involving ethyl acetate [14].

Fig. 1: Schematic of feed-forward network

Neural Network Modeling: A neural network consists ofinterconnected processing blocks of artificial neurons inwhich mathematical calculations are performed. Thenonlinear computations are carried out by transferfunctions in the neurons. The neurons are organized inlayouts comprising one input layer, hidden layer (s) andone output layer. Each neuron is connected to everyoutput from the preceding layer with weightedconnections, according to the schematic of feed-forwardnetwork illustrated in Figure 1. During the trainingalgorithm, the weights are adapted, based on theminimization of the mean square error (MSE), between thedesired and the actual-network outputs [15].

Three statistical criteria, namely correlationcoefficient (R ), root mean square error (RMSE) and2

standard error (SE) were used for assessing theperformance of the developed neural network modelsused in this study. These are defined as:

(1)

(2)

(3)

where, N is the number of observations and V and Vcalc,i exp,i

denote i output value of the neural network model andth

experimental data, respectively. Also, V is the meanexp

Iranica J. Energy & Environ., 4 (2): 161-170, 2013

162

( )( ) ( ) ( )( ) ( ) ( )

1 , 2 ,...., ,

1 , 2 ,....

y t y t y t ny t

u t u t u t m

− − −= − − −

( ) ( ) ( ) ( )( ))1 1 , 2 , Oil t-1 ( 2BS t BS t BS t Res Res Oli t− −= − − −

( )( ) ( )( )21 , 2 ,

Oil t-1 ( 2)

DCW t DCW tDCW t

Res Res Oli t− −

− −= −

( )( ) ( )( )3

1 , 2 ,

Oil t-1 ( 2)

FCU t FCU tFCU t


− −= −

( )( ) ( )

( )41 , 2 ,

Oil t-1 ( 2)

Nit t Nit tNit t


− −= − ( ) 2

2 11 xf x

e−= −

+

value of the experimental data. The small values of the the proposed neural network is that one neural networkRMSE and SE and the large value of the R indicate that should be trained for all outputs of the process, for each2

the neural network model is trained well. substrate. The numbers of neurons in the input andFor the dynamic modeling process, the feed-forward output layers are equal to 10 and 4, respectively.

neural network model is expressed as: The number of training data and inputs and outputs

(4) of hidden neurons. Moreover, it is known that the number

Where is an unknown nonlinear function which must be neurons may facilitate the training of the neural network,approximated by the neural network, t is the time, u and y it increases the computational load and memory spaceare the inputs and outputs of the process, respectively rapidly. On the other hand, a small number of hiddenand n and m are the orders of the dynamic model for the neurons make the neural network extremely difficult oroutput and the input, respectively. even impossible to be trained. Considering the above

The output and the input orders of the dynamic mentioned aspects, reaching a balance between themodel for biosurfactant production under the above number of hidden neurons and the desired performance ofdefined conditions are 2; therefore, the relationships the neural network, is a matter of great importance. Hence,between the outputs and their respective input variables in accordance to the neural network model proposed forare defined as: the modeling of the present process, the optimum number

is 8. (5) Neurons between adjacent layers, defining the neural

(6) to the fact that each of outputs is affected by some

(7) A hyperbolic tangent sigmoid transfer function with

(8)

In equations 5-8, to are nonlinear functions and

1 4

Res_Oil, BS, DCW, CFU and Nit indicate residual oil,biosurfactant, dry cell weight, colony forming unit andnitrate, respectively.

As shown in equations 5-8, for each output variable,its corresponding input variable is different unless thereare two input variables (i.e. Res_oil (t-1) and Res_Oil(t-2)). Therefore, a neural network should be developedfor each output variable if a fully connected neuralnetwork structure is selected for the prediction of theprocess model. On the other hand, four neural networksshould be developed for each substrate that can not besuitable. In order to overcome the aforementionedproblem, a partially connected feed-forward network Fig. 2: Structure of the proposed neural network modelconsisting of an input layer, a hidden layer and an output for the biosurfactant production from soybean oillayer was selected for this study. The main advantage of refinery wastes

of the neural network can significantly affect the number

of hidden neurons is crucial for successfulnetwork training. Although the large number of hidden

of hidden neurons estimated by the trial-and-error method

network structure (Figure 2), are not fully connected due

specific input (not all of them). For instance, the firstoutput (i.e. BS(t)) of the proposed neural network is afunction of four inputs, as presented in equation 5.

the following expression is applied to the hidden andoutput layers:

(9)


163

min

max min,normx x x

x x−

=−

In order to normalize the input and output between 0and 1, using equation 10 is thus essential to match theproposed transfer function.

(10)

where x , x and x are normalized, minimum andnorm min max

maximum values of variable x, respectively.

RESULTS AND DISCUSSION

The neural network used in this research wasdeveloped by MATLAB through the neural network Fig. 3: Performance error is based on the training andtoolbox. The total number of the experimental data was validation epochs100, of which 75% was used for training and the validationset and 25% was considered as the testing data. The [16]. Figure 3 shows that the individual condition isneural network was properly trained by using the gradient satisfied around iteration 1200.descent method. Data predicted by the neural network and

Variation in MSE with the number of iterations experiments with respect to time, are compared in(epochs) for training and validation of the assumed neural Figures 4(A), 4(B), 4(C) and 4(D) and correlatednetwork is presented in Figure 3.To prevent overfitting; by regression plots in Figures 5(A), 5(B), 5(C)the performance of the neural network was evaluated and 5(D). The presented results in Figures 4 and 5through the validation data after each iteration of the reveal the perfect model matching between theneural network training process. When the MSE of the predicted neural network values and the experimentalvalidation set begins to increase, the training is stopped data.

Fig. 4(A): Comparison between predicted neural network data and experimental data versus time in the presence ofsoybean oil


164

Fig. 4(B): Comparison between predicted neural network data and experimental data versus time in the presenceof acid oil

Fig. 4(C): Comparison between predicted neural network data and experimental data versus time in the presence ofdeodorizer


165

Fig. 4(D): Comparison between predicted neural network data and experimental data versus time in the presence of soapstock

Fig. 5(A): Regression analysis between predicted neural network values and experimental data in the presence ofsoybean oil


166

Fig. 5(B): Regression analysis between predicted neural network values and experimental data in the presence ofacid oil

Fig. 5(C): Regression analysis between predicted neural network values and experimental data in the presence ofdeodorizer


167

% *100%

change in outputSensitivity =change in input

Fig. 5(D): Regression analysis between predicted neural network values and experimental data in the presence ofsoap stock

Table 1: Statistical analysis between predicted neural network data and experimental data.

Soybean oil Acid oil Deodorizer Soap Stock

------------------------------------- ------------------------------------ -------------------------------- ------------------------------------Output R RMSE SE R RMSE SE R RMSE SE R RMSE SE2 2 2 2

BS 0.999 0.161 0.033 0.999 0.183 0.037 0.999 0.185 0.028 0.999 0.198 0.041DCW 0.998 0.034 0.007 0.993 0.123 0.025 0.999 0.069 0.014 0.997 0.062 0.013CFU 0.997 0.025 0.005 0.998 0.029 0.006 0.996 0.095 0.010 0.971 0.091 0.019

Nitrate 0.993 0.053 0.011 0.989 0.052 0.011 0.998 0.026 0.005 0.999 0.019 0.004

The performance of the proposed neural network During the analysis of sensitivity, an input wasmodel was evaluated by statistical parameters (Table 1), varied between the mean value ± one standard deviation,where an R value greater than 0.971 and RMSE and SE whereas other inputs were held constant at their mean2

values less than 0.198 and 0.041, respectively, indicate the value. The sensitivity of the input variables on the outputsuccess of the neural network developed for modeling the variables based on the experimental data and theproduction of biosurfactant by fermentation of soybean developed neural network model is illustrated as aoil and wastes. graphical bar in Figure 6. As shown in this figure, the

In this study, in order to evaluate the effectiveness of sensitivity results of the developed neural network modelthe input variables on the output variables, sensitivity are similar to the experimental data, with the percent erroranalysis was also conducted. Accordingly, the sensitivity of less than 4%. In addition, the residual oil with averageof each input is defined as: sensitivity of 49% has the significant effect on the

biosurfactant in comparison with other outputs.(11) Comparing the sensitivities showed higher sensitivity of

biosurfactant concentration to residual oil concentration


168

Fig. 6: Sensitivity analysis based on the proposed neural network model and the experimental data in the presence of(a) soybean oil, (b) acid oil, (c) deodorizer and (d) soap stock. (BS, biosurfactant; DCW, dry cell weight; CFU,colony forming unit)

which may concern presence of large number of fatty REFERENCESacids in biosurfactant molecule structure. In other words,this phenomenon seems to be originated from chemistry 1. Gautam, K.K. and V.K. Tyagi, 2006. Microbialof biosurfactant. surfactants: A review. J. Oleo. Sci., 55: 155-166.

CONCLUSION M.M. Resende and V.L. Cardoso, 2009. Biosurfactant

In this paper, predicted dynamic models based on the residual soybean oil. Appl. Biochem. Biotechnol.,neural network have been proposed for the production of 152: 156-168.biosurfactant by fermentation of soybean oil or refinery 3. Lotfabad, T.B., H. Abassi, R. Ahmadkhaniha,wastes, such as acid oil, deodorizer distillate and soap R. Roostaazad, F. Masoomi, H.S. Zahiri, G.stock. A three-layer neural network with 10 neurons in the Ahmadian, H. Vali and K.A. Noghabi, 2010.input layer, 8 neurons in the hidden layer and 4 neurons Structural characterization of a rhamnolipid-typein the output layer was developed for each substrate. The biosurfactant produced by Pseudomonasaccuracy of the performance of the proposed neural aeruginosa MR01: Enhancement of di-rhamnolipidnetwork for the predicted dynamic model of biosurfactant proportion using gamma irradiation. Colloids Surf. B,production was confirmed with an R of 0.971, RMSE of 81: 397-405.2

0.198 and SE of 0.041. The sensitivity analysis shows that 4. Omar, R., M.A. Abdullah, M.A. Hassan,the residual oil has the highest effect on the biosurfactant. M. Rosfarizan and M. Marziah, 2006. Kinetics andIn addition, the sensitivity results of the developed neural modeling of cell growth and substrate uptake innetwork model, as compared to those of the experimental centella asiatica cell culture. Biotechnol. Bioprocessdata, are acceptable with a percent error of less than 4%. Eng., 11: 223-229.

2. De Lima, C.J.B., E.J. Ribeiro, E.F.C. Sérvulo,

Production by Pseudomonas aeruginosa grown in


169

5. Sansonetti, S., T.J. Hobley, V. Calabrò, J. Villadsen 11. Lotfabad, T.B., M. Shourian, R. Roostaazad,and G. Sin, 2011. A biochemically structured model A.R. Najafabadi, M.R. Adelzadeh and K.A. Noghabi,for ethanol fermentation by Kluyveromyces 2009. An efficient biosurfactant-producing bacteriummarxianus: A batch fermentation and kinetic study. Pseudomonas aeruginosa MR01, isolated from oilBioresour. Technol., 102: 7513-7520. excavation areas in south of Iran. Colloids Surf B,

6. Medina-Moreno, S.A., D. Jiménez-Islas, J.N. Gracida- 69: 183-193.Rodríguez, M. Gutiérrez-Rojas and I.J. Díaz-Ramírez, 12. Müller, M.M., B. Hörmann, M. Kugel, C. Syldatk and2011. Modeling rhamnolipids production by R. Hausmann, 2011. Evaluation of rhamnolipidPseudomonas aeruginosa from immiscible carbon production capacity of Pseudomonas aeruginosasource in a batch system. Int. J. Environ. Sci. Tech., PAO1 in comparison to the rhamnolipid over-8(3): 471-482. producer strains DSM 7108 and DSM 2874. Appl.

7. Assidjo, N.E., D. Akaki, B.K. Yao and Y.T. Eboi, 2009. Microbiol. Biotechnol., 89: 585-592. A Hybrid Neural Network Approach for Batch 13. Jagessar, R.C. and L. Sooknundun,` 2011.Fermentation Simulation. Aust. J. Basic and Appl. Determination of nitrate anion in waste water fromSci., 3(4): 3930-3936. nine selected areas of coastal Guyana via a

8. Hur, W. and Y. Chung, 2006. An Artificial Neural spectrophotometric method. Int. J. Res. Rev. Appl.Network for Biomass Estimation from Automatic pH Sci., 7(2): 203-212.Control Signal. Biotechnol. Bioprocess Eng., 14. Wei, Y.H., C.L. Chou and J.S.H. Chang, 2005.11: 351-356. Rhamnolipid production by indigenous

9. Kittisupakorn, P., P. Tangteerasunun and Pseudomonas aeruginosa J4 originating fromP. Thitiyasook, 2005. Dynamic Neural Network petrochemical wastewater. Biochem. Eng. J.,Modeling for Hydrochloric Acid Recovery Process. 27: 146-154.Korean J. Chem. Eng., 226: 813-821. 15. Krose, B. and P.V.D. Smagt, 1996. An

10. Koprinkova, P., M. Petrova, T. Patarinska and introduction of neural networks. University ofM. Bliznakova, 1998. Neural Network Modelling of Amsterdam, Netherland.Fermentation Processes: Specific Kinetics Rate 16. Haykin, S., 1999. Neural Networks: A ComprehensiveModels. Cybernetics and Systems: An International Foundation. Prentice Hall Co.Journal, 29: 303-317.


170

Documents

Applying Neural Network to Dynamic Modeling of ... · consisting of an input layer, a hidden layer and an output for the biosurfactant production from soybean oil layer was selected