- Email: [email protected]
Application of artificial neural network to the determination of phenolic compounds in olive oil mill wastewater
Journal of Food Engineering 81 (2007) 544–552 www.elsevier.com/locate/jfoodeng
Application of artificial neural network to the determination of phenolic compounds in olive oil mill wastewater Jose´ S. Torrecilla a,*, Maria L. Mena b, Paloma Ya´n˜ez-Seden˜o b, Julia´n Garcı´a a a b
Department of Chemical Engineering, Faculty of Chemistry, Complutense University of Madrid, 28040 Madrid, Spain Department of Analytical Chemistry, Faculty of Chemistry, Complutense University of Madrid, 28040 Madrid, Spain Received 31 October 2006; received in revised form 2 December 2006; accepted 3 December 2006 Available online 25 January 2007
Abstract In this paper, a new computerised approach to the determination of concentrations of phenolic compounds is considered. In this approach, an integrated artificial neural network (ANN)/laccase biosensor is designed. The data collected (current signals) from amperometric detection of the laccase biosensor were transferred into an ANN trained computer for modelling and prediction of output. Such an integrated ANN/laccase biosensor system is capable of prediction of phenolic compounds concentration of olive oil mill wastewater, based on the created models and patterns, without any previous phenomenological knowledge. The predicted results using the ANN were compared with the amperometric detection of phenolic compounds obtained at a laccase biosensor in olive oil wastewater of 2004–2005 harvest season. The difference between the real and the predicted values was about 5%. Ó 2006 Elsevier Ltd. All rights reserved. Keywords: Biosensor; Olive oil mill wastewater; Phenolic compounds; Chemical analysis; Mathematical modelling
1. Introduction The extraction of olive oil is achieved through discontinuous (pressing) or continuous (centrifuging) processes in traditional mills or in modern units, respectively. Centrifugation, despite its high-water consumption (around 0.6 m3 per ton of olives processed), is still the most widely employed method for production of virgin olive oil, especially in countries that produce large amounts of olives in a short time (Benitez, Beltran-Heredia, Torregrosa, Acero, & Cercas, 1997). Usually, two by-products are obtained with either process: a solid residue and a brownish black colored effluent from the olive plus the wash water, i.e., the olive oil mill wastewater (OMW) (Borja, Alba, & Banks, 1997). This liquid effluent has a high-polluting organic load, due to a high-content of organic substances, including sugars, tannins, polyphenols, polyalcohols,
*
Corresponding author. Tel.: +34 394 42 40; fax: +34 394 42 43. E-mail address: [email protected] (J.S. Torrecilla).
0260-8774/$ - see front matter Ó 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.jfoodeng.2006.12.003
pectins and lipids (D’Annibale, Crestini, Vinciguerra, & Sermanni, 1998). This becomes a major environmental problem in the main olive-producing countries of the Mediterranean region. It is known that phenols are major contributors to the toxicity and the antibacterial activity of OMW, which limits its microbial degradability (Borja et al., 1997). Moreover, OMW may contain up to 10 g L 1 of phenols (D’Annibale et al., 1998), the maximum amount of phenols in wastewater allowed by the European Union being lower than 1 mg L 1 (Urban Water Directive 91/271/EC). Even though there is a plethora of analytical methods for the determination of phenolic compounds, including spectrometry (Bosch, Font, & Manes, 1987), high-performance liquid chromatography (Ong, Lee, & Li, 1989; Zhao & Lee, 2001), gas chromatography (Bartak, Frnkova, & Cap, 2000), and gas chromatography–mass spectrometry (GC–MS) (Angerosa, Dalessandro, Konstantinou, & Digiacinto, 1995; Tasioula-Margari & Okogeri, 2001), there is still a demand for relatively simple analytical devices, suitable for screening, and rapid assays of this type of
J.S. Torrecilla et al. / Journal of Food Engineering 81 (2007) 544–552
compounds in complex real samples, such as OMW. In this context, electrochemical biosensors with laccase as biological recognition element for the analysis of phenols have been developed by its immobilization on different electrode surfaces such as carbon fibers (Freire, Thongngamdee, Duran, Wang, & Kubota, 2002), glassy carbon (Freire, Duran, & Kubota, 2002; Leech & Daigle, 1998), graphite (Haghighi, Gorton, Ruzgas, & Jo¨nsson, 2003; Jarosz-Wilkolazka, Ruzgas, & Gorton, 2004; Kulys & Vidziunaite, 2003; Yaropolov, Kharybin, Emneus, Markovarga, & Gorton, 1995), carbon paste (Leite, Lupetti, Fatibello, Vieira, & Barbosa, 2003), polyethersulfone membranes on an universal sensors base electrode (Gomes & Rebelo, 2003), polyaniline-modified interdigitated Pt sensors (Timur, Pazarlioglu, Pilloton, & Telefoncu, 2004), PVP-gel deposited on a Clark electrode (Roy, Abraham, Abhijith, Sijithkumar, & Thakur, 2005), and gold surface (Vianello et al., 2004). This kind of determination commonly uses enzymatic reactions combined with amperometric detection of the resulting product. For phenolic compounds determination, laccase is used as the enzymatic recognition part. This enzyme is well known to reduce oxygen directly to water without intermediate formation of hydrogen peroxide in expense of oxidation of a variety of substrates, e.g. phenols. Amperometric reduction of the generated products is then used as the quantification method, by simply applying reduction potentials. Necessary reduction potentials for this process are very near to 0.0 V, which presents a great advantage as few substances interfere at this potential. The phenolic compound tested was caffeic acid because it is one of the major phenolic compounds present in OMW (GC–MS) and, because it is one of the most sensitive substrates of laccase (Haghighi et al., 2003). Because of the similarity in the produced oxidized species, amperometric signal overlapping in the reduction voltammograms is high. This fact transforms the quantification problem into a chemometric study – a case that ANN can be applied to. ANN is a mathematical algorithm which has the capability of relating the input and output parameters without requiring a prior knowledge of the relationships of the process parameters. Its structure is relatively simple, with connections in parallel and sequence between neurons. This means a short computing time and a high-potential of robustness and adaptive performance (Palancar, Arago´n, & Torrecilla, 1998). The ANN is able to model chemical processes based on linear or non-linear dynamics (Mellit, Benghanem, & Kalogirou, 2007). Focusing on the chemical processes, ANNs have been used to control/simulate nonlinear chemical process (Palancar et al., 1998), to estimate food treatment properties at high-pressure (less than 450 MPa) (Torrecilla, Otero, & Sanz, 2005), to predict variables values at relatively high temperature (less than 433 K) (Torrecilla, Arago´n, & Palancar, 2005). On the other hand, the use of ANNs in phenolic compounds quantification has also been reported in the literature (Gute´s, Ce´spedes, Alegret, & Valle, 2005; Trojanowicz, Jagielska, Rotkiewicz, & Kierzek, 1999).
545
In this paper, we present the application of ANN to the determination of phenolic compounds concentration on a laccase biosensor based on the approximation of a relationship between input and output. The proposed integrated ANN/laccase biosensor will enable the determination, of phenolic compounds concentration infield and online by an ANN trained computer. This would be very interesting for further applications to digital control, or measurement devices, which do not require any kind of mechanistic premises but only input and output variables. 2. Experimental 2.1. Reagents and solutions A stock 0.1 mol L 1 solution of 3-(3,4-dihydroxyphenyl)-2-propenoic acid (also 3,4-dihydroxycinnamic acid or caffeic acid), from Aldrich, was prepared daily by dissolving the appropriate amount methanol and was kept at 4 °C. More diluted standards were prepared by suitable dilution with the 0.1 mol L 1 citrate buffer solution (pH 5.0), which was also used as supporting electrolyte. Solutions of laccase (EC 1.10.3.2. from Trametes versicolor, Fluka, 23.75 U mg 1) were prepared in phosphate buffer solution of pH 6.5. A 4 mM dithiobis-N-succinimidylpropionate (DTSP) (Fluka), prepared in dimethyl sulfoxide (DMSO) (Panreac) and stored at 4 °C, was used in the formation of the monolayer. A 25% glutaraldehyde solution (Aldrich) was also used for laccase immobilization atop the modified electrode by cross-linking. All chemicals used were of analyticalreagent grade, and water was obtained from a Millipore Milli-Q purification system. 2.2. Instrumentation and preparation of laccase-NTSPmodified gold electrode A Metrohm 6.1204.020 gold disc electrode (AuE) (£ 3.0 mm) was used as the electrode substrate for further modification. A BAS MF-2063 Ag/AgCl/3 M KCl reference electrode and a Pt wire counter electrode were also employed. A 10 mL glass electrochemical cell was used. 2.2.1. Preparation of laccase-NTSP-modified gold electrode The pre-treatment of the gold surface before SAMs deposition is described elsewhere (Campuzano, Ga´lvez, Pedrero, Manuel de Villena, & Pingarro´n, 2002). Briefly, the gold disk electrode was polished with 3.0 lm diamond powder (BAS MF-2059) for 1 min, and immersed for 1 h in a hot 2 M KOH solution. Next, the electrode was rinsed with water, immersed in concentrated H2SO4 for 10 min, rinsed with water, immersed in concentrated HNO3 for 10 min, and rinsed again with deionised water. Finally, the electrode was dried with argon, and used immediately for the monolayer preparation. The clean AuE was immersed in a 4 mmol L 1 DTSP solution in DMSO for 1 h at room temperature (Darder,
546
J.S. Torrecilla et al. / Journal of Food Engineering 81 (2007) 544–552
Takada, Pariente, Lorenzo, & Abrun˜a, 1999). Afterwards, the electrode was thoroughly rinsed with water and finally with 0.1 mol L 1 citrate buffer (pH 5.0). Laccase was immobilized atop the modified SAM-AuE electrode by cross-linking with glutaraldehyde by deposition of 5 lL of a 1123.5 U mL 1 laccase solution on the DTSP-modified AuE, and subsequent immersion, once the electrode was dried out at ambient temperature, in a 25% glutaraldehyde solution for 1 h at 4 °C (Mena, Carralero, Gonza´lez-Corte´s, Ya´n˜ez-Seden˜o, & Pingarro´n, 2005). Amperometric measurement of phenolic compounds was carried out with an ECO Chemie Autolab PSTAT 10 potentiostat using the software package GPES 4.7 (General Purpose Electrochemical System). A P-Selecta Agimatic magnetic stirrer was also used. Amperometric sensing of phenolic compounds with the biosensor was carried out by adding different concentrations (1.0 10 7–1 10 5 mol L 1) of caffeic acid at 2 min intervals. Several data files were generated by injecting a measured amount of caffeic acid to a stirred solution in which the enzyme was immobilized on N-succinimidyl-3-thipropionate (NTSP)-modified gold electrode. Then the electrical current generated by the concentration changes was recorded with the application of a potential of +0.00 V (Mena et al., 2005). The collected amperometric data were pre-processed for the ANN learning and prediction process in order to demonstrate the ability of ANN in model prediction of the concentration of caffeic acid of the biosensor. 2.2.2. Samples OMW samples came from three different olive oil mills in Spain (Almendralejo, Badajoz, Martos, Jae´n and Villarejo de Salvane´s, Madrid), and corresponded to the 2004– 2005 harvest season. Olive oil was extracted using a centrifugation system in all cases. 2.2.3. Estimation of the total phenolic compounds content in olive oil mill wastewaters An appropriate aliquot of homogenized sample was diluted with 10 mL of citrate buffer and transferred to the electrochemical cell. Then, amperometric measurements in stirred solutions at 0.00 V using the laccase biosensor were carried out by recording the current and allowing the steady-state to be reached (Mena et al., 2005). For the biosensor, the content of phenolic compounds was estimated by applying the standard additions method, which implied the addition of five successive 20 lL aliquots from a 1.0 10 3 mol L 1 caffeic acid stock solution. For the ANN determination, the amperometric data corresponding to the addition of the sample was used to demonstrate the ability of ANN in model prediction of the concentration of phenolic compounds on the biosensor. 2.3. Artificial neural network The type of ANNs used in this work is a perceptron model also known as back-propagation perceptron. The
back-propagation perceptron is probably the most commonly used today. The ANN selected is a feed-forward network with a prediction horizon and supervised learning. It is characterised by layered architectures and feed-forward connections between neurons, or back connections, are allowed. Weights are assigned to these connections between the neurons of one layer and the following, in order to predict with the least possible error, these values must be optimized. This type of network has been selected because it is a good pattern classifier, signal filter and data compressor (Maren, Harston, & Pap, 1990). Specifically, the backpropagation multilayer perceptron is used to model systems based on non-linear dynamics (Palancar et al., 1998; Torrecilla, Otero, & Sanz, 2004, 2005; Torrecilla & Aragon et al., 2005). On the other hand, another important advantage is that knowledge of the system to be modelled is not necessary; therefore, the ANN has enormous applicability. A more detailed ANNs description can be found in the literature (Bhat & McAvoy, 1990; Parker, 1982; Psaltis, Sideris, & Yamamura, 1988; Rumelhart, Hinton, & Willians, 1986; Thibault & Grandjean, 1991; Van Breusegem, Thibault, & Che´ruy, 1991; Venkatasubramanian & Chan, 1989; Werbos, 1974). The ANN used were designed by Matlab version 7.01.24704 (R14). The statistical analyses were carried out by Statgraphics Plus (version 5.1). The ANN consists of two layers with connections to the outside world (an input layer where data are presented to the network and an output layer which holds the network response to given inputs) and one hidden layer (optimised afterwards). Given that the variables input to the ANN must be characteristic of the system, the input layer has four nodes, which corresponded to input variables (laccase units, pH value, applied potential and current signal). The output layer consists of one neuron for the output variable, i.e., the ANN predicts the caffeic acid concentration. This topology with a single hidden layer has been considered sufficient to solve similar or more complex problems (Palancar et al., 1998; Torrecilla & Otero et al., 2005; Torrecilla & Aragon et al., 2005). Moreover, more hidden layers may cause over-fitting (Ruan, Almaer, & Zhang, 1995). The transfer function used in ANN, the number of hidden neurons and other parameters of the ANN are optimised below. The learning and verification samples and processes are described below, and then, an optimisation process is exposed. 2.3.1. Learning and verification sample The learning sample was used to optimize the weights, i.e., select the adequate matrix of weights to predict, with the least possible error, the real value by the input values. On the other hand, the verification sample was used in the verification process. At this stage, the ANN uses the input values to predict the output values and no optimisation process of weights was carried out. The learning and verification samples were composed of data that characterises the process. Both samples had the
J.S. Torrecilla et al. / Journal of Food Engineering 81 (2007) 544–552
same format. These have as many rows as variables necessary to model the process and the same number of columns as the number of vectors to describe the system to measure. In this case, a pair of learning and verification samples was made. In the first case, the learning sample was composed of four rows one for each variable (laccase units, pH value, applied potential and current signal) and 2300 columns. They were 2300 vectors with dimension equal to four. The learning and verification samples show the same format. The only difference is that the verification sample has 577 columns. Taking into account that every datum (laccase units, pH value, applied potential and current signal) of the verification sample should be interpolated within learning range, the data were randomly distributed in both samples. 2.3.2. Learning and verification process of ANN The development of ANN involves two basic steps, learning and verification processes. Learning process: The learning sample set was presented to the network and a back-propagation algorithm automatically adjusted the weights; so that, the output response to input vector were as close as possible to the desired response (Ni & Gunasekaran, 1998). Each time an estimation was made, the result was compared to the corresponding desired value. Then, the estimation error (the difference between the estimated and desired values also called prediction error) was back distributed across the network in a manner that allowed the interconnection weights to be modified according to the scheme specified by the learning rule (training function). To optimise the ANN, several learning rules were tested, as described in the literature (Demuth, Beale, & Hagan, 2005). These are classified as a function of the algorithms type used to update the matrix of weights of the ANN. When the weights were modified, the next data set was fed to the network, and a new estimation was made. The estimation error was calculated again and back distributed across the network for the next modification. Simultaneously, using the verification sample, a verification test was carried out to determine the level of generalization produced by the learning set and to monitor ANN over-fitting (Ghaffari et al., 2006). When every data of the learning sample was used, an epoch was finished and other one began. To avoid the over-fitting of the neural network model, the learning process was repeated while the verification error decreased (Izadifar & Abdolahi, 2006). A matrix of weights was optimised for each training function (Demuth et al., 2005). Among other ANN parameters, the hidden neurons number (HNN) is related to the converging performance of the output error function during the learning process. The learning coefficient (Lc) controls the degree at which connection weights are modified during the learning phase. The learning coefficient decrease (Lcd) and learning coefficient increase (Lci) control the variation of Lc value. It varies as a function of performance of the ANN (the Lc decreases or increases with the mean square error).
547
Verification process: The objective of this step was to evaluate the competence of the trained network. For each training function, the matrix of weight optimised above was used. The verification sample was input into the ANN and predicted values were calculated. These were compared with the real ones to optimise the parameters of the ANN by statistical tools. In this process, no corrections of these weights were made and the ANN was only used to predict. 2.3.3. Optimisation process of the ANN In order to optimise the ANN, two stages were carried out. Firstly, the adequate training function was selected, and then, the main parameters of the ANN (using the adequate training function) were optimized. Selection of the training function: Using the adequate learning and verification samples, the training function was selected from among 14 different functions (Demuth et al., 2005). To investigate the effect of each training function, all other ANN parameters were set, i.e. the topology was 10 hidden neurons and the others were set as shown in literature (Demuth et al., 2005). With these conditions and for each training function, a learning process (using 100 epochs), and then, a verification process was developed. The predicted values were calculated in the verification process. This process was repeated to every training function, and then, all predicted values for each training function were compared one by one with the real values. These comparisons were carried out by prediction error, mean square error, statistical tests and correlation coefficient (predicted vs. real values). The statistical analysis was carried out to determine if there were significant differences between real data and those predicted by the ANN (at 95% confidence level). The null hypothesis assumes that statistical parameters of both series are equal. Otherwise, an alternative hypothesis is defined. pValue was used to check each hypothesis. Its threshold value was 0.05. If p-value is greater than this value the null hypothesis is fulfilled. Different parametric and non-parametric methods were applied based on either measures of central tendency (Kolmogorow–Smirnov test, Mann– Whitney–Wilcoxon test and Kruscal–Wallis test), on the variance (Kruscal–Wallis test, Cochran C test, Barllet’s test and Levene test) and inferential parametric test for significance (F-test and t-test). In every case, the p-values of statistical tests developed were higher than 0.05. Therefore, the mean of p-value can be calculated as a representative value. Given that the ANN must predict with the highest possible accuracy, the training function selection was carried out to obtain the least error (prediction error, mean square error) and the highest values of p-value and correlation coefficient (predicted vs. real values). Optimization of ANN parameters: Using the ANN with the training function selected above, the parameters of the ANN were optimized by experimental design. Given that the training function selected in every case was
548
J.S. Torrecilla et al. / Journal of Food Engineering 81 (2007) 544–552
TRAINLM, its parameters were optimised. The experimental design was a Screening Factorial with a centerpoint. In this design, the variables analyzed were the HNN, Lc, Lcd and Lci. The responses of experimental design were taken in the learning and verification processes. In the learning process, the epoch number necessary to optimize the matrix of weights and the mean square error were taken. In the verification process, the correlation coefficient (predicted vs. real values) and mean prediction error were used. The hidden neurons were tested between 5 and 25 neurons. When the HNN is smaller than 5, the ANN is not able to adapt to the process to be modelled. When it is higher than 25, the number of parameters to optimise in each epoch is higher than 150. The Lci was tested between 2 and 100, Lc and Lcd were tested between 0.001 and 1. These ranges were taken from Vacic (2005). A learning process is carried out in each run of the experimental design. Then, the verification process was carried out using the matrix of weights optimized in its learning process. Finally, the responses of experimental design were taken. The design was analyzed taking into account that the ANN must predict with the least prediction error and the correlation coefficient must be as close to unity as possible in the lowest iteration number. The error was measured by the prediction error and mean square error. The prediction error is the difference between an ANN’s output and real value. The mean square error is the performance function that calculates the average squared error between ANN outputs and the real value.
3. Results and discussion Firstly, two transfer functions (hyperbolical tangent and sigmoid) were tested in the ANN. In both cases, the training function was selected from among 14 algorithms (Demuth et al., 2005). The selection of adequate transfer function was carried out by comparing each one with the real values by statistical tools. Finally, the optimised ANN was validated with real values. 3.1. ANN optimisation The selection of adequate training and transfer function of the ANN and the optimization of its parameters were exposed in this section. The optimisation process of two ANNs using different transfer functions was carried out. The ANN using the hyperbolical tangent (ANNH) and sigmoid function (ANNS) as transfer function are described below. 3.1.1. Hyperbolical tangent transfer function In order to optimize the ANNH, laccase units, pH value, applied potential and current signal were input into the ANNH. The output of the ANNH was composed of caffeic acid concentration. The calculation process followed to optimize the ANNH is described in Section 2.3.3. In Fig. 1, the mean of p-values, the correlation coefficient and the mean square error versus training function are shown. The p-value of the training functions TRAINLM and TRAINBR were higher than 0.73. Focusing on these training functions, the correlation coefficients
Fig. 1. Training function selection (N correlation coefficient; d mean square error): (a) mean p-value calculated by ANNH predictions, (b) correlation coefficient and mean square error calculated by ANNH predictions, (c) mean p-value calculated by ANNS predictions and (d) correlation coefficient and mean square error calculated by ANNS predictions.
J.S. Torrecilla et al. / Journal of Food Engineering 81 (2007) 544–552
were similar (higher than 0.99) but the mean square error using TRAINLM (6.2 10 3) was the smallest. Therefore, the training function selected was TRAINLM. As mentioned above, the ANN with TRAINLM as training function was optimized by an experimental design. Taking into account the considerations described above, Lc equal to 1, Lcd equal to 0.001 and Lci equal to 2 and 25 hidden neurons were the optimal values of the ANNH, Table 1. Since that the HNN was the variable with the strongest influence on experimental design responses, it was carefully optimized. As can be seen in Fig. 2a, using 25 hidden neurons, the ANNH estimations present the least mean prediction error and one of the best correlation coefficients (0.996). Therefore, 25 hidden neurons were selected. The correlation coefficient was higher than 0.99,
Table 1 Prediction error and parameters of ANNH and ANNS Hyperbolical tangent Mean prediction error Sum of prediction error Correlation coefficient
8.53 10 4.91 10 0.996
Optimized parameters of the ANN Training function TRAINLM Hidden neurons number 25 Learning coefficient 1 Learning coefficient decrease 0.001 Learning coefficient increase 2
7 4
Sigmoid function 9.95 10 5.73 10 0.995
8 5
TRAINLM 25 1 0.001 100
549
the mean square error was 1.96 10 3 and the mean prediction error was 1.89 10 2, Table 1. On the other hand, the predicted and real values versus verification vector are shown in Fig. 3a. As can see, the predicted and real concentration values are similar. Therefore, the ANNH is able to predict concentration of caffeic acid concentration. 3.1.2. Sigmoid transfer function The only difference between ANNH (exposed above) and ANNS is the transfer function used. The calculation process was explained in ‘‘optimisation process of ANN” section. Following this method, the mean of p-values was calculated. The mean of p-values, the correlation coefficient and mean square error versus training function are shown in Figs. 1c and d, respectively. The mean of p-value of the TRAINLM training function was higher than 0.99. On the other hand, using this training function, the correlation coefficient was higher than 0.99 and the mean square error was the lowest (1.8 10 3). Therefore, the training function selected was TRAINLM. The experimental design was carried out in the same way and taking into account the same considerations as in the subsection above. The optimal values and the main results are shown in Table 1. Since the HNN was the variable with the strongest influence on experimental design responses, it was carefully optimized. As can be seen in Fig. 2b, the mean prediction error versus HNN was correlated by two regression lines (from 19 to 25 (R2 = 0.996) and from 25 to 35 (R2 = 0.956) hidden neurons). The
Normalized concentration
a
1.0
0.5
0.0
-0.5
-1.0
Normalized concentration
b
1.0 0.8 0.6 0.4 0.2 0.0 0
20
40
60
80
100
Verification vector
Fig. 2. Hidden neuron number selection: (a) hyperbolical tangent shape and (b) sigmoid shape.
Fig. 3. Predicted and real values of verification sample depending on transfer function (e, real concentration; , predicted concentration): (a) hyperbolical tangent shape and (b) sigmoid shape.
550
J.S. Torrecilla et al. / Journal of Food Engineering 81 (2007) 544–552
Table 2 Validation of optimized ANNS Almendralejo (Badajoz) Predicted value (mol L 1)
Real value (mol L 1) 1.51 10 1.40 10 1.47 10
6 6 6
1.42 10 1.32 10 1.38 10
6 6 6
Villarejo de Salvanes (Madrid)
Martos (Jae´n)
Real value (mol L 1)
Real value (mol L 1)
2.26 10 2.27 10 2.36 10
Predicted value (mol L 1) 6 6 6
regression line slope varies from 5 10 3 to 1 10 4 HNN 1, i.e. from 25 to 35 hidden neurons, the mean prediction error decreases more slowly. Given that the ANN generalization performance improves when the network size is minimized (Walczak & Cerpa, 1999), an ANNS with 25 hidden neurons was selected. The correlation coefficient was higher than 0.99, the mean square error was 4.79 10 4 and the mean prediction error was 9.95 10 3. Using the verification sample, the predicted and real values of caffeic acid concentration are shown in Fig. 3b. 3.1.3. Transfer function selection Given that the optimized ANN is used to predict caffeic acid concentration, the least prediction error was the criteria to select the adequate ANN (ANNH or ANNS). The optimized parameters and error are shown in Table 1. The correlation coefficient values were similar in both cases, the mean and the sum prediction error values (sum of prediction error values in the verification process) calculated by ANNS were less than the other one. Therefore, the ANN used to predict the polyaromatic concentration value is the ANNS. 3.2. Application of ANNS to a real case Finally, in order to validate the predictions capability of the optimised ANNS, two real samples were tested (Singleton & Rossi, 1965). These were taken from three different olive oil mills in Spain (Almendralejo, Badajoz; Martos, Jae´n; and Villarejo de Salvane´s, Madrid). The results are shown in Table 2. The predicted and real were compared by the statistical analysis and the mean p-values are 0.656, 0.791 and 0.388 to the mills of Badajoz, Jae´n and Madrid, respectively. Given that these mean p-values are higher than 0.05, there is not a statistically significant difference between the two distributions (predicted and real values) at the 95.0% confidence level. On the other hand, as can be seen in Table 2, the difference between real and predicted values in the three mills are about 5%. To sum up, the optimised ANNS is able to calculate the phenolic compounds concentration (expressed as caffeic acid) with adequate accuracy, when the ANNS is used within the range studied. Moreover, the ANN/laccase biosensor could be adapted to quantification of several analytes (caffeic acid, cathecol etc.,) and to deconvolute the
2.11 10 2.15 10 2.25 10
6 6 6
3.28 10 3.56 10 3.80 10
Predicted value (mol L 1) 6 6 6
3.11 10 3.38 10 3.62 10
6 6 6
contribution of each one. Given that this prototype is able to model the process to be measured with accuracy, it has the advantage of being more selective and more accurate than conventional methods. 4. Conclusions Therefore, the ANN with a sigmoid function as transfer function is able to predict the caffeic acid concentration. Given that the optimized ANN is used to predict caffeic acid concentration, the least prediction error was the criteria to select the adequate ANN (ANNH or ANNS). The ANNS developed and optimised was able to predict the caffeic acid concentration value in a waste stream from olive oil extraction. The optimised ANN used a TRAINLM training function (Lc of 1, Lcd of 0.001 and a Lci of 100), a topology of four input nodes (laccase units, pH value, applied potential, and current signals), 25 hidden neurons and one output (caffeic acid concentration). The difference between the real and the predicted values was about 5%. On the other hand, given the mean p-values calculated by statistical analysis in the three samples tested are 0.388, 0.791 and 0.656 (mills of Badajoz, Jae´n and Madrid, respectively), the predicted and real data were statistically similar. These confirmed that ANN is a good tool to predict the concentration of important waste chemicals, without any previous phenomenological knowledge. Acknowledgements The authors are grateful to the ‘‘Ministerio de Educacio´n y Ciencia” for financial support (Project CTQ2006-04644). Jose´ S. Torrecilla was supported by Ramo´n y Cajal research contracts from the Ministerio de Educacio´n y Ciencia in Spain. References Angerosa, F., Dalessandro, N., Konstantinou, P., & Digiacinto, L. (1995). GC-ris evaluation of phenolic-compounds in virgin olive oil. Journal of Agricultural Food Chemistry, 43(7), 1802–1807. Bartak, P., Frnkova, P., & Cap, L. (2000). Determination of phenols using simultaneous steam distillation–extraction. Journal of Chromatography A, 867(1–2), 281–287. Benitez, J., Beltran-Heredia, J., Torregrosa, J., Acero, J. L., & Cercas, V. (1997). Aerobic degradation of olive mill wastewater. Applied Microbiology and Biotechnology, 47(2), 185–188.
J.S. Torrecilla et al. / Journal of Food Engineering 81 (2007) 544–552 Bhat, N. V., & McAvoy, T. J. (1990). Use of neural nets for dynamic modelling and control of chemical process systems. Computers and Chemical Engineering, 14(4/5), 573–583. Borja, R., Alba, J., & Banks, C. J. (1997). Impact of the main phenolic compounds of olive mill wastewater (OMW) on the kinetics of acetoclastic methanogenesis. Process Biochemistry, 32(2), 121–133. Bosch, F., Font, G., & Manes, J. (1987). Ultraviolet spectrophotometric determination of phenols in natural and wastewaters with iodine monobromide. Analyst, 112(9), 1335–1337. Campuzano, S., Ga´lvez, R., Pedrero, M., Manuel de Villena, F. J., & Pingarro´n, J. M. (2002). Preparation, characterization and application of alkanethiol self-assembled monolayers modified with tetrathiafulvalene and glucose oxidase at a gold disk electrode. Journal of Electroanalytical Chemistry, 526, 92–100. D’Annibale, A., Crestini, C., Vinciguerra, V., & Sermanni, G. G. (1998). The biodegradation of recalcitrant effluents from an olive mill by a white-rot fungus. Journal of Biotechnology, 61(3), 209–218. Darder, M., Takada, K., Pariente, F., Lorenzo, E., & Abrun˜a, H. D. (1999). Dithiobissuccinimidyl propionate as an anchor for assembling peroxidases at electrodes surfaces and its application in a H2O2 biosensor. Analytical Chemistry, 71, 5530–5537. Demuth, H., Beale, M., & Hagan, M. (2005). MATLAB User’s Guide, v 4.0: Neural network toolbox. MathWorks Inc., Massahusetts, USA. Freire, R. S., Duran, N., & Kubota, L. T. (2002). Development of a laccase-based flow injection electrochemical biosensor for the determination of phenolic compounds and its application for monitoring remediation of Kraft E1 paper mill effluent. Analytica Chimica Acta, 463(2), 229–238. Freire, R. S., Thongngamdee, S., Duran, N., Wang, J., & Kubota, L. T. (2002). Mixed enzyme (laccase/tyrosinase)-based remote electrochemical biosensor for monitoring phenolic compounds. Analyst, 127(2), 258–261. Ghaffari, A., Abdollahi, H., Khoshayand, M. R., Soltani Bozchalooi, I., Dadgar, A., & Rafiee-Tehrani, M. (2006). Performance comparison of neural network training algorithms in modeling of bimodal drug delivery. International Journal of Pharmaceutics, 327(1–2), 126–138. Gomes, S. A. S. S., & Rebelo, M. J. F. (2003). A new laccase biosensor for polyphenols determination. Sensors, 3(6), 166–175. Gute´s, A., Ce´spedes, F., Alegret, S., & Valle, M. (2005). Determination of phenolic compounds by a polyphenol oxidase amperometric biosensor and artificial neural network analysis. Biosensors Bioelectronics, 20, 1668–1673. Haghighi, B., Gorton, L., Ruzgas, T., & Jo¨nsson, L. J. (2003). Characterization of graphite electrodes modified with laccase from Trametes versicolor and their use for bioelectrochemical monitoring of phenolic compounds in flow injection analysis. Analytica Chimica Acta, 487(1), 3–14. Izadifar, M., & Abdolahi, F. (2006). Comparison between neural network and mathematical modeling of supercritical CO2 extraction of black pepper essential oil. The Journal of Supercritical Fluids, 38(1), 37–43. Jarosz-Wilkolazka, A., Ruzgas, T., & Gorton, L. (2004). Use of laccasemodified electrode for amperometric detection of plant flavonoids. Enzyme and Microbial Technology, 35(2–3), 238–241. Kulys, J., & Vidziunaite, R. (2003). Amperometric biosensors based on recombinant laccases for phenols determination. Biosensors and Bioelectronics, 18(2–3), 319–325. Leech, D., & Daigle, F. (1998). Optimisation of a reagentless laccase electrode for the detection of the inhibitor azide. Analyst, 123(10), 1971–1974. Leite, O. D., Lupetti, K. O., Fatibello, O., Vieira, I. C., & Barbosa, A. D. (2003). Synergic effect studies of the bi-enzymatic system laccaseperoxidase in a voltammetric biosensor for catecholamines. Talanta, 59(5), 889–896. Maren, A. J., Harston, T., & Pap, R. P. (1990). Handbook of neural computing applications. Academic Press Inc. (pp. 323–324).
551
Mellit, A., Benghanem, M., & Kalogirou, S. A. (2007). Modeling and simulation of a stand-alone photovoltaic system using an adaptive artificial neural network: proposition for a new sizing procedure. Renewable Energy, 32(2), 285–313. Mena, M. L., Carralero, V., Gonza´lez-Corte´s, A., Ya´n˜ez-Seden˜o, P., & Pingarro´n, J. (2005). Laccase biosensor based on N-succinimidyl-3thiopropionate-functionalized gold electrodes. Electroanalysis, 17, 2147–2155. Ni, H., & Gunasekaran, S. (1998). Food quality prediction with neural networks. Food Technology, 52(10), 60–65. Ong, C. P., Lee, H. K., & Li, S. F. Y. (1989). Optimization of mobile phase-composition for high-performance liquid-chromatographic analysis of 11 priority substituted phenols. Journal of Chromatography, 464(2), 405–410. Palancar, M. C., Arago´n, J. M., & Torrecilla, J. S. (1998). pH-Control system based on artificial neural networks. Industrial and Engineering Chemistry Research, 37(7), 2729–2740. Parker, D. (1982). Learning-logic invention. Invention Report S81-64 File 1, Office of Technology Licensing, Stanford University, Stanford, CA. Psaltis, D., Sideris, A., & Yamamura, A. A. (1988). A multilayered neural network controller. IEEE Control Systems Magazine, 8(2), 17–21. Roy, J. J., Abraham, T. E., Abhijith, K. S., Sijithkumar, P. V., & Thakur, M. S. (2005). Biosensor for the determination of phenols based on cross-linked enzyme crystals (CLEC) of laccase. Biosensors and Bioelectronics, 21, 206–211. Ruan, R., Almaer, S., & Zhang, J. (1995). Prediction of dough rheological properties using neural networks. Cereal Chemistry, 72(3), 308–311. Rumelhart, D., Hinton, G., & Willians, R. (1986). Learning representations by backpropagating errors. Nature, 323, 533–536. Singleton, V. L., & Rossi, J. A. (1965). Colorimetry of total phenolics with phosphomolybdic-phosphotungstic acid reagents. American Journal of Enology and Viticulture, 16, 144–158. Tasioula-Margari, M., & Okogeri, O. (2001). Isolation and characterization of virgin olive oil phenolic compounds by HPLC/UV and GC– MS. Journal of Food Science, 66(4), 530–534. Thibault, J., & Grandjean, B. P. A. (1991). A neural network in process control. A survey. In Proceedings of the IFAC international symposium on advanced control of chemical processes, New York. Timur, S., Pazarlioglu, N., Pilloton, R., & Telefoncu, A. (2004). Thick film sensors based on laccases from different sources immobilized in polyaniline matrix. Sensors and Actuators, BChemical, 97(1), 132–136. Torrecilla, J. S., Arago´n, J. M., & Palancar, M. C. (2005). Modeling the drying of a high-moisture solid with an artificial neural network. Industrial and Engineering Chemistry Research, 44, 8057–8066. Torrecilla, J. S., Otero, L., & Sanz, P. D. (2004). A neural network approach for thermal/pressure food processing. Journal of Food Engineering, 62, 89–95. Torrecilla, J. S., Otero, L., & Sanz, P. D. (2005). Artificial neural networks: a promising tool to design and optimize high-pressure food processes. Journal of Food Engineering, 69, 299–306. Trojanowicz, M., Jagielska, A., Rotkiewicz, P., & Kierzek, A. (1999). flow-injection determination of phenols with tyrosinase amperometric biosensor and data processing by neural network. Chemia Analityczna, 44, 865–878. Urban Water Directive 91/271/EC, European Community. Vacic, V. (2005). Summary of the training functions in Matlab’s NN toolbox.. Van Breusegem, V., Thibault, J. T., & Che´ruy, A. (1991). Adaptive neural models for online prediction in fermentation. Canadian Journal of Chemical Engineering, 69, 481–487. Venkatasubramanian, V., & Chan, K. (1989). A neural Network methodology for process fault diagnosis. AICHE Journal, 35(12), 1993–2002. Vianello, F., Cambria, A., Ragusa, S., Cambria, M. T., Zennaro, L., & Rigo, A. (2004). A high sensitivity amperometric biosensor using a monomolecular layer of laccase as biorecognition element. Biosensors and Bioelectronics, 20, 315–321.
552
J.S. Torrecilla et al. / Journal of Food Engineering 81 (2007) 544–552
Walczak, S., & Cerpa, N. (1999). Heuristic principles for the design of artificial neural networks. Information and Software Technology., 41(2), 107–117. Werbos, P. (1974). Beyond regression: new tools for predictions and analysis in behavioral sciences. Ph.D. Thesis, Harvard University. Yaropolov, A. I., Kharybin, A. N., Emneus, J., Markovarga, G., & Gorton, L. (1995). Flow-injection analysis of phenols at a graphite
electrode modified with co-immobilized laccase and tyrosinase. Analytica Chimica Acta, 308(1–3), 137–144. Zhao, L. M., & Lee, H. K. (2001). Determination of phenols in water using liquid phase microextraction with back extraction combined with high-performance liquid chromatography. Journal of Chromatography A, 931(1–2), 95–105.
Application of artificial neural network to the determination of phenolic compounds in olive oil mill wastewater Jose´ S. Torrecilla a,*, Maria L. Mena b, Paloma Ya´n˜ez-Seden˜o b, Julia´n Garcı´a a a b
Department of Chemical Engineering, Faculty of Chemistry, Complutense University of Madrid, 28040 Madrid, Spain Department of Analytical Chemistry, Faculty of Chemistry, Complutense University of Madrid, 28040 Madrid, Spain Received 31 October 2006; received in revised form 2 December 2006; accepted 3 December 2006 Available online 25 January 2007
Abstract In this paper, a new computerised approach to the determination of concentrations of phenolic compounds is considered. In this approach, an integrated artificial neural network (ANN)/laccase biosensor is designed. The data collected (current signals) from amperometric detection of the laccase biosensor were transferred into an ANN trained computer for modelling and prediction of output. Such an integrated ANN/laccase biosensor system is capable of prediction of phenolic compounds concentration of olive oil mill wastewater, based on the created models and patterns, without any previous phenomenological knowledge. The predicted results using the ANN were compared with the amperometric detection of phenolic compounds obtained at a laccase biosensor in olive oil wastewater of 2004–2005 harvest season. The difference between the real and the predicted values was about 5%. Ó 2006 Elsevier Ltd. All rights reserved. Keywords: Biosensor; Olive oil mill wastewater; Phenolic compounds; Chemical analysis; Mathematical modelling
1. Introduction The extraction of olive oil is achieved through discontinuous (pressing) or continuous (centrifuging) processes in traditional mills or in modern units, respectively. Centrifugation, despite its high-water consumption (around 0.6 m3 per ton of olives processed), is still the most widely employed method for production of virgin olive oil, especially in countries that produce large amounts of olives in a short time (Benitez, Beltran-Heredia, Torregrosa, Acero, & Cercas, 1997). Usually, two by-products are obtained with either process: a solid residue and a brownish black colored effluent from the olive plus the wash water, i.e., the olive oil mill wastewater (OMW) (Borja, Alba, & Banks, 1997). This liquid effluent has a high-polluting organic load, due to a high-content of organic substances, including sugars, tannins, polyphenols, polyalcohols,
*
Corresponding author. Tel.: +34 394 42 40; fax: +34 394 42 43. E-mail address: [email protected] (J.S. Torrecilla).
0260-8774/$ - see front matter Ó 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.jfoodeng.2006.12.003
pectins and lipids (D’Annibale, Crestini, Vinciguerra, & Sermanni, 1998). This becomes a major environmental problem in the main olive-producing countries of the Mediterranean region. It is known that phenols are major contributors to the toxicity and the antibacterial activity of OMW, which limits its microbial degradability (Borja et al., 1997). Moreover, OMW may contain up to 10 g L 1 of phenols (D’Annibale et al., 1998), the maximum amount of phenols in wastewater allowed by the European Union being lower than 1 mg L 1 (Urban Water Directive 91/271/EC). Even though there is a plethora of analytical methods for the determination of phenolic compounds, including spectrometry (Bosch, Font, & Manes, 1987), high-performance liquid chromatography (Ong, Lee, & Li, 1989; Zhao & Lee, 2001), gas chromatography (Bartak, Frnkova, & Cap, 2000), and gas chromatography–mass spectrometry (GC–MS) (Angerosa, Dalessandro, Konstantinou, & Digiacinto, 1995; Tasioula-Margari & Okogeri, 2001), there is still a demand for relatively simple analytical devices, suitable for screening, and rapid assays of this type of
J.S. Torrecilla et al. / Journal of Food Engineering 81 (2007) 544–552
compounds in complex real samples, such as OMW. In this context, electrochemical biosensors with laccase as biological recognition element for the analysis of phenols have been developed by its immobilization on different electrode surfaces such as carbon fibers (Freire, Thongngamdee, Duran, Wang, & Kubota, 2002), glassy carbon (Freire, Duran, & Kubota, 2002; Leech & Daigle, 1998), graphite (Haghighi, Gorton, Ruzgas, & Jo¨nsson, 2003; Jarosz-Wilkolazka, Ruzgas, & Gorton, 2004; Kulys & Vidziunaite, 2003; Yaropolov, Kharybin, Emneus, Markovarga, & Gorton, 1995), carbon paste (Leite, Lupetti, Fatibello, Vieira, & Barbosa, 2003), polyethersulfone membranes on an universal sensors base electrode (Gomes & Rebelo, 2003), polyaniline-modified interdigitated Pt sensors (Timur, Pazarlioglu, Pilloton, & Telefoncu, 2004), PVP-gel deposited on a Clark electrode (Roy, Abraham, Abhijith, Sijithkumar, & Thakur, 2005), and gold surface (Vianello et al., 2004). This kind of determination commonly uses enzymatic reactions combined with amperometric detection of the resulting product. For phenolic compounds determination, laccase is used as the enzymatic recognition part. This enzyme is well known to reduce oxygen directly to water without intermediate formation of hydrogen peroxide in expense of oxidation of a variety of substrates, e.g. phenols. Amperometric reduction of the generated products is then used as the quantification method, by simply applying reduction potentials. Necessary reduction potentials for this process are very near to 0.0 V, which presents a great advantage as few substances interfere at this potential. The phenolic compound tested was caffeic acid because it is one of the major phenolic compounds present in OMW (GC–MS) and, because it is one of the most sensitive substrates of laccase (Haghighi et al., 2003). Because of the similarity in the produced oxidized species, amperometric signal overlapping in the reduction voltammograms is high. This fact transforms the quantification problem into a chemometric study – a case that ANN can be applied to. ANN is a mathematical algorithm which has the capability of relating the input and output parameters without requiring a prior knowledge of the relationships of the process parameters. Its structure is relatively simple, with connections in parallel and sequence between neurons. This means a short computing time and a high-potential of robustness and adaptive performance (Palancar, Arago´n, & Torrecilla, 1998). The ANN is able to model chemical processes based on linear or non-linear dynamics (Mellit, Benghanem, & Kalogirou, 2007). Focusing on the chemical processes, ANNs have been used to control/simulate nonlinear chemical process (Palancar et al., 1998), to estimate food treatment properties at high-pressure (less than 450 MPa) (Torrecilla, Otero, & Sanz, 2005), to predict variables values at relatively high temperature (less than 433 K) (Torrecilla, Arago´n, & Palancar, 2005). On the other hand, the use of ANNs in phenolic compounds quantification has also been reported in the literature (Gute´s, Ce´spedes, Alegret, & Valle, 2005; Trojanowicz, Jagielska, Rotkiewicz, & Kierzek, 1999).
545
In this paper, we present the application of ANN to the determination of phenolic compounds concentration on a laccase biosensor based on the approximation of a relationship between input and output. The proposed integrated ANN/laccase biosensor will enable the determination, of phenolic compounds concentration infield and online by an ANN trained computer. This would be very interesting for further applications to digital control, or measurement devices, which do not require any kind of mechanistic premises but only input and output variables. 2. Experimental 2.1. Reagents and solutions A stock 0.1 mol L 1 solution of 3-(3,4-dihydroxyphenyl)-2-propenoic acid (also 3,4-dihydroxycinnamic acid or caffeic acid), from Aldrich, was prepared daily by dissolving the appropriate amount methanol and was kept at 4 °C. More diluted standards were prepared by suitable dilution with the 0.1 mol L 1 citrate buffer solution (pH 5.0), which was also used as supporting electrolyte. Solutions of laccase (EC 1.10.3.2. from Trametes versicolor, Fluka, 23.75 U mg 1) were prepared in phosphate buffer solution of pH 6.5. A 4 mM dithiobis-N-succinimidylpropionate (DTSP) (Fluka), prepared in dimethyl sulfoxide (DMSO) (Panreac) and stored at 4 °C, was used in the formation of the monolayer. A 25% glutaraldehyde solution (Aldrich) was also used for laccase immobilization atop the modified electrode by cross-linking. All chemicals used were of analyticalreagent grade, and water was obtained from a Millipore Milli-Q purification system. 2.2. Instrumentation and preparation of laccase-NTSPmodified gold electrode A Metrohm 6.1204.020 gold disc electrode (AuE) (£ 3.0 mm) was used as the electrode substrate for further modification. A BAS MF-2063 Ag/AgCl/3 M KCl reference electrode and a Pt wire counter electrode were also employed. A 10 mL glass electrochemical cell was used. 2.2.1. Preparation of laccase-NTSP-modified gold electrode The pre-treatment of the gold surface before SAMs deposition is described elsewhere (Campuzano, Ga´lvez, Pedrero, Manuel de Villena, & Pingarro´n, 2002). Briefly, the gold disk electrode was polished with 3.0 lm diamond powder (BAS MF-2059) for 1 min, and immersed for 1 h in a hot 2 M KOH solution. Next, the electrode was rinsed with water, immersed in concentrated H2SO4 for 10 min, rinsed with water, immersed in concentrated HNO3 for 10 min, and rinsed again with deionised water. Finally, the electrode was dried with argon, and used immediately for the monolayer preparation. The clean AuE was immersed in a 4 mmol L 1 DTSP solution in DMSO for 1 h at room temperature (Darder,
546
J.S. Torrecilla et al. / Journal of Food Engineering 81 (2007) 544–552
Takada, Pariente, Lorenzo, & Abrun˜a, 1999). Afterwards, the electrode was thoroughly rinsed with water and finally with 0.1 mol L 1 citrate buffer (pH 5.0). Laccase was immobilized atop the modified SAM-AuE electrode by cross-linking with glutaraldehyde by deposition of 5 lL of a 1123.5 U mL 1 laccase solution on the DTSP-modified AuE, and subsequent immersion, once the electrode was dried out at ambient temperature, in a 25% glutaraldehyde solution for 1 h at 4 °C (Mena, Carralero, Gonza´lez-Corte´s, Ya´n˜ez-Seden˜o, & Pingarro´n, 2005). Amperometric measurement of phenolic compounds was carried out with an ECO Chemie Autolab PSTAT 10 potentiostat using the software package GPES 4.7 (General Purpose Electrochemical System). A P-Selecta Agimatic magnetic stirrer was also used. Amperometric sensing of phenolic compounds with the biosensor was carried out by adding different concentrations (1.0 10 7–1 10 5 mol L 1) of caffeic acid at 2 min intervals. Several data files were generated by injecting a measured amount of caffeic acid to a stirred solution in which the enzyme was immobilized on N-succinimidyl-3-thipropionate (NTSP)-modified gold electrode. Then the electrical current generated by the concentration changes was recorded with the application of a potential of +0.00 V (Mena et al., 2005). The collected amperometric data were pre-processed for the ANN learning and prediction process in order to demonstrate the ability of ANN in model prediction of the concentration of caffeic acid of the biosensor. 2.2.2. Samples OMW samples came from three different olive oil mills in Spain (Almendralejo, Badajoz, Martos, Jae´n and Villarejo de Salvane´s, Madrid), and corresponded to the 2004– 2005 harvest season. Olive oil was extracted using a centrifugation system in all cases. 2.2.3. Estimation of the total phenolic compounds content in olive oil mill wastewaters An appropriate aliquot of homogenized sample was diluted with 10 mL of citrate buffer and transferred to the electrochemical cell. Then, amperometric measurements in stirred solutions at 0.00 V using the laccase biosensor were carried out by recording the current and allowing the steady-state to be reached (Mena et al., 2005). For the biosensor, the content of phenolic compounds was estimated by applying the standard additions method, which implied the addition of five successive 20 lL aliquots from a 1.0 10 3 mol L 1 caffeic acid stock solution. For the ANN determination, the amperometric data corresponding to the addition of the sample was used to demonstrate the ability of ANN in model prediction of the concentration of phenolic compounds on the biosensor. 2.3. Artificial neural network The type of ANNs used in this work is a perceptron model also known as back-propagation perceptron. The
back-propagation perceptron is probably the most commonly used today. The ANN selected is a feed-forward network with a prediction horizon and supervised learning. It is characterised by layered architectures and feed-forward connections between neurons, or back connections, are allowed. Weights are assigned to these connections between the neurons of one layer and the following, in order to predict with the least possible error, these values must be optimized. This type of network has been selected because it is a good pattern classifier, signal filter and data compressor (Maren, Harston, & Pap, 1990). Specifically, the backpropagation multilayer perceptron is used to model systems based on non-linear dynamics (Palancar et al., 1998; Torrecilla, Otero, & Sanz, 2004, 2005; Torrecilla & Aragon et al., 2005). On the other hand, another important advantage is that knowledge of the system to be modelled is not necessary; therefore, the ANN has enormous applicability. A more detailed ANNs description can be found in the literature (Bhat & McAvoy, 1990; Parker, 1982; Psaltis, Sideris, & Yamamura, 1988; Rumelhart, Hinton, & Willians, 1986; Thibault & Grandjean, 1991; Van Breusegem, Thibault, & Che´ruy, 1991; Venkatasubramanian & Chan, 1989; Werbos, 1974). The ANN used were designed by Matlab version 7.01.24704 (R14). The statistical analyses were carried out by Statgraphics Plus (version 5.1). The ANN consists of two layers with connections to the outside world (an input layer where data are presented to the network and an output layer which holds the network response to given inputs) and one hidden layer (optimised afterwards). Given that the variables input to the ANN must be characteristic of the system, the input layer has four nodes, which corresponded to input variables (laccase units, pH value, applied potential and current signal). The output layer consists of one neuron for the output variable, i.e., the ANN predicts the caffeic acid concentration. This topology with a single hidden layer has been considered sufficient to solve similar or more complex problems (Palancar et al., 1998; Torrecilla & Otero et al., 2005; Torrecilla & Aragon et al., 2005). Moreover, more hidden layers may cause over-fitting (Ruan, Almaer, & Zhang, 1995). The transfer function used in ANN, the number of hidden neurons and other parameters of the ANN are optimised below. The learning and verification samples and processes are described below, and then, an optimisation process is exposed. 2.3.1. Learning and verification sample The learning sample was used to optimize the weights, i.e., select the adequate matrix of weights to predict, with the least possible error, the real value by the input values. On the other hand, the verification sample was used in the verification process. At this stage, the ANN uses the input values to predict the output values and no optimisation process of weights was carried out. The learning and verification samples were composed of data that characterises the process. Both samples had the
J.S. Torrecilla et al. / Journal of Food Engineering 81 (2007) 544–552
same format. These have as many rows as variables necessary to model the process and the same number of columns as the number of vectors to describe the system to measure. In this case, a pair of learning and verification samples was made. In the first case, the learning sample was composed of four rows one for each variable (laccase units, pH value, applied potential and current signal) and 2300 columns. They were 2300 vectors with dimension equal to four. The learning and verification samples show the same format. The only difference is that the verification sample has 577 columns. Taking into account that every datum (laccase units, pH value, applied potential and current signal) of the verification sample should be interpolated within learning range, the data were randomly distributed in both samples. 2.3.2. Learning and verification process of ANN The development of ANN involves two basic steps, learning and verification processes. Learning process: The learning sample set was presented to the network and a back-propagation algorithm automatically adjusted the weights; so that, the output response to input vector were as close as possible to the desired response (Ni & Gunasekaran, 1998). Each time an estimation was made, the result was compared to the corresponding desired value. Then, the estimation error (the difference between the estimated and desired values also called prediction error) was back distributed across the network in a manner that allowed the interconnection weights to be modified according to the scheme specified by the learning rule (training function). To optimise the ANN, several learning rules were tested, as described in the literature (Demuth, Beale, & Hagan, 2005). These are classified as a function of the algorithms type used to update the matrix of weights of the ANN. When the weights were modified, the next data set was fed to the network, and a new estimation was made. The estimation error was calculated again and back distributed across the network for the next modification. Simultaneously, using the verification sample, a verification test was carried out to determine the level of generalization produced by the learning set and to monitor ANN over-fitting (Ghaffari et al., 2006). When every data of the learning sample was used, an epoch was finished and other one began. To avoid the over-fitting of the neural network model, the learning process was repeated while the verification error decreased (Izadifar & Abdolahi, 2006). A matrix of weights was optimised for each training function (Demuth et al., 2005). Among other ANN parameters, the hidden neurons number (HNN) is related to the converging performance of the output error function during the learning process. The learning coefficient (Lc) controls the degree at which connection weights are modified during the learning phase. The learning coefficient decrease (Lcd) and learning coefficient increase (Lci) control the variation of Lc value. It varies as a function of performance of the ANN (the Lc decreases or increases with the mean square error).
547
Verification process: The objective of this step was to evaluate the competence of the trained network. For each training function, the matrix of weight optimised above was used. The verification sample was input into the ANN and predicted values were calculated. These were compared with the real ones to optimise the parameters of the ANN by statistical tools. In this process, no corrections of these weights were made and the ANN was only used to predict. 2.3.3. Optimisation process of the ANN In order to optimise the ANN, two stages were carried out. Firstly, the adequate training function was selected, and then, the main parameters of the ANN (using the adequate training function) were optimized. Selection of the training function: Using the adequate learning and verification samples, the training function was selected from among 14 different functions (Demuth et al., 2005). To investigate the effect of each training function, all other ANN parameters were set, i.e. the topology was 10 hidden neurons and the others were set as shown in literature (Demuth et al., 2005). With these conditions and for each training function, a learning process (using 100 epochs), and then, a verification process was developed. The predicted values were calculated in the verification process. This process was repeated to every training function, and then, all predicted values for each training function were compared one by one with the real values. These comparisons were carried out by prediction error, mean square error, statistical tests and correlation coefficient (predicted vs. real values). The statistical analysis was carried out to determine if there were significant differences between real data and those predicted by the ANN (at 95% confidence level). The null hypothesis assumes that statistical parameters of both series are equal. Otherwise, an alternative hypothesis is defined. pValue was used to check each hypothesis. Its threshold value was 0.05. If p-value is greater than this value the null hypothesis is fulfilled. Different parametric and non-parametric methods were applied based on either measures of central tendency (Kolmogorow–Smirnov test, Mann– Whitney–Wilcoxon test and Kruscal–Wallis test), on the variance (Kruscal–Wallis test, Cochran C test, Barllet’s test and Levene test) and inferential parametric test for significance (F-test and t-test). In every case, the p-values of statistical tests developed were higher than 0.05. Therefore, the mean of p-value can be calculated as a representative value. Given that the ANN must predict with the highest possible accuracy, the training function selection was carried out to obtain the least error (prediction error, mean square error) and the highest values of p-value and correlation coefficient (predicted vs. real values). Optimization of ANN parameters: Using the ANN with the training function selected above, the parameters of the ANN were optimized by experimental design. Given that the training function selected in every case was
548
J.S. Torrecilla et al. / Journal of Food Engineering 81 (2007) 544–552
TRAINLM, its parameters were optimised. The experimental design was a Screening Factorial with a centerpoint. In this design, the variables analyzed were the HNN, Lc, Lcd and Lci. The responses of experimental design were taken in the learning and verification processes. In the learning process, the epoch number necessary to optimize the matrix of weights and the mean square error were taken. In the verification process, the correlation coefficient (predicted vs. real values) and mean prediction error were used. The hidden neurons were tested between 5 and 25 neurons. When the HNN is smaller than 5, the ANN is not able to adapt to the process to be modelled. When it is higher than 25, the number of parameters to optimise in each epoch is higher than 150. The Lci was tested between 2 and 100, Lc and Lcd were tested between 0.001 and 1. These ranges were taken from Vacic (2005). A learning process is carried out in each run of the experimental design. Then, the verification process was carried out using the matrix of weights optimized in its learning process. Finally, the responses of experimental design were taken. The design was analyzed taking into account that the ANN must predict with the least prediction error and the correlation coefficient must be as close to unity as possible in the lowest iteration number. The error was measured by the prediction error and mean square error. The prediction error is the difference between an ANN’s output and real value. The mean square error is the performance function that calculates the average squared error between ANN outputs and the real value.
3. Results and discussion Firstly, two transfer functions (hyperbolical tangent and sigmoid) were tested in the ANN. In both cases, the training function was selected from among 14 algorithms (Demuth et al., 2005). The selection of adequate transfer function was carried out by comparing each one with the real values by statistical tools. Finally, the optimised ANN was validated with real values. 3.1. ANN optimisation The selection of adequate training and transfer function of the ANN and the optimization of its parameters were exposed in this section. The optimisation process of two ANNs using different transfer functions was carried out. The ANN using the hyperbolical tangent (ANNH) and sigmoid function (ANNS) as transfer function are described below. 3.1.1. Hyperbolical tangent transfer function In order to optimize the ANNH, laccase units, pH value, applied potential and current signal were input into the ANNH. The output of the ANNH was composed of caffeic acid concentration. The calculation process followed to optimize the ANNH is described in Section 2.3.3. In Fig. 1, the mean of p-values, the correlation coefficient and the mean square error versus training function are shown. The p-value of the training functions TRAINLM and TRAINBR were higher than 0.73. Focusing on these training functions, the correlation coefficients
Fig. 1. Training function selection (N correlation coefficient; d mean square error): (a) mean p-value calculated by ANNH predictions, (b) correlation coefficient and mean square error calculated by ANNH predictions, (c) mean p-value calculated by ANNS predictions and (d) correlation coefficient and mean square error calculated by ANNS predictions.
J.S. Torrecilla et al. / Journal of Food Engineering 81 (2007) 544–552
were similar (higher than 0.99) but the mean square error using TRAINLM (6.2 10 3) was the smallest. Therefore, the training function selected was TRAINLM. As mentioned above, the ANN with TRAINLM as training function was optimized by an experimental design. Taking into account the considerations described above, Lc equal to 1, Lcd equal to 0.001 and Lci equal to 2 and 25 hidden neurons were the optimal values of the ANNH, Table 1. Since that the HNN was the variable with the strongest influence on experimental design responses, it was carefully optimized. As can be seen in Fig. 2a, using 25 hidden neurons, the ANNH estimations present the least mean prediction error and one of the best correlation coefficients (0.996). Therefore, 25 hidden neurons were selected. The correlation coefficient was higher than 0.99,
Table 1 Prediction error and parameters of ANNH and ANNS Hyperbolical tangent Mean prediction error Sum of prediction error Correlation coefficient
8.53 10 4.91 10 0.996
Optimized parameters of the ANN Training function TRAINLM Hidden neurons number 25 Learning coefficient 1 Learning coefficient decrease 0.001 Learning coefficient increase 2
7 4
Sigmoid function 9.95 10 5.73 10 0.995
8 5
TRAINLM 25 1 0.001 100
549
the mean square error was 1.96 10 3 and the mean prediction error was 1.89 10 2, Table 1. On the other hand, the predicted and real values versus verification vector are shown in Fig. 3a. As can see, the predicted and real concentration values are similar. Therefore, the ANNH is able to predict concentration of caffeic acid concentration. 3.1.2. Sigmoid transfer function The only difference between ANNH (exposed above) and ANNS is the transfer function used. The calculation process was explained in ‘‘optimisation process of ANN” section. Following this method, the mean of p-values was calculated. The mean of p-values, the correlation coefficient and mean square error versus training function are shown in Figs. 1c and d, respectively. The mean of p-value of the TRAINLM training function was higher than 0.99. On the other hand, using this training function, the correlation coefficient was higher than 0.99 and the mean square error was the lowest (1.8 10 3). Therefore, the training function selected was TRAINLM. The experimental design was carried out in the same way and taking into account the same considerations as in the subsection above. The optimal values and the main results are shown in Table 1. Since the HNN was the variable with the strongest influence on experimental design responses, it was carefully optimized. As can be seen in Fig. 2b, the mean prediction error versus HNN was correlated by two regression lines (from 19 to 25 (R2 = 0.996) and from 25 to 35 (R2 = 0.956) hidden neurons). The
Normalized concentration
a
1.0
0.5
0.0
-0.5
-1.0
Normalized concentration
b
1.0 0.8 0.6 0.4 0.2 0.0 0
20
40
60
80
100
Verification vector
Fig. 2. Hidden neuron number selection: (a) hyperbolical tangent shape and (b) sigmoid shape.
Fig. 3. Predicted and real values of verification sample depending on transfer function (e, real concentration; , predicted concentration): (a) hyperbolical tangent shape and (b) sigmoid shape.
550
J.S. Torrecilla et al. / Journal of Food Engineering 81 (2007) 544–552
Table 2 Validation of optimized ANNS Almendralejo (Badajoz) Predicted value (mol L 1)
Real value (mol L 1) 1.51 10 1.40 10 1.47 10
6 6 6
1.42 10 1.32 10 1.38 10
6 6 6
Villarejo de Salvanes (Madrid)
Martos (Jae´n)
Real value (mol L 1)
Real value (mol L 1)
2.26 10 2.27 10 2.36 10
Predicted value (mol L 1) 6 6 6
regression line slope varies from 5 10 3 to 1 10 4 HNN 1, i.e. from 25 to 35 hidden neurons, the mean prediction error decreases more slowly. Given that the ANN generalization performance improves when the network size is minimized (Walczak & Cerpa, 1999), an ANNS with 25 hidden neurons was selected. The correlation coefficient was higher than 0.99, the mean square error was 4.79 10 4 and the mean prediction error was 9.95 10 3. Using the verification sample, the predicted and real values of caffeic acid concentration are shown in Fig. 3b. 3.1.3. Transfer function selection Given that the optimized ANN is used to predict caffeic acid concentration, the least prediction error was the criteria to select the adequate ANN (ANNH or ANNS). The optimized parameters and error are shown in Table 1. The correlation coefficient values were similar in both cases, the mean and the sum prediction error values (sum of prediction error values in the verification process) calculated by ANNS were less than the other one. Therefore, the ANN used to predict the polyaromatic concentration value is the ANNS. 3.2. Application of ANNS to a real case Finally, in order to validate the predictions capability of the optimised ANNS, two real samples were tested (Singleton & Rossi, 1965). These were taken from three different olive oil mills in Spain (Almendralejo, Badajoz; Martos, Jae´n; and Villarejo de Salvane´s, Madrid). The results are shown in Table 2. The predicted and real were compared by the statistical analysis and the mean p-values are 0.656, 0.791 and 0.388 to the mills of Badajoz, Jae´n and Madrid, respectively. Given that these mean p-values are higher than 0.05, there is not a statistically significant difference between the two distributions (predicted and real values) at the 95.0% confidence level. On the other hand, as can be seen in Table 2, the difference between real and predicted values in the three mills are about 5%. To sum up, the optimised ANNS is able to calculate the phenolic compounds concentration (expressed as caffeic acid) with adequate accuracy, when the ANNS is used within the range studied. Moreover, the ANN/laccase biosensor could be adapted to quantification of several analytes (caffeic acid, cathecol etc.,) and to deconvolute the
2.11 10 2.15 10 2.25 10
6 6 6
3.28 10 3.56 10 3.80 10
Predicted value (mol L 1) 6 6 6
3.11 10 3.38 10 3.62 10
6 6 6
contribution of each one. Given that this prototype is able to model the process to be measured with accuracy, it has the advantage of being more selective and more accurate than conventional methods. 4. Conclusions Therefore, the ANN with a sigmoid function as transfer function is able to predict the caffeic acid concentration. Given that the optimized ANN is used to predict caffeic acid concentration, the least prediction error was the criteria to select the adequate ANN (ANNH or ANNS). The ANNS developed and optimised was able to predict the caffeic acid concentration value in a waste stream from olive oil extraction. The optimised ANN used a TRAINLM training function (Lc of 1, Lcd of 0.001 and a Lci of 100), a topology of four input nodes (laccase units, pH value, applied potential, and current signals), 25 hidden neurons and one output (caffeic acid concentration). The difference between the real and the predicted values was about 5%. On the other hand, given the mean p-values calculated by statistical analysis in the three samples tested are 0.388, 0.791 and 0.656 (mills of Badajoz, Jae´n and Madrid, respectively), the predicted and real data were statistically similar. These confirmed that ANN is a good tool to predict the concentration of important waste chemicals, without any previous phenomenological knowledge. Acknowledgements The authors are grateful to the ‘‘Ministerio de Educacio´n y Ciencia” for financial support (Project CTQ2006-04644). Jose´ S. Torrecilla was supported by Ramo´n y Cajal research contracts from the Ministerio de Educacio´n y Ciencia in Spain. References Angerosa, F., Dalessandro, N., Konstantinou, P., & Digiacinto, L. (1995). GC-ris evaluation of phenolic-compounds in virgin olive oil. Journal of Agricultural Food Chemistry, 43(7), 1802–1807. Bartak, P., Frnkova, P., & Cap, L. (2000). Determination of phenols using simultaneous steam distillation–extraction. Journal of Chromatography A, 867(1–2), 281–287. Benitez, J., Beltran-Heredia, J., Torregrosa, J., Acero, J. L., & Cercas, V. (1997). Aerobic degradation of olive mill wastewater. Applied Microbiology and Biotechnology, 47(2), 185–188.
J.S. Torrecilla et al. / Journal of Food Engineering 81 (2007) 544–552 Bhat, N. V., & McAvoy, T. J. (1990). Use of neural nets for dynamic modelling and control of chemical process systems. Computers and Chemical Engineering, 14(4/5), 573–583. Borja, R., Alba, J., & Banks, C. J. (1997). Impact of the main phenolic compounds of olive mill wastewater (OMW) on the kinetics of acetoclastic methanogenesis. Process Biochemistry, 32(2), 121–133. Bosch, F., Font, G., & Manes, J. (1987). Ultraviolet spectrophotometric determination of phenols in natural and wastewaters with iodine monobromide. Analyst, 112(9), 1335–1337. Campuzano, S., Ga´lvez, R., Pedrero, M., Manuel de Villena, F. J., & Pingarro´n, J. M. (2002). Preparation, characterization and application of alkanethiol self-assembled monolayers modified with tetrathiafulvalene and glucose oxidase at a gold disk electrode. Journal of Electroanalytical Chemistry, 526, 92–100. D’Annibale, A., Crestini, C., Vinciguerra, V., & Sermanni, G. G. (1998). The biodegradation of recalcitrant effluents from an olive mill by a white-rot fungus. Journal of Biotechnology, 61(3), 209–218. Darder, M., Takada, K., Pariente, F., Lorenzo, E., & Abrun˜a, H. D. (1999). Dithiobissuccinimidyl propionate as an anchor for assembling peroxidases at electrodes surfaces and its application in a H2O2 biosensor. Analytical Chemistry, 71, 5530–5537. Demuth, H., Beale, M., & Hagan, M. (2005). MATLAB User’s Guide, v 4.0: Neural network toolbox. MathWorks Inc., Massahusetts, USA. Freire, R. S., Duran, N., & Kubota, L. T. (2002). Development of a laccase-based flow injection electrochemical biosensor for the determination of phenolic compounds and its application for monitoring remediation of Kraft E1 paper mill effluent. Analytica Chimica Acta, 463(2), 229–238. Freire, R. S., Thongngamdee, S., Duran, N., Wang, J., & Kubota, L. T. (2002). Mixed enzyme (laccase/tyrosinase)-based remote electrochemical biosensor for monitoring phenolic compounds. Analyst, 127(2), 258–261. Ghaffari, A., Abdollahi, H., Khoshayand, M. R., Soltani Bozchalooi, I., Dadgar, A., & Rafiee-Tehrani, M. (2006). Performance comparison of neural network training algorithms in modeling of bimodal drug delivery. International Journal of Pharmaceutics, 327(1–2), 126–138. Gomes, S. A. S. S., & Rebelo, M. J. F. (2003). A new laccase biosensor for polyphenols determination. Sensors, 3(6), 166–175. Gute´s, A., Ce´spedes, F., Alegret, S., & Valle, M. (2005). Determination of phenolic compounds by a polyphenol oxidase amperometric biosensor and artificial neural network analysis. Biosensors Bioelectronics, 20, 1668–1673. Haghighi, B., Gorton, L., Ruzgas, T., & Jo¨nsson, L. J. (2003). Characterization of graphite electrodes modified with laccase from Trametes versicolor and their use for bioelectrochemical monitoring of phenolic compounds in flow injection analysis. Analytica Chimica Acta, 487(1), 3–14. Izadifar, M., & Abdolahi, F. (2006). Comparison between neural network and mathematical modeling of supercritical CO2 extraction of black pepper essential oil. The Journal of Supercritical Fluids, 38(1), 37–43. Jarosz-Wilkolazka, A., Ruzgas, T., & Gorton, L. (2004). Use of laccasemodified electrode for amperometric detection of plant flavonoids. Enzyme and Microbial Technology, 35(2–3), 238–241. Kulys, J., & Vidziunaite, R. (2003). Amperometric biosensors based on recombinant laccases for phenols determination. Biosensors and Bioelectronics, 18(2–3), 319–325. Leech, D., & Daigle, F. (1998). Optimisation of a reagentless laccase electrode for the detection of the inhibitor azide. Analyst, 123(10), 1971–1974. Leite, O. D., Lupetti, K. O., Fatibello, O., Vieira, I. C., & Barbosa, A. D. (2003). Synergic effect studies of the bi-enzymatic system laccaseperoxidase in a voltammetric biosensor for catecholamines. Talanta, 59(5), 889–896. Maren, A. J., Harston, T., & Pap, R. P. (1990). Handbook of neural computing applications. Academic Press Inc. (pp. 323–324).
551
Mellit, A., Benghanem, M., & Kalogirou, S. A. (2007). Modeling and simulation of a stand-alone photovoltaic system using an adaptive artificial neural network: proposition for a new sizing procedure. Renewable Energy, 32(2), 285–313. Mena, M. L., Carralero, V., Gonza´lez-Corte´s, A., Ya´n˜ez-Seden˜o, P., & Pingarro´n, J. (2005). Laccase biosensor based on N-succinimidyl-3thiopropionate-functionalized gold electrodes. Electroanalysis, 17, 2147–2155. Ni, H., & Gunasekaran, S. (1998). Food quality prediction with neural networks. Food Technology, 52(10), 60–65. Ong, C. P., Lee, H. K., & Li, S. F. Y. (1989). Optimization of mobile phase-composition for high-performance liquid-chromatographic analysis of 11 priority substituted phenols. Journal of Chromatography, 464(2), 405–410. Palancar, M. C., Arago´n, J. M., & Torrecilla, J. S. (1998). pH-Control system based on artificial neural networks. Industrial and Engineering Chemistry Research, 37(7), 2729–2740. Parker, D. (1982). Learning-logic invention. Invention Report S81-64 File 1, Office of Technology Licensing, Stanford University, Stanford, CA. Psaltis, D., Sideris, A., & Yamamura, A. A. (1988). A multilayered neural network controller. IEEE Control Systems Magazine, 8(2), 17–21. Roy, J. J., Abraham, T. E., Abhijith, K. S., Sijithkumar, P. V., & Thakur, M. S. (2005). Biosensor for the determination of phenols based on cross-linked enzyme crystals (CLEC) of laccase. Biosensors and Bioelectronics, 21, 206–211. Ruan, R., Almaer, S., & Zhang, J. (1995). Prediction of dough rheological properties using neural networks. Cereal Chemistry, 72(3), 308–311. Rumelhart, D., Hinton, G., & Willians, R. (1986). Learning representations by backpropagating errors. Nature, 323, 533–536. Singleton, V. L., & Rossi, J. A. (1965). Colorimetry of total phenolics with phosphomolybdic-phosphotungstic acid reagents. American Journal of Enology and Viticulture, 16, 144–158. Tasioula-Margari, M., & Okogeri, O. (2001). Isolation and characterization of virgin olive oil phenolic compounds by HPLC/UV and GC– MS. Journal of Food Science, 66(4), 530–534. Thibault, J., & Grandjean, B. P. A. (1991). A neural network in process control. A survey. In Proceedings of the IFAC international symposium on advanced control of chemical processes, New York. Timur, S., Pazarlioglu, N., Pilloton, R., & Telefoncu, A. (2004). Thick film sensors based on laccases from different sources immobilized in polyaniline matrix. Sensors and Actuators, BChemical, 97(1), 132–136. Torrecilla, J. S., Arago´n, J. M., & Palancar, M. C. (2005). Modeling the drying of a high-moisture solid with an artificial neural network. Industrial and Engineering Chemistry Research, 44, 8057–8066. Torrecilla, J. S., Otero, L., & Sanz, P. D. (2004). A neural network approach for thermal/pressure food processing. Journal of Food Engineering, 62, 89–95. Torrecilla, J. S., Otero, L., & Sanz, P. D. (2005). Artificial neural networks: a promising tool to design and optimize high-pressure food processes. Journal of Food Engineering, 69, 299–306. Trojanowicz, M., Jagielska, A., Rotkiewicz, P., & Kierzek, A. (1999). flow-injection determination of phenols with tyrosinase amperometric biosensor and data processing by neural network. Chemia Analityczna, 44, 865–878. Urban Water Directive 91/271/EC, European Community. Vacic, V. (2005). Summary of the training functions in Matlab’s NN toolbox.
552
J.S. Torrecilla et al. / Journal of Food Engineering 81 (2007) 544–552
Walczak, S., & Cerpa, N. (1999). Heuristic principles for the design of artificial neural networks. Information and Software Technology., 41(2), 107–117. Werbos, P. (1974). Beyond regression: new tools for predictions and analysis in behavioral sciences. Ph.D. Thesis, Harvard University. Yaropolov, A. I., Kharybin, A. N., Emneus, J., Markovarga, G., & Gorton, L. (1995). Flow-injection analysis of phenols at a graphite
electrode modified with co-immobilized laccase and tyrosinase. Analytica Chimica Acta, 308(1–3), 137–144. Zhao, L. M., & Lee, H. K. (2001). Determination of phenols in water using liquid phase microextraction with back extraction combined with high-performance liquid chromatography. Journal of Chromatography A, 931(1–2), 95–105.