Monitoring of solid-state fermentation of protein feed by electronic nose and chemometric analysis

Process Biochemistry 49 (2014) 583–588

Contents lists available at ScienceDirect

Process Biochemistry journal homepage: www.elsevier.com/locate/procbio

Short communication

Monitoring of solid-state fermentation of protein feed by electronic nose and chemometric analysis Hui Jiang a,∗ , Quansheng Chen b , Guohai Liu a a b

School of Electrical and Information Engineering, Jiangsu University, Zhenjiang 212013, PR China School of Food and Biological Engineering, Jiangsu University, Zhenjiang 212013, PR China

a r t i c l e

i n f o

Article history: Received 18 October 2013 Received in revised form 25 December 2013 Accepted 11 January 2014 Available online 23 January 2014 Keywords: Electronic nose Protein feed Solid-state fermentation Sensors Chemometric analysis

a b s t r a c t To achieve the real-time smell monitoring of solid-state fermentation (SSF) of protein feed associated with its degree of fermentation. Electronic nose (e-nose) technique, with the help of chemometric analysis, was attempted in this study. Linear discriminant analysis (LDA), K-nearest neighbors (KNN), and support vector machines (SVM) were respectively used to calibrate discrimination models in order to evaluate the influences of different linear and non-linear classification algorithms on the identification results. Experimental results showed that the predictive precision of SVM model was superior to those of the others two, and the optimum SVM model was obtained when five PCs were included. The discrimination rates of the SVM model were 97.14% and 91.43% in the training and testing sets, respectively. The overall results sufficiently demonstrate excellent promise for the e-nose technique combined with an appropriate chemometric method to be applied in the SSF industry. © 2014 Elsevier Ltd. All rights reserved.

1. Introduction Agricultural residues are the most abundant resource in China with an annual production rate of about 700 million tons which can be microbiologically converted into protein feed products for animals [1]. Most of the agricultural by-products are poor in nutrition such as protein and vitamin and are rich in fiber with low digestibility and are not suitable for non-ruminant animals [2]. Under such circumstances, a potential solution is available by the utilization of microorganisms to convert agricultural wastes into obtain products with higher nutritive value and digestibility, especially in regard to protein content. Bioconversion of these materials by the process of solid-state fermentation (SSF) is often used due to its low effluent generation, requirement for simple fermentation equipment and the direct applicability of the fermented product for feeding [3,4]. SSF is a complex process in which raw materials are transformed into high-value product (e.g. agricultural residue into protein feed). The fermentation process is, however, sensitive to many factors, which can cause the target product to deteriorate. Smells of the products are vital factors in the fermentation industry and subsequently much time and effort are spent on methods that can estimate and measure these factors [5]. In fermentation, only gas chromatography, in some cases combined with mass spectrometry is commonly used [6–9]. Head space gas chromatography combined with mass spectrometry

∗ Corresponding author. Tel.: +86 0511 88791960; fax: +86 0511 88780088. E-mail address: [email protected] (H. Jiang). 1359-5113/$ – see front matter © 2014 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.procbio.2014.01.006

(GC/MS) can identify and quantify the composition of volatile organic compounds (VOCs), giving a complex pattern. The problem is that even if the composition is known, it is still difficult to couple this pattern to the quality of the fermentation. Thus, it appears that a simple but yet powerful objective method for the description of fermentation parameters related to smell would be very valuable. There is today a great interest in using an electronic nose (e-nose) system for detection and discrimination of volatile compounds. The sensor array in the e-nose system consists of some non-specific sensors, and an odor stimulus generates a characteristic fingerprint from the sensor array [10]. The e-nose technique as an increasingly fast, reliable and robust technology has been successfully applied in different fields such as food [11–14], clinical diagnostics [15–17], pharmaceutical [18], environmental control [19–21]. Recently, the e-nose technique, most noteworthy, has been also employed in recognition and quality analysis of various fermentation [22–27]. These studies mentioned above show that the e-nose technique has high potential in discrimination and quality monitoring of the fermentation process. However, little attention, up to the present, has been reported on the process monitoring of SSF of protein feed by the use of e-nose technique; additionally, the experiments were performed in the commercial e-nose instrument; and they have also not systemically studied different linear and non-linear discrimination algorithms in the solution to the identification of fermenting degrees by the use of the e-nose data. The e-nose device usually consists of an array of different metal oxide semiconductor (MOS) gas sensors with overlapping sensitivities toward volatile gas components. Normally, the gas sensors

584

H. Jiang et al. / Process Biochemistry 49 (2014) 583–588

have non-specific sensitivity toward volatile gas components in SSF of protein feed. Thus, there are significant correlations existing in two different gas sensor data, and this data information is often called the redundant information. The performance of the discrimination model based on e-nose data might be weakened because of involving too much redundant information. However, this problem can be effectively solved with the help of multivariate data calibration technique. In multivariate calibration, a discrimination model based on supervised pattern recognition method (that is, a method with a priori knowledge about the category membership of samples) is used for identification [28]. First, a discrimination model is developed by the use of a training set of samples with known categories; then, the model performance is evaluated by means of the independent samples from a testing set by comparing the prediction with the true categories [29]. Supervised pattern recognition methods are numerous, and the crucial problem is to select the most appropriate method. In this study, we evaluate the use of an e-nose technique for monitoring of the fermented process of SSF of protein feed by calibrating an identification model. The specific research work was arranged as follows: (1) an e-nose system based on the gas sensor array was developed for data acquisition of fermented samples; (2) latent variables were extracted from the responses of gas sensors and (3) three different linear and non-linear discrimination tools, which were linear discriminant analysis (LDA), K-nearest neighbors (KNN) and support vector machine (SVM), were used to develop the discrimination models, respectively.

Storage Lab of Jiangsu University. Based on the fermentation, a series of trials were carried out using a number of commercially available metals oxide semiconductor (MOS) gas sensors. From the response sensitivity of individual sensor toward the VOCs of fermented materials, a set of eleven gas sensors from Figaro Co. Ltd., Japan (i.e. universal sensors of TGS2602, TGS2610, TGS2611, TGS813, and TGS822, special sensors of TGS822TF, TGS825, TGS826, TGS880, TGS4160, and TGS5042) were eventually selected for odor capture in the SSF process of protein feed. In the e-nose system, the gas sensors are very sensitive to temperature, which strongly influences the electrical properties of MOS. Thus, a temperature sensor was embedded in the sensor chamber. There is an adjustable temperature controller existing in the e-nose system which can control the sampling temperature of the sensor array. During the experiment, the ambient temperature is pre-set at 22 ◦ C in the laboratory. In order to maintain the temperature at around 22 ◦ C in the sensor chamber, the temperature controller begins to work while the temperature in the sensor chamber exceeds the ambient temperature, i.e. 22 ◦ C, and cannot stop until the temperature is going back to 22 ◦ C. In addition, the humidity is kept at a level of 65% in the laboratory. The e-nose experimental cycle consists of the automated sequence of internal operations for each sample contained the following three stages: (i) headspace generation, (ii) sampling, and (iii) purifying. The experimental conditions of the e-nose system in this study are given as follows: amount of each fermented sample = 6 g, temperature = 22 ± 1 ◦ C, headspace generation time = 120 s, sampling time = 120 s, purging time = 180 s, and airflow rate = 3 mL/s. The maximum value of the response of each sensor was extracted as the latent variables. Thus, eleven latent variables can be obtained from raw e-nose data, which were marked as p1 , p2 , . . ., p11 , respectively. 2.3. Software Software of e-nose data acquisition was compiled by us based on Delphi 7 (Borland, Scotts Valley, USA). All algorithms were implemented in PASW Statistics 18 (IBM, New York, USA) and Matlab R2010a (Mathworks, Natick, USA) under Windows 7 in data processing.

3. Results and discussion 2. Materials and methods

3.1. Principal component analysis (PCA) 2.1. Sample preparation Samples were prepared at different times from four runs of SSF of protein feed trials. The rice chaff was obtained from the Zhenjiang city of Jiangsu Province of China. It was mixed with corn flour and wheat bran (rice chaff:corn flour:wheat bran was 7:2:1), and then the mixtures were ground by use of a crushing machine with 40 mesh screen. Finally, the basal substrates, which were made up of mixture of effective microorganisms (EM) bacterial liquid, water and the pretreated mixtures in the ratio of 1:200:500, were loaded in GTG-100 bioreactor (100 L) with a 40% occupancy coefficient of the bioreactor volume, and cultures were incubated by anaerobic fermentation at 30 ± 2 ◦ C for six days. The samples were obtained from the four runs of fermentation, and in each run, 35 samples were collected. Every day, 5 samples were taken out for signal acquisition of the e-nose system, thus 140 samples date were obtained in this process. In this study, all 140 fermented samples were divided into two subsets (i.e. training set and testing set). The training set contained 105 samples from the first three runs of fermentation experiments, and the remaining 35 samples from the last run of fermentation trial constituted the testing set. 2.2. Data acquisition and latent variable extraction The original data were obtained by the use of an electronic nose (e-nose) system, which was designed and developed by the Agricultural Product Processing and

In the e-nose system, eleven non-specific metal oxide semiconductor gas sensors have cross-sensitivity toward VOCs in fermented substrate. Thus, there are collinear variables existing in the eleven latent variables, which also can be proved by the use of the results of Pearson correlation analysis (see Table 1). This problem could be solved with the help of principal component analysis (PCA). PCA is a method of describing the unique variances in a set of original variables using linear combinations of the original variables (principal components, PCs), and these PCs are orthogonal [30]. In this study, PCA was performed on eleven latent variables, and several of the PCs were extracted as the input of supervised pattern recognition. To visualize the cluster trends of all 140 samples, a scatter plot (also called a score plot) was obtained using the top two PCs issued from PCA based on eleven latent variables. Fig. 1 shows a two-dimensional (2D) space of all fermented samples represented by PC1 and PC2. Investigated from Fig. 1, seven sample groups appeared in cluster trend along two principal component axes, confirming the presence of seven different clusters

Table 1 Pearson correlations among eleven latent variables. Variables

p1 p2 p3 p4 p5 p6 p7 p8 p9 p10 p11 a b

Pearson correlations p1

p2

p3

1

−0.742 1

a

p4 a

0.727 −0.234a 1

Correlation is significant at the 0.01 level (2-tailed). Correlation is significant at the 0.05 level (2-tailed).

p5

−0.357 0.554a −0.280 1 a

p6 a

0.791 −0.303a 0.942a 0.122 1

p7 a

0.729 −0.205b 0.963a 0.118 0.984a 1

p8

−0.275 0.732a 0.377a 0.738a 0.347a 0.430a 1 a

p9 a

0.483 0.053 0.897a 0.391a 0.900a 0.927a 0.688a 1

p10 a

0.234 −0.243a 0.136 −0.070 0.154 0.138 −0.117 0.083 1

p11 a

0.640 −0.653a 0.530a −0.675a 0.383a 0.374a −0.372a 0.211b 0.230a 1

−0.240a 0.705a 0.388a 0.730a 0.383a 0.461a 0.979a 0.686a −0.145 −0.433a 1

H. Jiang et al. / Process Biochemistry 49 (2014) 583–588

585

Fig. 1. Score cluster plot with the top two principal components (PCs) for all 140 samples.

just associated with their degree of fermentation (i.e. day 0, day 1, day 2, day 3, day 4, day 5, and day 6). PC1 can explain 45.38% variance, and PC2 can explain 37.55% variance. The cumulative variance contribution rate of PC1 and PC2 was 82.93%. In other word, the 2D representation of the PC1 and PC2 scores for all samples can interpret 82.93% information from the eleven latent variables, which covered most of useful information. Moreover, as can be seen from Fig. 1, samples from “day 0”, “day 1”, and “day 3” could be separated directly by PCA. Nevertheless, the separation of the samples from “day 2” and “day 4” was not clear, especially, more overlap can be observed from the samples of “day 5” and “day 6”. It shows that the samples of “day 5” is similar to the samples of “day 6” in their internal ingredients, and can be inferred that the fermentation process has already finished basically when the fermentation to the fifth day. 3.2. Discrimination models of supervised pattern recognition The geometrical exploration of 2D plot by PCA can give the cluster trend of all 140 fermented samples and cannot be used as a discrimination tool. The aim of this study is to discriminate the degree of fermentation by means of the e-nose technique with the help of the supervised pattern recognition tools. Consequently, The key issue is to select the suitable identification tool to build a discrimination model. In this study, three linear or non-linear supervised pattern recognition algorithms (i.e. LDA, KNN, and SVM) were attempted to construct the discrimination models, respectively. 3.2.1. Linear discriminant analysis (LDA) LDA is applied to find the linear combination of latent vectors, which best separate two or more classes of object or event. The resulting combinations may be used as a linear classifier, or more commonly in dimensionality reduction before later classification. This method maximizes the ratio of between-class variance to the within-class variance in any particular data set thereby guaranteeing maximal separability [31]. In LDA model calibration, the number of PCs is crucial to the predictive precision of the LDA discrimination model. Thus, in this study, the number of PCs was optimized by five-fold cross-validation in calibrating LDA model, and the optimum number of PCs was determined according to the

Fig. 2. The parameter optimization of different discriminant models by crossvalidation. (a) Discrimination rates of LDA model with the different number of PCs; (b) discrimination rates of the KNN model with different PCs and K values; (c) results of grid search for the optimal C and  values of the SVM model with RBF kernel.

highest discrimination rate in training set by the cross-validation. Fig. 2a shows the discrimination rates of LDA model in training set by the cross-validation with the different number of PCs. Investigated from Fig. 2a, the maximum discrimination rate of LDA model by the cross-validation was found when the corresponding number of PCs is seven. Thus, the optimum LDA model was obtained when

586

H. Jiang et al. / Process Biochemistry 49 (2014) 583–588

Table 2 Results and comparison of the best LDA, KNN, and SVM models. Model

LDA KNN SVM a b c

Optimum parameters

Discrimination results

PCs

Other parameters

Training set

7 7 5

– K = 3a C = 256b ,  = 0.1768c

Testing

Ratio

Percentage

Ratio

Percentage

93/105 96/105 102/105

88.57 91.43 97.14

26/35 30/35 32/35

74.29 85.71 91.43

K: parameter K of KNN algorithm. C: penalty parameter of SVM algorithm. : width parameter of RBF-kernel function.

seven PCs were included, and the discrimination rates were 88.57% and 74.29% in the training and testing sets, respectively. 3.2.2. K-nearest neighbors (KNN) The K-nearest neighbor (KNN) method was first described by Fix and Hodges [32], and the method is a non-parametric supervised pattern recognition method that places the objects of the testing set in the same multi-dimensional space as those of the training set. One problem with this method is that, in principle, there exists an optimum choice for the value of the parameter K, which describes the best performance of the classifier. The number of PCs also should be optimized in this study. Therefore, 10 K values (K = 1, 2, . . ., 10) and 11 PCs (PCs = 1, 2, . . ., 11) were simultaneously optimized by five-fold cross-validation. Fig. 2b shows the discrimination rates in the training set by the cross-validation with the different number of PCs and the various values of parameter K. As shown in Fig. 2b, the maximum discrimination rate in the training set by the cross-validation is 91.43% when K = 3 and PCs = 7. Thus, the optimum KNN model was achieved when the value of parameter K was equal to three and seven PCs were included, and the discrimination rates were 91.43% and 85.71% in the training and testing sets, respectively. 3.2.3. Support vector machines (SVM) Considering that the linear supervised pattern recognition method may not provide a complete solution to the classification problem, non-linear approach, support vector machines (SVM) was used in this study. The original SVM algorithm was invented by Vapnik and the current standard incarnation was proposed by Cortes and Vapnik [33]. The basic concept of SVM is mapping the original data set into a high or infinite dimensional feature space, and then constructing a hyperplane or set of hyperplanes which can be used for classification. The transformation into higher-dimensional

space is implemented by a kernel function [34]. Selection of kernel function has a high influence on the performance of the SVM discrimination model. The RBF kernel nonlinearly map samples into a higher dimensional space, so it, unlike the linear kernel, can handle the case when the relation between class labels and attributes is nonlinear [35]. Therefore, the RBF kernel function was used in this study. Through the implementation of PCA, the accumulated variance contribution rate was up to 99.04% for the top five PCs. Therefore, the top five PCs were inputted into the SVM classifiers as characteristic variables. In order to obtain a good performance, the penalty parameter C and kernel parameter  while using RBF kernels in SVM model have to be optimized. In this study, a “grid-search” on parameter C and  was used by five-fold crossvalidation. Basically pairs of (C, ) were tried and the one with the best identification accuracy by the cross-validation was picked. The practice showed that trying to exponentially grow sequences of C and  is a practical method to identify good parameters (in this study, C = 2−10 , 2−9 , . . ., 29 , 210 , respectively and  = 2−5 , 2−4.5 , . . ., 24.5 , 25 , respectively). Fig. 2c shows a contour map for the optimum C and  values of the SVM model with RBF kernel with different discrimination rates obtained by the cross-validation based on the training set, and discrimination rates are indicated by different color. It can be seen from Fig. 2c, the optimum combination of C and  were obtained with [256,0.1768] when the maximum discrimination rate is 97.14% of the training set by the cross-validation, which was marked by a blue asterisk as shown in Fig. 2c. The discrimination rates of the optimum SVM model were 97.14% and 91.43% in the training and testing sets, respectively. 3.3. Discussion of discrimination results In order to obtain a good predictive precision in the identification of fermentation degree by the e-nose technique, the

Table 3 Detailed discrimination results of the best SVM model. Subsets

Training set

Testing set

Sample number

Fermentation time

Detailed discrimination results

Discrimination rate/%

Day 0

Day 1

Day 2

Day 3

Day 4

Day 5

Day 6

0 15 0 0 0 0 0

0 0 15 0 0 0 0

0 0 0 15 0 0 0

0 0 0 0 15 0 0

0 0 0 0 0 14 2

0 0 0 0 0 1 13

97.14

0 5 0 0 0 0 0

0 0 5 0 0 0 0

0 0 0 5 0 0 0

0 0 0 0 5 0 0

0 0 0 0 0 3 1

0 0 0 0 0 2 4

91.43

15 15 15 15 15 15 15

Day 0 Day 1 Day 2 Day 3 Day 4 Day 5 Day 6

15 0 0 0 0 0 0

5 5 5 5 5 5 5

Day 0 Day 1 Day 2 Day 3 Day 4 Day 5 Day 6

5 0 0 0 0 0 0

H. Jiang et al. / Process Biochemistry 49 (2014) 583–588

multivariate calibration and parameter optimization have been systematically studied in model calibration. Table 2 shows the discrimination results of LDA, KNN, and SVM models in the training and testing sets. As seen from Table 2, the discrimination results of SVM model are superior to those of LDA and KNN models. Seen from the number of input variable (i.e. PCs) of different models, compared with LDA and KNN models, SVM model has a simple structure due to the minimum PCs included in the model (i.e. five PCs). As for the reasons why e-nose technique combined with SVM classification tool could obtain such good discrimination results, the main reasons could be summarized from the following two aspects. Firstly, on the principles of e-nose system, in this study, the e-nose consists of eleven MOS gas sensors with different selectivities. During the fermentation, the composition of fermented materials could be changed with the consumption of solid-state substrate for microbial growth and product synthesization, which result in great changes in the composition of VOCs in bioreactor that can be related to some certain responses by the eleven MOS gas sensors. Thus, the differences between the odor of two samples with different fermentation degrees can be present in the response curve of e-nose system. Although these differences in the e-nose response curve are difficult to visible only by naked eyes, they can be differentiated with the help of an appropriate chemometric method. Secondly, on the principles of statistical learning theory, the algorithm of SVM has its own unique advantages in contrast to the other two algorithms (i.e. LDA and KNN) in this study. SVM is a non-linear pattern recognition algorithm; while, LDA and KNN are linear ones. Considering that microbial SSF of protein feed is a complex process which involves the growth of microorganisms during fermentation, the response curve differences of gas sensors as fermentation degree are also very complex; so the linear tools might not provide a complete solution to such classification problem. Generally, non-linear approach is stronger than linear method in the level of self-learning and self-adjust; besides, the topological architecture of SVM might be more suitable for the solution to this classification problem in this study. Therefore, non-linear model often has a simpler structure and a higher predictive precision of discrimination. Table 3 shows the detailed discrimination results of the optimum SVM model. As can be seen from Table 3, misclassifications often occur between “day 5” samples and “day 6” samples in both the training set and testing set. As for the reasons, we can infer that there are only slight differences existing in “day 5” samples and “day 6” samples so that some samples cannot be differentiated by the e-nose technique. Usually, along with the fermentation, the whole process can be divided into three stages, i.e. lag phase, exponential phase, and stationary phase. These changes of fermented substrate tissue are very slow or even stagnant because the fermentation has entered into stationary phase when the fermentation to the fifth day. Therefore, the misclassifications only occur between “day 5” samples and “day 6” samples.

4. Conclusions The overall results sufficiently demonstrate that the e-nose technique coupled with an appropriate chemometric method could be successfully used in monitoring of SSF process. Three different linear and non-linear discrimination tools (LDA, KNN, and SVM) were attempted comparatively to develop the discrimination model in this study. Among the three discrimination models, the predictive precision of SVM model is superior to those of the others two models. It can be concluded that e-nose technique coupled with SVM classification tool has high potential to monitor other fermentation process in a non-invasive way.

587

Acknowledgements The authors gratefully acknowledge the financial support provided by the Medium and Small Scale Enterprises Innovation of China (Grant no. 12C26213202207), and the Advanced Talents Science Foundation of Jiangsu University (Grant No. 13JDG094). We also wish to thank many of our colleagues for many stimulating discussions in this field. Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at doi:10.1016/j.procbio.2014.01.006. References [1] Jiang H, Liu G, Xiao X, Mei C, Ding Y, Yu S. Monitoring of solid-state fermentation of wheat straw in a pilot scale using FT-NIR spectroscopy and support vector data description. Microchem J 2012;102:68–74. [2] Rajesh N, Imelda-Joseph, Paul Raj R. Value addition of vegetable wastes by solidstate fermentation using Aspergillus niger for use in aquafeed industry. Waste Manage 2010;30:2223–7. [3] Yang X, Chen H, Gao H, Li Z. Bioconversion of corn straw by coupling ensiling and solid-state fermentation. Bioresour Technol 2001;78:277–80. [4] Jiang H, Liu G, Mei C, Yu S, Xiao X, Ding Y. Rapid determination of pH in solid-state fermentation of wheat straw by FT-NIR spectroscopy and efficient wavelengths selection. Anal Bioanal Chem 2012;404:603–11. [5] Eklöv T, Johansson G, Winquist F, Lundström I. Monitoring sausage fermentation using an electronic nose. J Sci Food Agric 1998;76:525–32. [6] Kim Y, Goodner KL, Park J-D, Choi J, Talcott ST. Changes in antioxidant phytochemicals and volatile composition of Camellia sinensis by oxidation during tea fermentation. Food Chem 2011;129:1331–42. [7] Gómez García-Carpintero E, Gómez Gallego MA, Sánchez-Palomo E, González ˜ MA. Impact of alternative technique to ageing using oak chips in alcoholic Vinas or in malolactic fermentation on volatile and sensory composition of red wines. Food Chem 2012;134:851–63. [8] Hyˇspler R, Tichá A, Indrová M, Zadák Z, Hyˇsplerová L, Gaspariˇc J, et al. A simple, optimized method for the determination of sulphide in whole blood by GC–MS as a marked of bowel fermentation processes. J Chromatogr B 2002;770:255–9. [9] Mallouchos A, Komaitis M, Koutinas A, Kanellaki M. Wine fermentations by immobilized and freecells at different temperatures. Effect of immobilization and temperature on volatile by-products. Food Chem 2003;80:109–13. [10] Peris M, Escuder-Gilabert L. A 21st century technique for food control: Electronic noses. Anal Chim Acta 2009;638:1–15. [11] Chen Q, Zhao J, Chen Z, Lin H, Zhao D-A. Discrimination of green tea quality using the electronic nose technique and the human panel test, comparison of linear and nonlinear classification tools. Sensor Actuat B: Chem 2011;159:294–300. [12] Concina I, Falasconi M, Gobbi E, Bianchi F, Musci M, Mattarozzi M, et al. Early detection of microbial contamination in processed tomatoes by electronic nose. Food Control 2009;20:873–80. [13] Capone S, Epifani M, Quaranta F, Siciliano P, Taurino A, Vasanelli L. Monitoring of rancidity of milk by means of an electronic nose and a dynamic PCA analysis. Sensor Actuat B: Chem 2001;78:174–9. [14] Panigrahi S, Balasubramanian S, Gu H, Logue C, Marchello M. Neural-networkintegrated electronic nose system for identification of spoiled beef. LWT – Food Sci Technol 2006;39:135–45. [15] Lin Y-J, Guo H-R, Chang Y-H, Kao M-T, Wang H-H, Hong R-I. Application of the electronic nose for uremia diagnosis. Sensor Actuat B: Chem 2001;76:177–80. [16] Pavlou AK, Magan N, Jones JM, Brown J, Klatser P, Turner APF. Detection of Mycobacterium tuberculosis (TB) in vitro and in situ using an electronic nose in combination with a neural network system. Biosens Bioelectron 2004;20:538–44. [17] Dragonieri S, Schot R, Mertens BJA, Le Cessie S, Gauw SA, Spanevello A, et al. An electronic nose in the discrimination of patients with asthma and controls. J Allergy Clin Immunol 2007;120:856–62. [18] Zhu L, Seburg RA, Tsai E, Puech S, Mifsud J-C. Flavor analysis in a pharmaceutical oral solution formulation using an electronic-nose. J Pharmaceut Biomed 2004;34:453–61. [19] Stuetz RM, Fenner RA, Engin G. Characterisation of wastewater using an electronic nose. Water Res 1999;33:442–52. [20] De Cesare F, Pantalei S, Zampetti E, Macagnano A. Electronic nose and SPME techniques to monitor phenanthrene biodegradation in soil. Sensor Actuat B: Chem 2008;131:63–70. [21] Delpha C, Siadat M, Lumbreras M. An electronic nose for the discrimination of forane 134a and carbon dioxide in a humidity controlled atmosphere. Sensor Actuat B: Chem 2001;78:49–56. [22] Bhattacharyya N, Seth S, Tudu B, Tamuly P, Jana A, Ghosh D, et al. Detection of optimum fermentation time for black tea manufacturing using electronic nose. Sensor Actuat B: Chem 2007;122:627–34. [23] García-Martínez T, Bellincontro A, de Lerma MdlNL, Peinado RA, Mauricio JC, Mencarelli F, et al. Discrimination of sweet wines partially fermented by two

588

[24]

[25]

[26]

[27]

[28]

H. Jiang et al. / Process Biochemistry 49 (2014) 583–588 osmo-ethanol-tolerant yeasts by gas chromatographic analysis and electronic nose. Food Chem 2011;127:1391–6. Bhattacharyya N, Seth S, Tudu B, Tamuly P, Jana A, Ghosh D, et al. Monitoring of black tea fermentation process using electronic nose. J Food Eng 2007;80:1146–56. Pinheiro C, Rodrigues CM, Schäfer T, Crespo JG. Monitoring the aroma production during wine-must fermentation with an electronic nose. Biotechnol Bioeng 2002;77:632–40. Bachinger T, Mandenius C-F. Physiologically motivated monitoring of fermentation processes by means of an electronic nose. Eng Life Sci 2001;1: 33–42. Bhattacharya N, Tudu B, Jana A, Ghosh D, Bandhopadhyaya R, Bhuyan M. Preemptive identification of optimum fermentation time for black tea using electronic nose. Sensor Actuat B: Chem 2008;131:110–6. Jiang H, Liu G, Xiao X, Yu S, Mei C, Ding Y. Classification of Chinese soybean paste by Fourier transform near-infrared (FT-NIR) spectroscopy and different supervised pattern recognition. Food Anal Method 2012;5: 928–34.

[29] Roggo Y, Duponchel L, Huvenne J-P. Comparison of supervised pattern recognition methods with McNemar’s statistical test: application to qualitative analysis of sugar beet by near-infrared spectroscopy. Anal Chim Acta 2003;477:187–200. [30] Jalali-Heravi M, Masoum S, Shahbazikhah P. Simulation of 13 C nuclear magnetic resonance spectra of lignin compounds using principal component analysis and artificial neural networks. J Magn Reson 2004;171:176–85. [31] Sánchez MS, Sarabia LA. Efficiency of multi-layered feed-forward neural networks on classification in relation to linear discriminant analysis, quadratic discriminant analysis and regularized discriminant analysis. Chemom Intell Lab Syst 1995;28:287–303. [32] Berrueta L, Alonsosalces R, Heberger K. Supervised pattern recognition in food analysis. J Chromatogr A 2007;1158:196–214. [33] Cortes C, Vapnik V. Support-vector networks. Mach Learn 1995;20:273–97. [34] Thissen U. Comparing support vector machines to PLS for spectral regression applications. Chemom Intell Lab Syst 2004;73:169–79. [35] Sathiya Keerthi S, Lin C-J. Asymptotic behaviors of support vector machines with Gaussian kernel. Neural Comput 2003;15:1667–89.