- Email: [email protected]
Monitoring storage time and quality attribute of egg based on electronic nose
Analytica Chimica Acta 650 (2009) 183–188
Contents lists available at ScienceDirect
Analytica Chimica Acta journal homepage: www.elsevier.com/locate/aca
Monitoring storage time and quality attribute of egg based on electronic nose Wang Yongwei, Jun Wang ∗ , Bo Zhou, Qiujun Lu Department of Biosystems Engineering, Zhejiang University, 268 Kaixuan Road, Hangzhou 310029, PR China
a r t i c l e
i n f o
Article history: Received 2 May 2009 Received in revised form 19 July 2009 Accepted 20 July 2009 Available online 24 July 2009 Keywords: Egg Electronic nose Quality attribute Storage time Neural network
a b s t r a c t The objective of this study was to investigate the potential of an electronic nose (E-nose) technique for monitoring egg storage time and quality attributes. An electronic nose was used to distinguish eggs under cool and room-temperature storage by means of principal component analysis (PCA), linear discriminant analysis (LDA), BP neural network (BPNN) and the combination of a genetic algorithm and BP neural network (GANN). Results showed that the E-nose could distinguish eggs of different storage time under cool and room-temperature storage by LDA, PCA, BPNN and GANN; better prediction values were obtained by GANN than by BPNN. Relationships were established between the E-nose signal and egg quality indices (Haugh unit and yolk factor) by quadratic polynomial step regression (QPSR). The prediction models for Haugh unit and yolk factor indicated a good prediction performance. The Haugh unit model had a standard error of prediction of 3.74 and correlation coefficient 0.91; the yolk factor model had a 0.02 SEP and 0.93 correlation coefficient between predicted and measured values respectively. © 2009 Elsevier B.V. All rights reserved.
1. Introduction Eggs are generally considered as a tasty, wholesome nutritious but perishable food. The commodity value of an egg directly is determined by its quality during commercial circulation. There is a requirement therefore to determine egg quality using nondestructive methods. Many research reports have been concerned with egg quality inspection method in recent years. For example, machine image vision techniques for automatically detecting cracks have been investigated [1–4]. Image analysis by machine vision has been studied as a component of automatic egg inspection and camera inspection works excellently for detection of dirty shells (excretions and blood), broken shells and odd sharps [1,3,4]. Accuracy of vision methods depends on the resolution of the camera, the sorting algorithm, the colour of the shell, and so on; results indicate detection of freshness is more difficult and the inspection correction classification rate was too low. Some research on acoustic impulse responses of egg was reported [5,6]. The eggs were impacted, resulting sound signals were picked up by microphone some peak frequencies were identified and the correlation between egg cracks and peak frequency (or frequency domain) was found. However, the results did not indicate that the egg internal quality can be detected by acoustic impulse response.
∗ Corresponding author. Tel.: +86 571 86971881; fax: +86 571 86971139. E-mail address: [email protected] (J. Wang). 0003-2670/$ – see front matter © 2009 Elsevier B.V. All rights reserved. doi:10.1016/j.aca.2009.07.049
An electronic nose (E-nose) system is a sensor-based technology which considers total headspace volatiles and creates a unique smell print. E-nose does not resolve the sample’s volatiles into their individual components but responds to the whole set of volatiles in a unique digital pattern. The pattern is a signature of the particular set of aromatic compounds. For each process or application of interest, a database of such digitised patterns is created, called the training set. When an unknown sample is exposed to the Enose sensors, the E-nose first digitises the sample’s volatiles and then compares the resulting pattern with the existing training set. E-nose technology is being investigated for discrimination of different quality fruits and other agri-products. For example, Gomez et al. found that E-nose technology using metal oxide sensors could discriminate storage shelf-life for mandarin [7,8]. Savels et al. demonstrated that an E-nose with quartz microbalance sensors could be used to evaluate the optimal harvest date for Jonagold and Braeburn apples [9]. Brezmes et al. found that the olfactory system was able to classify fruit samples into three different states of ripeness (green, ripe and overripe) with very good accuracy for peaches and pears and showed a success rate above 92% [10]. Some researchers have attempted to identify volatile components that contribute to the egg unique flavours and aromas, working with different extraction and analytical techniques (steamdistillation, solvent extraction, purge and trap, etc.): several aldehydes, aromatic compounds and sulphur compounds were identified in greatest concentrations [11]. In particular, methylsulphide compounds are closely related to deterioration and perception of unacceptable odours in whole eggs [12]. An alternative strategy to sensing the global profile of organic volatiles emitted by eggs may potentially be achieved using arti-
184
W. Yongwei et al. / Analytica Chimica Acta 650 (2009) 183–188
Fig. 1. Diagrammatic layout of electronic nose.
ficial olfactory systems (AOS) [13]. Dutta et al. reported on the use of an E-nose, employing an array of four tin oxide sensors, in combination with a pattern recognition engine, to predict the freshness of eggs. This method produced results in good agreement with the three categories of egg freshness determined from the ‘use by date’ of egg samples and the time periods recorded during the course of the experiments [13]. AOS based on common metal oxide sensors was presented as an interesting tool to carry out, in a simple and rapid way, the freshness assessment of industrial egg products. Contemporaneously, the correlation between AOS responses and the corresponding chemical (lactic and succinic acids) and microbiological (total viable mesophilic bacteria and Enterobacteriaceae counts) parameters, which are legal references to attest the quality of the egg products, has been demonstrated [14]. However, little detailed information is available on inspecting the internal quality of egg based on E-nose technology. The aim of this research was to study the feasibility of the E-nose technique for inspecting internal quality in eggs after different storage times and methods and to develop prediction models for egg internal quality using the E-nose signal.
2. Materials and methods 2.1. Electronic nose (E-nose) The electronic nose consists of an array of gas sensors with different selectivity patterns, a signal-collecting unit and pattern recognition software. Fig. 1 shows a schematic diagram of such a system. The array is composed of 8 metal oxide semiconductor (MOS) type chemical sensors. Table 1 lists all 8 of the sensors used and their main applications. A/D data acquisition card (DAQ): transformed analogue parameters to digital parameters. DAQ type was NI USB-6009, 8 ways and 14 bits D/A conversion. The experimental setup consisted of sensor channels, sampling channels and a personal computer (PC). This E-nose had the ability for auto-regulation, auto-calibration and autoconcentration.
2.2. Experimental sample Eggs were newly laid samples (M: >53 and <63 g;) from the barn housing systems (Hangzhou Hennery, China) and were collected in 15 September 2007. Four hundred eggs with intact shell were carefully collected under the magnifying glass inspection, were divided into two groups (200 eggs each group) and stored at 22 ± 1 ◦ C and 4 ± 0.5 ◦ C (except for intraday experiment: fresh egg), respectively. Each group was divided into 5 sets (and put in 5 boxes, 40 eggs each box). One set was removed at 0, 1, 2, 3 and 4 weeks and evaluated using E-nose and other measurements, respectively. Eggs were newly-laid samples collected in Hangzhou Hennery, China, on 15 September 2007. Four hundred eggs with intact shells were carefully collected following magnifying glass inspection, divided into two groups (200 eggs each group) and stored at 22 ± 1 ◦ C and 4 ± 0.5 ◦ C (except for intraday experiment: fresh egg), respectively. Each group was divided into 5 sets and put into 5 boxes (40 eggs per box). One set was removed at 0, 1, 2, 3 and 4 weeks and evaluated using E-nose and other measurements respectively. After removal, eggs stored at 4 ◦ C were placed in the laboratory for 3–4 h until the egg’s temperature reached room temperature. Then, each egg was placed in an airtight glass jar with a volume of 250 mL (concentration chamber). The glass jar was closed for 30 min and headspace reached static equilibrium (Fig. 1). During the experiments, 24 eggs were removed from each box at random for examination by the E-nose and obtained correlation eigenvalue for the training, the other 16 eggs were used for testing for Artificial Neural Network and Quadratic Polynomial Step Regression. So, the 200 samples (40 duplicates for each of the 5 sets) were divided into two parts: 120 samples (24 samples of each part) for the training set and the rest 80 samples (16 samples of each part) for the training set during 4-week experiments. 2.3. Experiment procedure The E-nose experiment was started when the resistance of the gas sensors remained stable at a high value. A fan mixed the air in the test chamber. A little amount of the headspace gas to be tested was introduced into the test chamber. Then, the data acquisition
Table 1 Sensors used and their main applications in electronic nose. Number in array
Sensor-name
General description
Reference
S1 S2 S3 S4 S5 S6 S7 S8
TGS822 MQ-3 TGS825 TGS800 TGS824 TGS813 TGS880 MQ-7
Mainly sensitive to hydrogen, carbon monoxide, hydrocarbon Detects alcohols, broad range Sensitive to sulphur organic compounds Reacts on low concentrations gasoline of ventilating arrangement Very sensitive to ammonia Detects combustible gas Detects aromatics compounds, moisture Detects carbon monoxide
H2 C2H5OH H2 S Gasoline NH3 CH4 Benzene CO
W. Yongwei et al. / Analytica Chimica Acta 650 (2009) 183–188
started and lasted 90 s, time enough for sensors to reach a stable value. When a measurement was completed, a cleaning phase was activated with clean air being introduced into the test chamber. The main purpose of this cleaning was to clean the test chamber and return sensor values to their baseline. The cleaning phase lasted 90 s. E-nose was used at a temperature of 22 ± 1 ◦ C and 50–60% RH during all experiments. The experiments involved three replicates for each egg. After evaluation by E-nose, the egg was removed from the jar and the internal quality was measured. 2.4. Internal quality measurement 2.4.1. Haugh unit The most widely used measurement of albumen quality is the Haugh unit. Haugh unit is related to the albumen height (H) and egg weight (W) (Eq. (1)): HU = 100 log(H + 7.57 − 1.7W 0.37 )
(1)
2.4.2. Yolk coefficient The Yolk coefficient is related to the yolk height (H) and width (D) (Eq. (2)): Y (%) =
H × 100. D
(2)
2.5. Pattern recognition 2.5.1. PCA and LDA analysis Pattern-recognition techniques used were PCA and LDA. Principal component analysis is a chemometric linear, unsupervised and pattern recognition technique used for analyzing, classifying and reducing the dimensionality of numerical datasets in a multivariate problem. This method permits extraction of useful information from the data, and exploration of the data structure, the relationship between objects, the relationship between objects and variables, and the global correlation of the variables. Linear discriminant analysis (LDA) is one of the most used classification procedure. The method maximizes the variance between categories and minimizes the variance within categories. It merely looks for a sensible rule to discriminate between them by forming linear functions of the data maximizing the ratio of the between-set sum of squares to the within-set sum of squares. 2.5.2. Typical back-propagation neural network Back-propagation neural network (BPNN) is a type of artificial neural networks that most widely used to solve problems in modeling and classification. The typical back-propagation network consists of an input layer, an output layer and at least one hidden layer. Each layer contains neurons and each neuron is a simple micro-processing unit which receives and combines signals from other neurons. Each neuron has weighted inputs, summation function, transfer function and output. The behavior of a back-propagation network is mainly determined by the transfer functions of its neurons. 2.5.3. Genetic algorithm and BP neural network (GANN) Genetic algorithms are interesting because they are inspired by biological evolution and they seem applicable to a wide range of optimization problems. The data processed by the algorithm consists of a set (population) of strings which represent multiple points in a search space. A string with a finite length in which each bit is called an allele, is defined as a solution (individual) having the objective function value of a point in the search space. The function to be minimized by the algorithm is converted to a fitness value that determines the probability of the individual undergoing transitional operators. The operators are analogous to the biological
185
terms of crossover, mutation and selection. The GA search technique adopted here is advantageous in that it does not require any analytical properties of a function to be optimized, and in that it searches a solution more globally in a large, irregularly shaped space. The GA is used to determine one optimized set of all BPNN training factors. 2.5.4. Quadratic polynomial step regression (QPSR) The QPST algorithm establishes a model that describes the relationship between sensor signals and the egg quality indices. All variables left in the models are significant at the 0.01 level. The regression equation should be checked: (1) if P ≤ 0.05, in the analysis of variance or the regression equation cannot be used; (2) whether the significant level of the partial correlation coefficient for all regression coefficients is less than 0.05; (3) whether Durbin–Watson (DW) does approach to 2 [15]. DW is defined as follows: DW =
n
(ei − ei−1 )2
2
where ei is the predictive error of i sample. The quality of the model is defined by the accuracy of F (variance analysis valve), P (significant level), SEC (standard error of calibration), SEP (standard error prediction) and R (correlation coefficient) between the predicted and measured parameters.
Ic 1 SEC = (ˆyi − yi )2 Ic − 1
i=1
Ip 1 2 SEP = (ˆyi − yi ) IP − 1
i=1
yˆ i is the predicted value of the ith observation, yi is the measured value of the ith observation, Ic is the number of observations in the testing set and Ip is the number of observations in the training set. 3. Results and discussion 3.1. Electronic nose response to egg volatile Fig. 2 shows a typical signal of eight sensors for two roomtemperature storage times in which each curve represents a different sensor response with time. The ordinate represents sensor response signal; this is the gas response R/R0 , in which R and R0 express the resistance of a sensor in a detected gas and in clean air respectively. It was apparent that, for most sensors, after an initial period of low conductivity, the conductivity increased sharply and then stabilised. In this research, the signal of each sensor after 60 s was used in analysis. The value of sensor response signals differed with different storage times (Fig. 2). Fig. 2a shows that the value R/R0 (1.0–2.0) of egg stored for 1 week was lower than R/R0 (1.5–3.0) of the egg stored for 2 weeks, indicating that the sensor response of the E-nose varied with storage time and quality attributes of the egg. This also implied that the E-nose technique might distinguish eggs of different storage times and quality attributes. 3.2. Classification of storage time using PCA and LDA 3.2.1. Discrimination of cool storage time Fig. 3 shows the results of PCA and LDA applied to the egg dataset. The first principal component (PC1, LD1) and the second principal component (PC2, LD2) account together for more than 90% of the
186
W. Yongwei et al. / Analytica Chimica Acta 650 (2009) 183–188
Fig. 2. Sensor response curves for egg volatiles (a: egg after 1 week storage; b: egg after 2 weeks storage).
Fig. 3. Discrimination of eggs after different storage times at 4 ◦ C (a: PCA, b: LDA).
variance in the dataset. On the basis of sample scores on PC 1 & 2 (Fig. 3a), some clustering of egg samples on the basis of storage time may be detected. This clustering is incomplete, however, in the case of fresh eggs and eggs stored for 1 week set and between 1-week old eggs and their 2-week counterparts. Using LDA, Fig. 3b shows that each set can be clearly distinguished from the other sets except for fresh and 1-week old eggs which are partially overlapping.
Fig. 3a and b show that LDA produces better results than PCA as might be expected. LDA is a well established statistical technique for classification and discrimination purposes that works well under normality assumption provided that all types are strictly homogeneous. Whereas, the PCA is a projection method of the original variables onto new ones, orthogonal and arranged according to their eigenvalue.
Fig. 4. Discrimination of eggs after different storage times at room-temperature (a: PCA, b: LDA).
W. Yongwei et al. / Analytica Chimica Acta 650 (2009) 183–188
187
Table 2 Discrimination of eggs with different storage times (4 ◦ C) by ANN. Network style
Correct rate of training set
BPNN GANN
95% 100%
Correct rate of testing set Fresh
One-week
Two-week
Three-week
Four-week
64(80%) 71(88.75%)
65(81.25%) 70(87.5%)
64(80%) 71(88.75%)
66(82.5%) 72(90%)
66(82.5%) 73(91.25%)
These results can be related to the volatiles emitted from egg samples. Cool storage under 4 ◦ C is the better storage method. Under these storage conditions, the variation of egg internal quality and the variation of volatiles are limited after 1 week. Thus the difference in volatiles between fresh egg and 1-week old egg are not obvious and cause some overlapped or slightly overlapped clusters by both PCA and LDA. Egg quality changed slightly after 2 weeks, egg internal quality changed faster and the volatiles emitted from egg also increased after 3 weeks, with the result that nearly all sets could be distinguished by PCA or LDA. 3.2.2. Discrimination room-temperature storage time Analytical results are shown in Fig. 4. The first principal component (PC1, LD1) and second principal component (PC2, LD2) accounted together for more than 95% of the variance in the dataset, implying that these methods could be used for differentiation. Each sample set was clearly distinguishable from the other sets, except for the 1-week set and 2-week sets after PCA (Fig. 4a). Five sets were clearly distinguishable from each other by LDA (Fig. 4b). This is because PCA estimated the correlation structure of the variables and investigating how many components (a linear combination of original features), to explain the greater part of variance with a minimum loss of information, whereas, LDA supplies a number of orthogonal linear discriminant functions, equal to the number of categories minus one, that permit the samples to be classified in one or another category. These results are similar to those obtained by Liu et al. [16] who reported that egg quality was AA grade (the best grade) after 1 week storage, A grade at 2 weeks storage, B grade at 4 weeks storage and C grade after 4 weeks storage at room-temperature. These results indicated that egg internal quality changed faster under room-temperature storage than under cool storage and that volatiles emitted from egg also changed more quickly under roomtemperature storage than under cool storage. Different storage times cause the volatile changes which affect the E-nose sensor response signal and facilitate storage time differentiation by PCA and LDA. Comparing Figs. 3 and 4, the differentiation was better following room-temperature storage than cool storage because egg internal quality and volatiles emitted from egg change more rapidly under conditions of room-temperature storage.
It can be seen that the correct classification rate of BPNN for the learning samples does not reach 100% and this rate for the test samples is 80–82.5%. In contrast, the correct classification rate of the GANN for the learning samples was 100% and 88.75–91.25% for the test samples. GANN therefore performed better than BPNN. 3.3.2. Discrimination of the room-temperature storage time It can be seen in Table 3 that the correct classification rates of BPNN and GANN for learning samples were both 100%. Recognition accuracy of BPNN for test samples was 90–93.8%, while for GANN, a corresponding value >93.75% was achieved. Comparing Tables 2 and 3, the models performed better with eggs stored at room temperature than in cool storage. This may be because there are greater changes in egg internal quality and volatiles emitted from egg under the room-temperature storage for any given storage time. Results also show higher classification accuracy by GANN than BPNN. 3.4. Egg internal quality prediction Egg internal quality is directly related to emission of volatiles which increase in quantity with declining egg quality. When eggs deteriorate, they develop off-odours and lot of sulphides are produced. Therefore, the sensor response signals of the E-nose may potentially be used for the prediction of egg internal quality. Eight sensor response signals of the E-nose after 60 s measurement time were treated as independent variables while egg internal quality parameters (Haugh unit, HU; Yolk coefficient, Y) were dependent variables. Two hundred and forty (240) egg samples chosen from the two storage trials (total 400 eggs) were selected as the training set to establish a regression model; the remaining 160 egg samples were used as a test set to evaluate models developed. The quadratic polynomial step regression algorithm develops a model that describes the relationship between sensor signals and the egg quality indices. In this analysis, backward stepwise regression, with p-values of 0.10, was used to exclude variables that had little or no influence on the trait under analysis. Thus all variables left in the models are significant at the 0.01 level. The predictive models for HU and Y thus developed are given below: HU = −18.24 − 143.05S1 + 250.31S2 + 96.57S4 − 18.44S42 + 62.16S1 × S5 − 136.40S2 × S5
3.3. Discrimination storage time by ANN 3.3.1. Discrimination of the cool storage time In this study, BPNN and GANN were employed. The response value of the E-nose at 60 s was used as the input vector of ANN. The network topology was designed as 8-12-5 and results are shown in Table 2.
(3)
Eq. (3): F = 169.43, P < 10−4 , R2 = 0.91, SEC = 1.15, Durbin–Watson (DW) = 2.09. Y = −10.80 + 2.36S1 + 2.73S2 + 5.01S3 + 20.30S6 − 16.98S8 − 0.98S12 + 2.93S42 − 1.54S52 + 13.88S72 − 9.14S82 − 7.68S1
Table 3 Discrimination of egg with different storage time under room-temperature by ANN. Network style
BPNN GANN
Correct rate of training set
100% 100%
Correct rate of testing set Fresh
One-week
Two-week
Three-week
Four-week
72(90%) 76(95%)
73(91.25%) 76(95%)
73(91.25%) 75(93.75%)
75(93.75%) 77(96.25%)
75(93.75%) 76(95%)
188
W. Yongwei et al. / Analytica Chimica Acta 650 (2009) 183–188
Fig. 5. Predicted versus measured values of egg quality indices from QPSR models (a: Haugh unit, b: Yolk factor).
Table 4 Prediction results for egg quality on the base of E-nose signal. Quality indices
QPST Training set
HU Y
Testing set
R2
SEC
ERR%
R2
SEP
ERR%
0.93 0.94
2.15 0.01
2.55 1.5
0.91 0.93
3.74 0.02
5.37 2.89
× S3 + 4.69S1 × S4 + 0.88S1 × S5 − 4.0S1 × S7 + 6.8S1 × S8
Acknowledgements
− 0.2S2 × S5 − 2.25S2 × S6 − 2.74S3 × S4 − 6.26S3 × S6 + 11.78S3 × S8 − 1.89S4 × S5 − 4.57S4 × S8 + 3.21S5 × S8 − 25.45S6 × S7 + 14.32S6 × S8
(3) BPNN and GANN produced good predictions for egg storage time. GANN demonstrated better correction classification rates than BPNN. (4) The quadratic polynomial step regression (QPSR) algorithm established models that described the relationship between sensor signals and egg quality indices. The QPST models appeared to be of high predictive ability, with high correlation coefficients (R2 = 0.91, 0.93) between predicted and measured values.
(4)
Eq. (4): F = 8.69, P < 10−4 , R2 = 0.93, SEC = 0.08, Durbin–Watson (DW) = 2.18. The following three aspects of test to the regression equation should be checked to evaluate a model’s quality: (1) if P = 0.05 in the analysis of variance or the regression equation cannot be used; (2) whether the significant level of the partial correlation coefficient for all regression coefficients is less than 0.05; (3) whether Durbin–Watson (DW) approaches a value of 2. The inspection found that the significance level of each regression equation was less than 0.05 and DW was also close to 2; therefore, all regression equations were considered statistically valid. The test set was used to validate the prediction capabilities of the models developed (Fig. 5). The correlation coefficient between measurement and predictive values of HU and Y was 0.91 and 0.93 respectively while the standard error of prediction was 3.74 and 0.02 respectively. Results for two models developed are displayed in Table 4. The QPST models appeared to be of high predictive ability. 4. Conclusions (1) Sensor response signals in an electronic nose differed with different storage times in the case of eggs. (2) Good distinction between eggs stored for different times were obtained by PCA and LDA. Results by LDA were better than those obtained by PCA.
The authors acknowledge the financial support of the Chinese National Foundation of Nature and Science through Project 30771246, the National High Technology Research and Development Program of China through Project 2006AA10Z212, the Research Fund for the Doctoral Program of Chinese National Higher Education through Project 20060335060. References [1] V.C. Patel, R.W. McClendon, J.W. Goodrum, Artif. Intell. Rev. 12 (1998) 163–176. [2] N.D. Overfield, Egg grading as a form of quality control, Poultry-Misset 23 (1987) 10–13. [3] B. De Ketelaere, F. Bamelis, B. Kemps, E. Decuypere, J. De Baerdemaeker, World’s Poultry Sci. J. 60 (2004) 289–302. [4] V.C. Patel, R.W. McClendon, J.W. Goodrum, Appl. Eng. Agric. 12 (1996) 253–258. [5] B. De Ketelaere, P. Coucke, J. De Baerdemaeker, J. Agric. Eng. Res. 76 (2000) 157–163. [6] H.K. Cho, W.K. Choi, J.H. Paek, Trans. ASAE 43 (2000) 1921–1926. [7] A.H. Gomez, J. Wang, G.X. Hu, A.G. Pereira, Sens. Actuators B 13 (2006) 347–353. [8] A.H. Gomez, J. Wang, G.X. Hu, A.G. Pereira, LWT-Food Sci. Technol. 40 (2007) 681–689. [9] S. Saevels, J. Lammertyn, A.Z. Berna, C.D. Natale, B.M. Nicolaı, Postharvest Biol. Technol. 30 (2003) 3–4. [10] J. Brezmes, E. Llobet, X. Vilanova, G. Saiz, X. Correig, Sens. Actuators B 69 (2000) 223–229. [11] M.W. Warren, D.K. Larick, H.R. Ball, J. Food Sci. 60 (1995) 79–84. [12] M.L. Brown, D.M. Holbrook, E.F. Hoerning, M.G. Legendre, A.J. St. Angelo, Poultry Sci. 65 (1986) 1925–1933. [13] R. Dutta, E.L. Hines, J.W. Gardner, D.D. Udrea, P. Boilot, Meas. Sci. Technol. 14 (2003) 190–198. [14] M.A. Suman, G.B. Riani, E.B. Dalcanale, Sens. Actuators B 125 (2007) 40–47. [15] H. Zhang, J. Wang, S. Ye, Anal. Chim. Acta 606 (2008) 112–118. [16] Y. Liu, Z. Qiao, Acta Agric. Univ. Jiangxiensis 24 (2002) 12–16.
Contents lists available at ScienceDirect
Analytica Chimica Acta journal homepage: www.elsevier.com/locate/aca
Monitoring storage time and quality attribute of egg based on electronic nose Wang Yongwei, Jun Wang ∗ , Bo Zhou, Qiujun Lu Department of Biosystems Engineering, Zhejiang University, 268 Kaixuan Road, Hangzhou 310029, PR China
a r t i c l e
i n f o
Article history: Received 2 May 2009 Received in revised form 19 July 2009 Accepted 20 July 2009 Available online 24 July 2009 Keywords: Egg Electronic nose Quality attribute Storage time Neural network
a b s t r a c t The objective of this study was to investigate the potential of an electronic nose (E-nose) technique for monitoring egg storage time and quality attributes. An electronic nose was used to distinguish eggs under cool and room-temperature storage by means of principal component analysis (PCA), linear discriminant analysis (LDA), BP neural network (BPNN) and the combination of a genetic algorithm and BP neural network (GANN). Results showed that the E-nose could distinguish eggs of different storage time under cool and room-temperature storage by LDA, PCA, BPNN and GANN; better prediction values were obtained by GANN than by BPNN. Relationships were established between the E-nose signal and egg quality indices (Haugh unit and yolk factor) by quadratic polynomial step regression (QPSR). The prediction models for Haugh unit and yolk factor indicated a good prediction performance. The Haugh unit model had a standard error of prediction of 3.74 and correlation coefficient 0.91; the yolk factor model had a 0.02 SEP and 0.93 correlation coefficient between predicted and measured values respectively. © 2009 Elsevier B.V. All rights reserved.
1. Introduction Eggs are generally considered as a tasty, wholesome nutritious but perishable food. The commodity value of an egg directly is determined by its quality during commercial circulation. There is a requirement therefore to determine egg quality using nondestructive methods. Many research reports have been concerned with egg quality inspection method in recent years. For example, machine image vision techniques for automatically detecting cracks have been investigated [1–4]. Image analysis by machine vision has been studied as a component of automatic egg inspection and camera inspection works excellently for detection of dirty shells (excretions and blood), broken shells and odd sharps [1,3,4]. Accuracy of vision methods depends on the resolution of the camera, the sorting algorithm, the colour of the shell, and so on; results indicate detection of freshness is more difficult and the inspection correction classification rate was too low. Some research on acoustic impulse responses of egg was reported [5,6]. The eggs were impacted, resulting sound signals were picked up by microphone some peak frequencies were identified and the correlation between egg cracks and peak frequency (or frequency domain) was found. However, the results did not indicate that the egg internal quality can be detected by acoustic impulse response.
∗ Corresponding author. Tel.: +86 571 86971881; fax: +86 571 86971139. E-mail address: [email protected] (J. Wang). 0003-2670/$ – see front matter © 2009 Elsevier B.V. All rights reserved. doi:10.1016/j.aca.2009.07.049
An electronic nose (E-nose) system is a sensor-based technology which considers total headspace volatiles and creates a unique smell print. E-nose does not resolve the sample’s volatiles into their individual components but responds to the whole set of volatiles in a unique digital pattern. The pattern is a signature of the particular set of aromatic compounds. For each process or application of interest, a database of such digitised patterns is created, called the training set. When an unknown sample is exposed to the Enose sensors, the E-nose first digitises the sample’s volatiles and then compares the resulting pattern with the existing training set. E-nose technology is being investigated for discrimination of different quality fruits and other agri-products. For example, Gomez et al. found that E-nose technology using metal oxide sensors could discriminate storage shelf-life for mandarin [7,8]. Savels et al. demonstrated that an E-nose with quartz microbalance sensors could be used to evaluate the optimal harvest date for Jonagold and Braeburn apples [9]. Brezmes et al. found that the olfactory system was able to classify fruit samples into three different states of ripeness (green, ripe and overripe) with very good accuracy for peaches and pears and showed a success rate above 92% [10]. Some researchers have attempted to identify volatile components that contribute to the egg unique flavours and aromas, working with different extraction and analytical techniques (steamdistillation, solvent extraction, purge and trap, etc.): several aldehydes, aromatic compounds and sulphur compounds were identified in greatest concentrations [11]. In particular, methylsulphide compounds are closely related to deterioration and perception of unacceptable odours in whole eggs [12]. An alternative strategy to sensing the global profile of organic volatiles emitted by eggs may potentially be achieved using arti-
184
W. Yongwei et al. / Analytica Chimica Acta 650 (2009) 183–188
Fig. 1. Diagrammatic layout of electronic nose.
ficial olfactory systems (AOS) [13]. Dutta et al. reported on the use of an E-nose, employing an array of four tin oxide sensors, in combination with a pattern recognition engine, to predict the freshness of eggs. This method produced results in good agreement with the three categories of egg freshness determined from the ‘use by date’ of egg samples and the time periods recorded during the course of the experiments [13]. AOS based on common metal oxide sensors was presented as an interesting tool to carry out, in a simple and rapid way, the freshness assessment of industrial egg products. Contemporaneously, the correlation between AOS responses and the corresponding chemical (lactic and succinic acids) and microbiological (total viable mesophilic bacteria and Enterobacteriaceae counts) parameters, which are legal references to attest the quality of the egg products, has been demonstrated [14]. However, little detailed information is available on inspecting the internal quality of egg based on E-nose technology. The aim of this research was to study the feasibility of the E-nose technique for inspecting internal quality in eggs after different storage times and methods and to develop prediction models for egg internal quality using the E-nose signal.
2. Materials and methods 2.1. Electronic nose (E-nose) The electronic nose consists of an array of gas sensors with different selectivity patterns, a signal-collecting unit and pattern recognition software. Fig. 1 shows a schematic diagram of such a system. The array is composed of 8 metal oxide semiconductor (MOS) type chemical sensors. Table 1 lists all 8 of the sensors used and their main applications. A/D data acquisition card (DAQ): transformed analogue parameters to digital parameters. DAQ type was NI USB-6009, 8 ways and 14 bits D/A conversion. The experimental setup consisted of sensor channels, sampling channels and a personal computer (PC). This E-nose had the ability for auto-regulation, auto-calibration and autoconcentration.
2.2. Experimental sample Eggs were newly laid samples (M: >53 and <63 g;) from the barn housing systems (Hangzhou Hennery, China) and were collected in 15 September 2007. Four hundred eggs with intact shell were carefully collected under the magnifying glass inspection, were divided into two groups (200 eggs each group) and stored at 22 ± 1 ◦ C and 4 ± 0.5 ◦ C (except for intraday experiment: fresh egg), respectively. Each group was divided into 5 sets (and put in 5 boxes, 40 eggs each box). One set was removed at 0, 1, 2, 3 and 4 weeks and evaluated using E-nose and other measurements, respectively. Eggs were newly-laid samples collected in Hangzhou Hennery, China, on 15 September 2007. Four hundred eggs with intact shells were carefully collected following magnifying glass inspection, divided into two groups (200 eggs each group) and stored at 22 ± 1 ◦ C and 4 ± 0.5 ◦ C (except for intraday experiment: fresh egg), respectively. Each group was divided into 5 sets and put into 5 boxes (40 eggs per box). One set was removed at 0, 1, 2, 3 and 4 weeks and evaluated using E-nose and other measurements respectively. After removal, eggs stored at 4 ◦ C were placed in the laboratory for 3–4 h until the egg’s temperature reached room temperature. Then, each egg was placed in an airtight glass jar with a volume of 250 mL (concentration chamber). The glass jar was closed for 30 min and headspace reached static equilibrium (Fig. 1). During the experiments, 24 eggs were removed from each box at random for examination by the E-nose and obtained correlation eigenvalue for the training, the other 16 eggs were used for testing for Artificial Neural Network and Quadratic Polynomial Step Regression. So, the 200 samples (40 duplicates for each of the 5 sets) were divided into two parts: 120 samples (24 samples of each part) for the training set and the rest 80 samples (16 samples of each part) for the training set during 4-week experiments. 2.3. Experiment procedure The E-nose experiment was started when the resistance of the gas sensors remained stable at a high value. A fan mixed the air in the test chamber. A little amount of the headspace gas to be tested was introduced into the test chamber. Then, the data acquisition
Table 1 Sensors used and their main applications in electronic nose. Number in array
Sensor-name
General description
Reference
S1 S2 S3 S4 S5 S6 S7 S8
TGS822 MQ-3 TGS825 TGS800 TGS824 TGS813 TGS880 MQ-7
Mainly sensitive to hydrogen, carbon monoxide, hydrocarbon Detects alcohols, broad range Sensitive to sulphur organic compounds Reacts on low concentrations gasoline of ventilating arrangement Very sensitive to ammonia Detects combustible gas Detects aromatics compounds, moisture Detects carbon monoxide
H2 C2H5OH H2 S Gasoline NH3 CH4 Benzene CO
W. Yongwei et al. / Analytica Chimica Acta 650 (2009) 183–188
started and lasted 90 s, time enough for sensors to reach a stable value. When a measurement was completed, a cleaning phase was activated with clean air being introduced into the test chamber. The main purpose of this cleaning was to clean the test chamber and return sensor values to their baseline. The cleaning phase lasted 90 s. E-nose was used at a temperature of 22 ± 1 ◦ C and 50–60% RH during all experiments. The experiments involved three replicates for each egg. After evaluation by E-nose, the egg was removed from the jar and the internal quality was measured. 2.4. Internal quality measurement 2.4.1. Haugh unit The most widely used measurement of albumen quality is the Haugh unit. Haugh unit is related to the albumen height (H) and egg weight (W) (Eq. (1)): HU = 100 log(H + 7.57 − 1.7W 0.37 )
(1)
2.4.2. Yolk coefficient The Yolk coefficient is related to the yolk height (H) and width (D) (Eq. (2)): Y (%) =
H × 100. D
(2)
2.5. Pattern recognition 2.5.1. PCA and LDA analysis Pattern-recognition techniques used were PCA and LDA. Principal component analysis is a chemometric linear, unsupervised and pattern recognition technique used for analyzing, classifying and reducing the dimensionality of numerical datasets in a multivariate problem. This method permits extraction of useful information from the data, and exploration of the data structure, the relationship between objects, the relationship between objects and variables, and the global correlation of the variables. Linear discriminant analysis (LDA) is one of the most used classification procedure. The method maximizes the variance between categories and minimizes the variance within categories. It merely looks for a sensible rule to discriminate between them by forming linear functions of the data maximizing the ratio of the between-set sum of squares to the within-set sum of squares. 2.5.2. Typical back-propagation neural network Back-propagation neural network (BPNN) is a type of artificial neural networks that most widely used to solve problems in modeling and classification. The typical back-propagation network consists of an input layer, an output layer and at least one hidden layer. Each layer contains neurons and each neuron is a simple micro-processing unit which receives and combines signals from other neurons. Each neuron has weighted inputs, summation function, transfer function and output. The behavior of a back-propagation network is mainly determined by the transfer functions of its neurons. 2.5.3. Genetic algorithm and BP neural network (GANN) Genetic algorithms are interesting because they are inspired by biological evolution and they seem applicable to a wide range of optimization problems. The data processed by the algorithm consists of a set (population) of strings which represent multiple points in a search space. A string with a finite length in which each bit is called an allele, is defined as a solution (individual) having the objective function value of a point in the search space. The function to be minimized by the algorithm is converted to a fitness value that determines the probability of the individual undergoing transitional operators. The operators are analogous to the biological
185
terms of crossover, mutation and selection. The GA search technique adopted here is advantageous in that it does not require any analytical properties of a function to be optimized, and in that it searches a solution more globally in a large, irregularly shaped space. The GA is used to determine one optimized set of all BPNN training factors. 2.5.4. Quadratic polynomial step regression (QPSR) The QPST algorithm establishes a model that describes the relationship between sensor signals and the egg quality indices. All variables left in the models are significant at the 0.01 level. The regression equation should be checked: (1) if P ≤ 0.05, in the analysis of variance or the regression equation cannot be used; (2) whether the significant level of the partial correlation coefficient for all regression coefficients is less than 0.05; (3) whether Durbin–Watson (DW) does approach to 2 [15]. DW is defined as follows: DW =
n
(ei − ei−1 )2
2
where ei is the predictive error of i sample. The quality of the model is defined by the accuracy of F (variance analysis valve), P (significant level), SEC (standard error of calibration), SEP (standard error prediction) and R (correlation coefficient) between the predicted and measured parameters.
Ic 1 SEC = (ˆyi − yi )2 Ic − 1
i=1
Ip 1 2 SEP = (ˆyi − yi ) IP − 1
i=1
yˆ i is the predicted value of the ith observation, yi is the measured value of the ith observation, Ic is the number of observations in the testing set and Ip is the number of observations in the training set. 3. Results and discussion 3.1. Electronic nose response to egg volatile Fig. 2 shows a typical signal of eight sensors for two roomtemperature storage times in which each curve represents a different sensor response with time. The ordinate represents sensor response signal; this is the gas response R/R0 , in which R and R0 express the resistance of a sensor in a detected gas and in clean air respectively. It was apparent that, for most sensors, after an initial period of low conductivity, the conductivity increased sharply and then stabilised. In this research, the signal of each sensor after 60 s was used in analysis. The value of sensor response signals differed with different storage times (Fig. 2). Fig. 2a shows that the value R/R0 (1.0–2.0) of egg stored for 1 week was lower than R/R0 (1.5–3.0) of the egg stored for 2 weeks, indicating that the sensor response of the E-nose varied with storage time and quality attributes of the egg. This also implied that the E-nose technique might distinguish eggs of different storage times and quality attributes. 3.2. Classification of storage time using PCA and LDA 3.2.1. Discrimination of cool storage time Fig. 3 shows the results of PCA and LDA applied to the egg dataset. The first principal component (PC1, LD1) and the second principal component (PC2, LD2) account together for more than 90% of the
186
W. Yongwei et al. / Analytica Chimica Acta 650 (2009) 183–188
Fig. 2. Sensor response curves for egg volatiles (a: egg after 1 week storage; b: egg after 2 weeks storage).
Fig. 3. Discrimination of eggs after different storage times at 4 ◦ C (a: PCA, b: LDA).
variance in the dataset. On the basis of sample scores on PC 1 & 2 (Fig. 3a), some clustering of egg samples on the basis of storage time may be detected. This clustering is incomplete, however, in the case of fresh eggs and eggs stored for 1 week set and between 1-week old eggs and their 2-week counterparts. Using LDA, Fig. 3b shows that each set can be clearly distinguished from the other sets except for fresh and 1-week old eggs which are partially overlapping.
Fig. 3a and b show that LDA produces better results than PCA as might be expected. LDA is a well established statistical technique for classification and discrimination purposes that works well under normality assumption provided that all types are strictly homogeneous. Whereas, the PCA is a projection method of the original variables onto new ones, orthogonal and arranged according to their eigenvalue.
Fig. 4. Discrimination of eggs after different storage times at room-temperature (a: PCA, b: LDA).
W. Yongwei et al. / Analytica Chimica Acta 650 (2009) 183–188
187
Table 2 Discrimination of eggs with different storage times (4 ◦ C) by ANN. Network style
Correct rate of training set
BPNN GANN
95% 100%
Correct rate of testing set Fresh
One-week
Two-week
Three-week
Four-week
64(80%) 71(88.75%)
65(81.25%) 70(87.5%)
64(80%) 71(88.75%)
66(82.5%) 72(90%)
66(82.5%) 73(91.25%)
These results can be related to the volatiles emitted from egg samples. Cool storage under 4 ◦ C is the better storage method. Under these storage conditions, the variation of egg internal quality and the variation of volatiles are limited after 1 week. Thus the difference in volatiles between fresh egg and 1-week old egg are not obvious and cause some overlapped or slightly overlapped clusters by both PCA and LDA. Egg quality changed slightly after 2 weeks, egg internal quality changed faster and the volatiles emitted from egg also increased after 3 weeks, with the result that nearly all sets could be distinguished by PCA or LDA. 3.2.2. Discrimination room-temperature storage time Analytical results are shown in Fig. 4. The first principal component (PC1, LD1) and second principal component (PC2, LD2) accounted together for more than 95% of the variance in the dataset, implying that these methods could be used for differentiation. Each sample set was clearly distinguishable from the other sets, except for the 1-week set and 2-week sets after PCA (Fig. 4a). Five sets were clearly distinguishable from each other by LDA (Fig. 4b). This is because PCA estimated the correlation structure of the variables and investigating how many components (a linear combination of original features), to explain the greater part of variance with a minimum loss of information, whereas, LDA supplies a number of orthogonal linear discriminant functions, equal to the number of categories minus one, that permit the samples to be classified in one or another category. These results are similar to those obtained by Liu et al. [16] who reported that egg quality was AA grade (the best grade) after 1 week storage, A grade at 2 weeks storage, B grade at 4 weeks storage and C grade after 4 weeks storage at room-temperature. These results indicated that egg internal quality changed faster under room-temperature storage than under cool storage and that volatiles emitted from egg also changed more quickly under roomtemperature storage than under cool storage. Different storage times cause the volatile changes which affect the E-nose sensor response signal and facilitate storage time differentiation by PCA and LDA. Comparing Figs. 3 and 4, the differentiation was better following room-temperature storage than cool storage because egg internal quality and volatiles emitted from egg change more rapidly under conditions of room-temperature storage.
It can be seen that the correct classification rate of BPNN for the learning samples does not reach 100% and this rate for the test samples is 80–82.5%. In contrast, the correct classification rate of the GANN for the learning samples was 100% and 88.75–91.25% for the test samples. GANN therefore performed better than BPNN. 3.3.2. Discrimination of the room-temperature storage time It can be seen in Table 3 that the correct classification rates of BPNN and GANN for learning samples were both 100%. Recognition accuracy of BPNN for test samples was 90–93.8%, while for GANN, a corresponding value >93.75% was achieved. Comparing Tables 2 and 3, the models performed better with eggs stored at room temperature than in cool storage. This may be because there are greater changes in egg internal quality and volatiles emitted from egg under the room-temperature storage for any given storage time. Results also show higher classification accuracy by GANN than BPNN. 3.4. Egg internal quality prediction Egg internal quality is directly related to emission of volatiles which increase in quantity with declining egg quality. When eggs deteriorate, they develop off-odours and lot of sulphides are produced. Therefore, the sensor response signals of the E-nose may potentially be used for the prediction of egg internal quality. Eight sensor response signals of the E-nose after 60 s measurement time were treated as independent variables while egg internal quality parameters (Haugh unit, HU; Yolk coefficient, Y) were dependent variables. Two hundred and forty (240) egg samples chosen from the two storage trials (total 400 eggs) were selected as the training set to establish a regression model; the remaining 160 egg samples were used as a test set to evaluate models developed. The quadratic polynomial step regression algorithm develops a model that describes the relationship between sensor signals and the egg quality indices. In this analysis, backward stepwise regression, with p-values of 0.10, was used to exclude variables that had little or no influence on the trait under analysis. Thus all variables left in the models are significant at the 0.01 level. The predictive models for HU and Y thus developed are given below: HU = −18.24 − 143.05S1 + 250.31S2 + 96.57S4 − 18.44S42 + 62.16S1 × S5 − 136.40S2 × S5
3.3. Discrimination storage time by ANN 3.3.1. Discrimination of the cool storage time In this study, BPNN and GANN were employed. The response value of the E-nose at 60 s was used as the input vector of ANN. The network topology was designed as 8-12-5 and results are shown in Table 2.
(3)
Eq. (3): F = 169.43, P < 10−4 , R2 = 0.91, SEC = 1.15, Durbin–Watson (DW) = 2.09. Y = −10.80 + 2.36S1 + 2.73S2 + 5.01S3 + 20.30S6 − 16.98S8 − 0.98S12 + 2.93S42 − 1.54S52 + 13.88S72 − 9.14S82 − 7.68S1
Table 3 Discrimination of egg with different storage time under room-temperature by ANN. Network style
BPNN GANN
Correct rate of training set
100% 100%
Correct rate of testing set Fresh
One-week
Two-week
Three-week
Four-week
72(90%) 76(95%)
73(91.25%) 76(95%)
73(91.25%) 75(93.75%)
75(93.75%) 77(96.25%)
75(93.75%) 76(95%)
188
W. Yongwei et al. / Analytica Chimica Acta 650 (2009) 183–188
Fig. 5. Predicted versus measured values of egg quality indices from QPSR models (a: Haugh unit, b: Yolk factor).
Table 4 Prediction results for egg quality on the base of E-nose signal. Quality indices
QPST Training set
HU Y
Testing set
R2
SEC
ERR%
R2
SEP
ERR%
0.93 0.94
2.15 0.01
2.55 1.5
0.91 0.93
3.74 0.02
5.37 2.89
× S3 + 4.69S1 × S4 + 0.88S1 × S5 − 4.0S1 × S7 + 6.8S1 × S8
Acknowledgements
− 0.2S2 × S5 − 2.25S2 × S6 − 2.74S3 × S4 − 6.26S3 × S6 + 11.78S3 × S8 − 1.89S4 × S5 − 4.57S4 × S8 + 3.21S5 × S8 − 25.45S6 × S7 + 14.32S6 × S8
(3) BPNN and GANN produced good predictions for egg storage time. GANN demonstrated better correction classification rates than BPNN. (4) The quadratic polynomial step regression (QPSR) algorithm established models that described the relationship between sensor signals and egg quality indices. The QPST models appeared to be of high predictive ability, with high correlation coefficients (R2 = 0.91, 0.93) between predicted and measured values.
(4)
Eq. (4): F = 8.69, P < 10−4 , R2 = 0.93, SEC = 0.08, Durbin–Watson (DW) = 2.18. The following three aspects of test to the regression equation should be checked to evaluate a model’s quality: (1) if P = 0.05 in the analysis of variance or the regression equation cannot be used; (2) whether the significant level of the partial correlation coefficient for all regression coefficients is less than 0.05; (3) whether Durbin–Watson (DW) approaches a value of 2. The inspection found that the significance level of each regression equation was less than 0.05 and DW was also close to 2; therefore, all regression equations were considered statistically valid. The test set was used to validate the prediction capabilities of the models developed (Fig. 5). The correlation coefficient between measurement and predictive values of HU and Y was 0.91 and 0.93 respectively while the standard error of prediction was 3.74 and 0.02 respectively. Results for two models developed are displayed in Table 4. The QPST models appeared to be of high predictive ability. 4. Conclusions (1) Sensor response signals in an electronic nose differed with different storage times in the case of eggs. (2) Good distinction between eggs stored for different times were obtained by PCA and LDA. Results by LDA were better than those obtained by PCA.
The authors acknowledge the financial support of the Chinese National Foundation of Nature and Science through Project 30771246, the National High Technology Research and Development Program of China through Project 2006AA10Z212, the Research Fund for the Doctoral Program of Chinese National Higher Education through Project 20060335060. References [1] V.C. Patel, R.W. McClendon, J.W. Goodrum, Artif. Intell. Rev. 12 (1998) 163–176. [2] N.D. Overfield, Egg grading as a form of quality control, Poultry-Misset 23 (1987) 10–13. [3] B. De Ketelaere, F. Bamelis, B. Kemps, E. Decuypere, J. De Baerdemaeker, World’s Poultry Sci. J. 60 (2004) 289–302. [4] V.C. Patel, R.W. McClendon, J.W. Goodrum, Appl. Eng. Agric. 12 (1996) 253–258. [5] B. De Ketelaere, P. Coucke, J. De Baerdemaeker, J. Agric. Eng. Res. 76 (2000) 157–163. [6] H.K. Cho, W.K. Choi, J.H. Paek, Trans. ASAE 43 (2000) 1921–1926. [7] A.H. Gomez, J. Wang, G.X. Hu, A.G. Pereira, Sens. Actuators B 13 (2006) 347–353. [8] A.H. Gomez, J. Wang, G.X. Hu, A.G. Pereira, LWT-Food Sci. Technol. 40 (2007) 681–689. [9] S. Saevels, J. Lammertyn, A.Z. Berna, C.D. Natale, B.M. Nicolaı, Postharvest Biol. Technol. 30 (2003) 3–4. [10] J. Brezmes, E. Llobet, X. Vilanova, G. Saiz, X. Correig, Sens. Actuators B 69 (2000) 223–229. [11] M.W. Warren, D.K. Larick, H.R. Ball, J. Food Sci. 60 (1995) 79–84. [12] M.L. Brown, D.M. Holbrook, E.F. Hoerning, M.G. Legendre, A.J. St. Angelo, Poultry Sci. 65 (1986) 1925–1933. [13] R. Dutta, E.L. Hines, J.W. Gardner, D.D. Udrea, P. Boilot, Meas. Sci. Technol. 14 (2003) 190–198. [14] M.A. Suman, G.B. Riani, E.B. Dalcanale, Sens. Actuators B 125 (2007) 40–47. [15] H. Zhang, J. Wang, S. Ye, Anal. Chim. Acta 606 (2008) 112–118. [16] Y. Liu, Z. Qiao, Acta Agric. Univ. Jiangxiensis 24 (2002) 12–16.