Вы находитесь на странице: 1из 10

Measurement 148 (2019) 106936

Contents lists available at ScienceDirect

Measurement
journal homepage: www.elsevier.com/locate/measurement

Physicochemical properties of Iranian ziziphus honey and emerging


approach for predicting them using electronic nose
Saeed Faal ⇑, Mohammad Loghavi, Saadat Kamgar
Department of Biosystems Engineering, School of Agriculture, Shiraz University, Shiraz 71441-65186, Iran

a r t i c l e i n f o a b s t r a c t

Article history: The aim of this paper is development and evaluation of an electronic nose as a non-destructive system for
Received 29 May 2019 predicting the physicochemical parameters of honey including ash content (AC), free acidity (FA), mois-
Received in revised form 27 July 2019 ture content (MC) and pH based on its aroma. To verify the authenticity of the honey samples the HPLC
Accepted 8 August 2019
analysis carried out. Artificial neural network (ANN) and Support Vector Machine (SVM) models were
Available online 12 August 2019
applied to the features extracted from the sensors responses to predict AC, FA, MC and pH. The developed
models for prediction of the physicochemical parameters showed high performance based on the honey
Keywords:
headspace gases and the best results for predicting the AC, FA, MC and pH obtained from the ANN model
Electronic Nose
Honey
which the R2 values were 0.838, 0.918, 0.926 and 0.933, respectively. The results proved feasibility and
Metal oxide semiconductor applicability of using the developed e-nose as a non-destructive system for predicting the physicochem-
Physicochemical attributes ical properties of honey.
Ó 2019 Elsevier Ltd. All rights reserved.

1. Introduction (sucrose, glucose and fructose), color and hydroxyl methyl furfural
(HMF) [6,7] which have been used to differentiate between differ-
Usually, the sweet taste of honey makes it easy to forget about ent species of honey or adulteration detection. The precise of these
the health benefits of that and its replacement with white sugar as techniques have been proven but these techniques are expensive,
a natural alternative. Honey has a wide range of health benefits for time-consuming and destructive which they can’t be used in the
people who consume it consistently and it contains wonderful real time sorters. In the industrial quality control applications,
properties that can significantly enhance the overall health [1]. managers should utilize the type of test procedures that can handle
Honey quality is a complex concept referring to chemical com- the large number of samples with low price. E-nose is one of the
pounds that make honey acceptable for global standards. The analytical tool, that recently proposed by numerous researchers
chemical compounds and specific aroma of honey are affected by for evaluating the food quality in the food industry [8]. E-nose is
many different factors, such as geographical origin, environmental a device that identifies the specific components of an odor coupled
conditions, botanical and climate. Basically, food quality evaluation with pattern recognition methods to extracts sample specifications
performs by using the conventional methods such as testing its and characterizations.
structure (moisture content, color, flavor and texture) or by means For the sensory properties of foods, aroma plays an important
of advanced analytical techniques such as gas chromatography role and several studies are reporting the potential utilization of
(GC), atomic absorption spectrophotometry, high performance liq- the e-nose in food industry such as identifying the geographical
uid chromatography (HPLC) and gas chromatography-mass spec- origin of the virgin olive oil [9], microbial contamination in pro-
trometry (GC-MS) analysis with solid phase micro extraction [2– cessed tomatoes [10], detecting the adulteration in milk [11], esti-
5]. Usually the honey quality parameters are measured by the con- mating the saffron adulteration [12], prediction of rapeseed quality
ventional analytical techniques and there are many studies report- [13], Real-time aroma monitoring of mint (Mentha spicata L.) [14],
ing on the measurement of physicochemical properties of honey, predicting the peach freshness [15] tracing the tomato-juice qual-
such as moisture content, pH, ash content, enzyme activities of ity [16], identification of tea quality [17], detecting the adulter-
invertase, electrical conductivity, acidity, sugar composition ation in honey [18], discriminating the honey based on their
geographical and botanical [19], apple quality control [20] and pre-
dicting the banana quality properties [21].
⇑ Corresponding author. Considering the successful reports researchers have been
E-mail addresses: s.faal@shirazu.ac.ir (S. Faal), Loghavi@shirazu.ac.ir reported on using the e-nose in the food industry the aim of this
(M. Loghavi), Kamgar@shirazu.ac.ir (S. Kamgar).

https://doi.org/10.1016/j.measurement.2019.106936
0263-2241/Ó 2019 Elsevier Ltd. All rights reserved.
2 S. Faal et al. / Measurement 148 (2019) 106936

paper was to evaluate the potential of an e-nose for finding the honey is the content of all free acids, expressed in milliequivalents
relationships between AC, FA, MC, pH and honey headspace gases kg1. All the experiments were repeated in ten replications.
and utilizing the chemometric methods to predict AC, FA, MC and
pH. 2.3. E-nose system

The main components of the developed e-nose system are: sen-


2. Materials and methods
sor chamber, gas sensors, pumps and pipes for circulating the air
and aroma, sample chamber, data preprocessing program written
2.1. Honey samples and chemical materials
in LabVIEW software and data acquisition system (Fig. 1a). For
designing sensor array about twenty different MOS gas sensors
Ziziphus honey is one of the best and most desirable honey in
available in Iran market were purchased and each of them evalu-
the southern regions of Iran. One hundred eleven different honey
ated by means of different honey samples and finally eight sensors
Samples directly collected from beekeepers from major ziziphus
which were more sensitive to the honey head space gases were
honey producing regions (Zahedan (S1), Bushehr (S2), Abadan
selected and were assembled inside the sensor chamber. The sen-
(S3), Ahvaz (S4), Bandar Abbas (S5), Iran Shahr (S6), Jahrom (S7),
sor array contains two TGS sensors made in by Figaro Engineering
Jiroft (S8), Shiraz (S9) and Kerman (S10)) in the summer season
(Osaka, Japan) and six MQ sensors made in HANWEI Electronics
of 2016. All chemicals and solvents used for analysis were pur-
Co., Ltd. (Henan, China). Table 1 shows the sensors specifications.
chased from a reputable company: Sodium hydroxide, Phenolph-
A digital temperature and humidity sensor (DHT-11) were added
thalein and HPLC-grade water were obtained from Merck
inside the sensor chamber to check up the conditions of the sensor
(Darmstadt, Germany).
chamber. A scheme of the test station is presented in Fig. 1b that
shows how the system works. The experimental conditions for E-
2.2. Physicochemical analysis nose are given as follows it is divided into two phases: (1) purging
and (2) saturation and measurement. About 250 g of the honey was
2.2.1. HPLC analysis poured in an Erlenmeyer flask (500 ml) at the temperature of
To verify the authenticity of the honey purchased from bee- 25 °C ± 3 °C and was sealed by plastic for the headspace generation.
keepers, the analysis of detecting glucose, fructose and sucrose A computer program was wrote to control the solenoid valves to
contents were performed by using the Knauer HPLC equipped with circulate the air through different paths based on the system oper-
the 1000 pump, a manual chromatographic valve, an on-line vac- ation stage. Before the sensor chamber was exposed to sample
uum degasser and a thermostatic column compartment. The col- volatiles it was cleaned with the flow of fresh dry air. In this stage,
umn was a Eurocat PB column (300–8 mm, 10 lm). The mobile solenoid valves 1, 2 and 4 were opened and the solenoid valve 3
phase was HPLC-grade water and the flow-rate was maintained was closed and all the pumps were running. During this stage
at 0.5 ml min1 and the column was operated at 75 °C. Every injec- the dry air flowed through the sample and sensor chambers to
tion was performed by injecting 40 lL of the prepared sample and clean them and at the end was exhausted from air outlet pipe to
it took twenty minutes between the two injections for re- the atmosphere. This operation takes about 600 (s). In the next
equilibration. stage, pump 1 was turned off but the pumps 2 and 3 were still run-
ning and solenoid valves 1 and 4 were closed and solenoid valves 2
2.2.2. pH and 3 were opened. During this stage the honey volatiles became
For measuring the pH, it was evaluated in a solution of 10 g of saturated and were circulated throw the sensor chamber. The sen-
honey in 75 ml of CO2 free distilled water, using a FisherbrandTM sors were exposed to sample volatiles and the changes in sensors’
accumetTM AE150 pH Meter with custom buffers, all the experi- responses were acquired by the data acquisition system for 600 (s),
ments were repeated in ten replications. within this time not only the headspace gasses were circulated
through the sensor chamber they also became saturated. The
steady-state sensor responses occurred at around 550 (s) and this
2.2.3. Moisture content
time was considered as the response time. After this phase again
The moisture content was measured by refractometry, using a
the sensors and sample chambers that were saturated with the
WYA-2S Digital Abbe refractometry. All the measurements were
honey headspace gases were cleaned by the fresh dry air for
carried out at the 20 °C and all the experiments were repeated in
600 s so the sensors response return to their baseline value.
ten replications. For each measurement, two grams of sample
was poured on the glass surface of refractometry.
2.4. LabVIEW program

2.2.4. Ash content Sensors output signals were acquired using an Arduino Mega
To determine the ash content, a crucible was weighed with a 2560 microcontroller. The program, for communicating the data
±0.0001 g scale and after it reached the constant weight, 5 g of with the computer was NI LabVIEW 2017 to acquire data from
honey samples were weighed and then samples burned on the the Arduino and convert the analog output to digital for prepro-
direct flame. After samples were completely burned and they did cessing them and a user-friendly graphical user interface (GUI)
not produce any smoke they were placed in a furnace (Model was designed inside the LabView program. Fig. 2(a and b) show
LEF-304P, Lab Tech, South Korea) and were calcined overnight at the designed GUI and the algorithm of the LabVIEW program that
600 °C. All the experiments were repeated in ten replications. includes: A) Read count (ms) to determine the reading intervals. B)
VIZA COM port to select and activate the serial port. C) Start button
2.2.5. Free acidity to start reading data from Arduino. D) Raw data from Arduino
Free acidity was determined by potentiometric titration and which displays raw information. E) Data array which represents
according to the harmonized methods of the European Honey data after converting from string to array. F) Real-time chart and
Commission [22]. 10 g sample was solved in 75 ml of CO2 free dis- measurement gauges (these two parts display simultaneously the
tilled water in a 250 ml beaker and it was titrated with 0.1 M NaOH values of the data on the chart and measurement gauges). G) Stop
to pH 8.30 (operator have to wait 120 s of starting the titration to button to stop the processing and storing all the data as an excel
determine that the read number is constant). The free acidity of file on the system.
S. Faal et al. / Measurement 148 (2019) 106936 3

Fig. 1. a) The developed e-nose system b) schematic of developed e-nose system.

Table 1
Specifications of the sensors utilized in the sensor chamber.

Sensor Detection range (PPM) Detect


TGS-2610 500–10000 High sensitivity to LP and its component gases
TGS-2620 500–5000 For toluene, alcohol, xylene, other volatile organic vapors
MQ-2 200–5000 Butane, Methane, LPG, smoke
MQ-3 0.05–10 Ethanol, Alcohol, smoke
MQ-4 200–10000 CNG Gas, Methane
MQ-5 200–10000 LPG, Natural gas
MQ-6 200–10000 butane gas, LPG
MQ-136 1–200 Hydrogen Sulfide gas

2.5. Signal preprocessing ized [23]. Eq. (1) represents the fractional method, in this equation
Xs (0) (sensors minimum responses when they exposed to dry air)
The purpose of data preprocessing is to extract useful informa- was subtracted from the Xs (t) (sensor responses while exposed
tion from the sensors responses and using them as input data for with honey headspace gases) and then the calculated value was
pattern recognition analysis. The preprocessing technique usually divided by the Xs (0).
depends on the sensors type, three general stages can be identified:
xs ðtÞ  xs ð0Þ
baseline manipulation, normalization and compression [23]. ys ðtÞ ¼ ð1Þ
xs ð0Þ
2.5.1. Baseline manipulation
Three baseline manipulation methods are commonly used: dif- 2.5.2. Compression
ferential, relative, and fractional. In this study the fractional The sensor-array response should be compressed down to a few
method was chosen as the baseline manipulation because the out- descriptors. In Parameter-extraction methods, two features of the
put of sensor responses (Ys (t)) became dimensionless and normal- signal obtained from particular sensors were extracted: the maxi-
4 S. Faal et al. / Measurement 148 (2019) 106936

Fig. 2. a) GUI designed in LabVIEW b) diagram of the Code written in the LabVIEW.

mum value of response time for each sensor (MVRT) and the area 2.5.3. Normalization
under the response time curve for each sensor (AURTC). Each mea- Normalization operates across the entire database for a single
surement was described by sixteen parameters (i.e., 8 sensors  2 sensor, and are generally employed to compensate for differences
parameters per sensor) and all the measurements repeated ten ðkÞ
in sensor scaling. It means it will denote by xs the response of sen-
time for each sample. sors to the k-th example in the database. After data normalization,
S. Faal et al. / Measurement 148 (2019) 106936 5

the value for each individual sensor normalized between [0, 1] Moisture content like other physicochemical properties is
ðkÞ related to botanical origin, climatic conditions and the season of
(ys ). This is simply done by subtracting the minimum and dividing
by the range of the sensor across the entire database (Eq. (2)) [23]. the year. In the summer season the climate of Bushehr (S2), Aba-
dan (S3), Ahvaz (S4) and Bandar Abbas (S5) are sultry and other
ðkÞ ðkÞ
xs  min8k ½xs  cities (Zahedan (S1), Iran Shahr (S6), Jahrom (S7), Jiroft (S8), Shiraz
yðkÞ
s ¼
h i ð2Þ
ðkÞ ðkÞ (S9) and Kerman (S10)) have warm-dry climate, according to the
max8k xs  min8k ½xs 
results presented in Table 2 the moisture content has affected by
climate change and the moisture content of sultry cities are higher
than the cities with warm-dry climate. The values of moisture con-
2.6. Prediction and classification using machine learning techniques tent analysis were between 15.68% and 21.3% (mean
value = 17.87), the obtained values are similar with those obtained
In this study, artificial neural network (ANN) and support vector by [29] who found the range of moisture content between 16.6%
machine (SVM), was utilized for predicting the physicochemical and 23.3% for thirty different honey samples. According to the
parameter of honey samples and finally, all the results were com- Codex Alimentarius Commission [28], the moisture content of
pared to find the best classifier. For both SVM and ANN classifiers honey should be less than specific value (>21%) because of safety
10-fold cross-validation technique was performed to compare the and shelf-life of honey. By increasing the moisture content the
accuracy of the models. probability of yeast fermentation during the storage time
increases, therefore, the amount of acidity increases [26]. Accord-
2.6.1. Support vector machine (SVM) ing to the moisture content results presented in Table 2 some sam-
In the SVM analysis, a hyper plane is designed and all the train- ples (one sample in the group of S2, and three samples in the group
ing vectors classify into different classes and the best result accurse of S5) exceeded the limitation suggested by Codex Alimentarius
when the hyper plane has the maximum margin from each class Commission [28].
and least error. In machine learning, one of the known branches It can be seen from the results of ash content, all the results are
of SVM is the kernel method that is a class of algorithms for pattern within 0.07% and 0.29% (mean value = 0.182) that are lower than
recognition analysis [24]. In this study different types of kernels the limit allowed for floral honey (0.6%), and indicate originality
have been used in SVM and compared with each other to recognize of honey samples [30]. The similar values were reported by [26],
which kernel is better for prediction physicochemical properties. who found the range of 0.02–0.46% for the ash content.
Although the sugar content of honey samples measured with
HPLC experiment are presented in Table 2, the values of summa-
2.6.2. Artificial neuron network (ANN)
tion of glucose and fructose were between 70.55% and 79.13%
In this study two features: MVRP and AURTC were utilized for
(mean value = 75.22) and the sucrose values were between 2.94%
the input layer neurons and they were used to predict the AC,
and 5.14% (mean value = 4.20). The similar values were reported
FA, MC and pH. In the hidden layer for choosing the most efficient
by [31,32], who found the range of 29.80–65.50%, 25.93–35.98%
number of neurons, the training step started with one neuron and
and 0.4–8.80% for fructose, glucose and sucrose respectively.
was stopped at fifteenth neurons. According to the results, the low-
According to the Codex Alimentarius Commission the summation
est root mean square error value (RMSE) was expressed the best
of fructose and glucose should be above 65% and the sucrose values
number of nodes in the hidden layer. All the SVM and ANN analysis
should be below 5%, the minimum amount of fructose and glucose
were performed using Weka version 3.9 software.
content for honey samples was 70.55% that was in the standard
range but for the sucrose results there were some samples (three
3. Results and discussion samples in the group of S6) which exceeded the limitation sug-
gested by Codex Alimentarius Commission [28].
3.1. Physicochemical parameters On the basic of the results discussed above it concluded to elim-
inate the samples which exceeded the limitation suggested by
The physicochemical parameters obtained from the experi- Codex Alimentarius Commission [28], so the final dataset consisted
ments are presented in Table 2. The pH is a significant factor that of 104 samples.
affects the acidity, texture (since acidity can influence its stability).
pH of honey can be changed by the storage and harvesting condi- 3.2. Honey aroma and its relationship with AC, FA, MC and PH
tions and effects to the shelf-life of honey [25]. The limitation for
pH has not yet been discussed by any regulatory committees [7], The relationships (Pearson correlation coefficients) between
the results of pH analysis showed that the pH value was between honey aroma and AC, FA, MC and PH are presented in the Table 3.
3.83 and 5.06 (mean value = 4.28). These values are similar with In this case, two features MVRT and AURTC were analyzed. Accord-
the pH values (3.34–5.33) reported by Silva, Sousa and Taveira ing to the results, all the sensors outputs showed significant corre-
[26]. lations between aroma and AC, FA, MC and pH except MQ-3 gas
Free acidity is one of the important compounds of honey and it’s sensor that showed no significant correlations between honey
effected by the presence of lactone, inorganic ions, organic acids aroma and MC and pH for MVRT feature and no significant corre-
and internal esters [7], although the climatic factors, harvesting lations between honey aroma and AC, MC and pH for AURTC fea-
season and botanical origin effect on the physicochemical proper- ture. On the basis of the above results it can be concluded that in
ties of honey, these can explain the difference in acidity between the sensor array, the MQ-3 gas sensor considering the fabrication
different samples [27]. According to the Codex Alimentarius Com- cost and reduction in computational process time can be elimi-
mission [28], the maximum value permitted for the acidity is nated from the sensor array in further researches.
50 meq kg1 and the results of free acidity measurement of all
honey samples ranged from 17.24 meq kg1 to 35.78 meq kg1 3.3. Predicting physicochemical parameters by e-nose response
(mean value = 28.37) and according to the results all of the tested
honey samples had the lower value comparing to the permitted The main goal of this research was to demonstrate the applica-
level (50 meq kg1). Similar results for honey free acidity were bility of the developed e-nose system as a simple and non-
reported by Silva, Sousa and Taveira [26]. destructive tool for honey aroma measurement and prediction
6 S. Faal et al. / Measurement 148 (2019) 106936

Table 2
Results of physicochemical parameters of pure honey samples (mean ± standard deviation).

Sample Sucrose Glucose Fructose pH Moisture Free acidity (meq/kg) Ash


(g 100 g1) (g 100 g1) (g 100 g1) (g 100 g1) (g 100 g1)
S1 (n = 12) Mean ± SD 3.98 ± 0.2 27.57 ± 1 48.84 ± 0.88 4.03 ± 0.1 16.57 ± 0.04 20.17 ± 0.53 0.12 ± 0.02
MIN 3.5 25.84 47.51 3.94 16.52 19.68 0.1
MAX 4.21 28.22 49.82 4.23 16.65 21.57 0.14
Samples exceeding limits – – – – – – –
S2 (n = 11) Mean ± SD 4.64 ± 0.23 27.31 ± 0.92 47.44 ± 0.81 4.47 ± 0.07 19.43 ± 0.52 33.83 ± 0.72 0.15 ± 0.01
MIN 4.23 25.49 45.44 4.38 19.38 32.99 0.12
MAX 4.72 28.15 48.39 4.63 21.09 35.18 0.18
Samples exceeding limits – – – – 1 – –
S3 (n = 11) Mean ± SD 4.53 ± 0.27 30.33 ± 0.8 44.54 ± 0.85 4.56 ± 0.09 19.49 ± 0.04 24.43 ± 0.79 0.14 ± 0.01
MIN 4.09 28.77 42.67 4.46 19.44 23.51 0.12
MAX 4.74 31.16 45.18 4.71 19.55 25.84 0.16
Samples exceeding limits – – – – – – –
S4 (n = 13) Mean ± SD 4.15 ± 0.26 24.9 ± 0.82 49.34 ± 0.7 4.71 ± 0.09 19.88 ± 0.05 30.87 ± 0.69 0.18 ± 0.01
MIN 3.82 22.95 47.51 4.63 19.86 29.97 0.15
MAX 4.22 25.87 49.58 4.9 19.98 31.91 0.21
Samples exceeding limits – – – – – – –
S5 (n = 12) Mean ± SD 3.61 ± 0.27 28.45 ± 0.9 45.34 ± 0.64 4.86 ± 0.09 20.07 ± 0.48 30.17 ± 0.83 0.21 ± 0.02
MIN 3.24 27.56 43.35 4.8 20.02 29.23 0.18
MAX 3.91 29.41 46.3 5.06 21.3 31.05 0.24
Samples exceeding limits – – – – 3 – –
S6 (n = 12) Mean ± SD 4.86 ± 0.21 33.5 ± 0.81 44.04 ± 1.05 3.95 ± 0.11 16.06 ± 0.05 33.2 ± 0.64 0.23 ± 0.02
MIN 4.54 32.43 42.91 3.97 16.02 32.36 0.21
MAX 5.14 34.46 44.93 4.13 16.15 34.7 0.26
Samples exceeding limits 3 – – – – – –
S7 (n = 10) Mean ± SD 3.84 ± 0.27 26.37 ± 0.89 46.18 ± 0.86 3.91 ± 0.07 17.67 ± 0.04 34.7 ± 0.82 0.26 ± 0.02
MIN 3.45 24.4 44.58 3.83 17.64 33.88 0.24
MAX 4.09 27.29 46.63 4.03 17.77 35.78 0.29
Samples exceeding limits – – – – – – –
S8 (n = 10) Mean ± SD 4.41 ± 0.28 33.28 ± 0.96 43.62 ± 0.92 4.13 ± 0.08 17.27 ± 0.05 26.9 ± 0.66 0.19 ± 0.02
MIN 3.96 31.57 41.76 4.08 17.23 26.07 0.16
MAX 4.68 34.19 43.74 4.27 17.37 28.12 0.21
Samples exceeding limits – – – – – – –
S9 (n = 11) Mean ± SD 3.35 ± 0.16 26.73 ± 0.93 50.8 ± 0.9 4.15 ± 0.1 15.73 ± 0.03 18.21 ± 0.72 0.1 ± 0.02
MIN 2.94 24.76 49.02 4.07 15.68 17.24 0.07
MAX 3.56 27.22 51.4 4.33 15.83 19.27 0.13
Samples exceeding limits – – – – – – –
S10 (n = 9) Mean ± SD 4.66 ± 0.1 28.04 ± 0.95 45.62 ± 0.98 4.02 ± 0.08 16.3 ± 0.06 31.23 ± 0.67 0.24 ± 0.03
MIN 4.16 26.67 44.24 3.95 16.28 30.26 0.21
MAX 4.79 28.96 45.94 4.2 16.4 31.92 0.27
Samples exceeding limits – – – – – – –
Total (n = 111)

the AC, FA, MC and pH based on the data obtained from the honey structure obtained by using the best results for predicting the
aroma. Accordingly, ANN and SVM machine learning techniques pH, AC, FA and MC features.
were used for Prediction and classification.

3.3.1. Artificial neural network results 3.3.2. Support vector machine results
In the input layer of the ANN analysis the preprocessed data It is important to select the right kernel of SVM analysis which
were utilized in order to predict the physicochemical parameters enhances SVM’s performance for honey classification. In this study,
of honey. In the hidden layer for choosing the most efficient num- four common types of kernel functions were used including poly-
ber of neurons, the training step started with one neuron and was nomial, normalized polynomial, Pearson VII universal kernel
stopped at fifteenth neurons, each training step was repeated ten (PUK) and radial basis function (RBF) [33]. For analyzing the data
times and at the end the mean value was calculated. The results set each training set repeated ten times and the mean value was
of ANN structures using the MVRT and AURTC features as input calculated, the results of SVM analyzed with the different kernel
data for predicting the physicochemical parameters are presented using the MVRT and AURTC features as input data for prediction
in Table 4. According to the results the hidden layer with 2, 7, 13 of the physicochemical parameters are presented in Table 5. The
and 10 neurons utilizing MVRT feature and hidden layer with 10, results of SVM for MVRT feature reveals that the RBF function pre-
5, 2 and 6 neurons utilizing AURTC feature had the lowest RMSE sents lowest RMSE (0.195 ± 0.06), and highest accuracy
and highest accuracy with best performance for predicting the (88.85 ± 0.73) for prediction of AC, also the PUK function compar-
AC, FA, MC and pH, respectively. Comparing the best result of both ing to the other kernels offers the lowest RMSE (1.272 ± 0.4,
features showed that prediction of the AC, FA, MC and pH utilizing 0.752 ± 0.28) and the highest accuracy (94.8 ± 1.02, 94.72 ± 0.94)
MVRT feature have lower RMSE and higher accuracy against the for predicting the FA and MC, respectively and the normalized
AURTC feature and suggests better performance for predicting polynomial function presents the lowest RMSE (0.75 ± 0.15), and
the physicochemical parameters of honey, Fig. 3 shows the ANN the highest accuracy (94.2 ± 1.07) for predicting the PH.
S. Faal et al. / Measurement 148 (2019) 106936 7

Table 3
The correlation between eight sensors and physicochemical parameters for different features.

The maximum value of response time for each sensor (MVRT)


pH MC FA AC MQ 6 MQ 2 MQ 3 MQ 5 MQ 136 MQ 4 TGS 2610 TGS 2620
pH 1 0.871** 0.956** 0.936** 0.765** 0.713** 0.169 0.865** 0.849** 0.825** 0.856** 0.864**
MC 0.871** 1 0.813** 0.804** 0.674** 0.635** 0.273 0.767** 0.755** 0.720** 0.748** 0.746**
FA 0.956** 0.813** 1 0.911** 0.802** 0.754** 0.528* 0.887** 0.878** 0.851** 0.884** 0.884**
AC 0.936** 0.804** 0.911** 1 0.693** 0.657** 0.381* 0.794** 0.781** 0.755** 0.781** 0.789**
MQ 6 0.765** 0.674** 0.802** 0.693** 1 0.963** 0.804** 0.864** 0.954** 0.850** 0.957** 0.853**
MQ 2 0.713** 0.635** 0.754** 0.657** 0.963** 1 0.814** 0.764** 0.923** 0.744** 0.883** 0.747**
MQ 3 0.169 0.273 0.528* 0.381* 0.804** 0.814** 1 0.511** 0.673** 0.491** 0.697** 0.533**
MQ 5 0.865** 0.767** 0.887** 0.794** 0.864** 0.764** 0.511** 1 0.940** 0.995** 0.958** 0.987**
MQ 136 0.849** 0.755** 0.878** 0.781** 0.954** 0.923** 0.673** 0.940** 1 0.925** 0.977** 0.935**
MQ 4 0.825** 0.720** 0.851** 0.755** 0.850** 0.744** 0.491** 0.995** 0.925** 1 0.943** 0.979**
TGS 2610 0.856** 0.748** 0.884** 0.781** 0.957** 0.883** 0.697** 0.958** 0.977** 0.943** 1 0.962**
TGS 2620 0.864** 0.746** 0.884** 0.789** 0.853** 0.747** 0.533** 0.987** 0.935** 0.979** 0.962** 1
The area under the response time curve (AURTC)
pH 1 0.87** 0.95** 0.93** 0.74** 0.60** 0.19 0.83** 0.80** 0.79** 0.82** 0.83**
MC 0.871** 1 0.813** 0.804** 0.633** 0.556** 0.04 0.729** 0.715** 0.677** 0.694** 0.717**
FA 0.956** 0.813** 1 0.911** 0.795** 0.666** 0.295* 0.880** 0.848** 0.839** 0.865** 0.870**
AC 0.936** 0.804** 0.911** 1 0.670** 0.562** 0.15 0.772** 0.740** 0.729** 0.748** 0.764**
MQ 6 0.749** 0.633** 0.795** 0.670** 1 0.903** 0.733** 0.834** 0.958** 0.802** 0.940** 0.870**
MQ 2 0.609** 0.556** 0.666** 0.562** 0.903** 1 0.796** 0.626** 0.913** 0.578** 0.760** 0.664**
MQ 3 0.19 0.04 0.295* 0.15 0.733** 0.796** 1 0.316* 0.615** 0.291* 0.519** 0.376**
MQ 5 0.839** 0.729** 0.880** 0.772** 0.834** 0.626** 0.316* 1 0.879** 0.995** 0.966** 0.991**
MQ 136 0.800** 0.715** 0.848** 0.740** 0.958** 0.913** 0.615** 0.879** 1 0.846** 0.940** 0.904**
MQ 4 0.795** 0.677** 0.839** 0.729** 0.802** 0.578** 0.291* 0.995** 0.846** 1 0.952** 0.983**
TGS 2610 0.822** 0.694** 0.865** 0.748** 0.940** 0.760** 0.519** 0.966** 0.940** 0.952** 1 0.980**
TGS 2620 0.831** 0.717** 0.870** 0.764** 0.870** 0.664** 0.376** 0.991** 0.904** 0.983** 0.980** 1

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


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

Table 4
Results of ANN analysis of different number of nodes in the hidden layer (NHL) for predicting the different physicochemical parameters of honey (mean ± standard deviation).

NHL RMSE R (%)


pH MC FA AC pH MC FA AC
MVRT feature
1 0.535 ± 0.13 0.736 ± 0.06 1.372 ± 0.3 0.194 ± 0.07 96.12 ± 1.35 94.83 ± 1.36 95.64 ± 0.92 87.24 ± 1.36
2 0.461 ± 0.13 0.593 ± 0.11 1.184 ± 0.36 0.144 ± 0.06 96.26 ± 1.03 95.63 ± 1.18 95.72 ± 0.73 91.57 ± 1.41
3 0.438 ± 0.08 0.613 ± 0.08 1.493 ± 0.27 0.155 ± 0.05 96.36 ± 0.97 95.46 ± 1.48 95.84 ± 0.96 90.76 ± 1.41
4 0.411 ± 0.13 0.595 ± 0.12 1.157 ± 0.36 0.147 ± 0.05 96.5 ± 0.88 95.49 ± 1.27 95.75 ± 0.54 91.43 ± 1.68
5 0.4 ± 0.15 0.477 ± 0.1 1.112 ± 0.21 0.152 ± 0.06 96.42 ± 1.2 96.09 ± 1.17 95.82 ± 0.94 91.08 ± 1.66
6 0.436 ± 0.1 0.459 ± 0.1 1.163 ± 0.27 0.165 ± 0.04 96.29 ± 1.24 96.06 ± 1.3 95.75 ± 0.63 89.88 ± 1.17
7 0.405 ± 0.06 0.467 ± 0.09 1.099 ± 0.3 0.147 ± 0.07 96.39 ± 1.13 96.05 ± 1.56 95.85 ± 1.08 91.42 ± 1.5
8 0.397 ± 0.11 0.457 ± 0.13 1.101 ± 0.25 0.149 ± 0.06 96.46 ± 0.96 96.07 ± 1.19 94.91 ± 0.67 91.24 ± 1.71
9 0.413 ± 0.13 0.445 ± 0.12 1.109 ± 0.31 0.151 ± 0.06 96.45 ± 1.47 96.21 ± 0.98 94.86 ± 0.86 91.1 ± 1.78
10 0.389 ± 0.08 0.448 ± 0.13 1.174 ± 0.28 0.147 ± 0.06 96.64 ± 1.06 95.26 ± 0.79 95.75 ± 0.7 91.44 ± 1.02
11 0.399 ± 0.08 0.465 ± 0.11 1.129 ± 0.3 0.156 ± 0.06 96.54 ± 1.24 96.14 ± 1.27 95.65 ± 0.66 90.75 ± 1.47
12 0.423 ± 0.09 0.508 ± 0.13 1.187 ± 0.32 0.147 ± 0.04 96.48 ± 1.01 96.1 ± 1.2 95.75 ± 0.93 91.43 ± 1.39
13 0.403 ± 0.14 0.434 ± 0.15 1.117 ± 0.31 0.163 ± 0.05 96.47 ± 0.94 96.26 ± 0.94 95.35 ± 0.84 90.07 ± 1.15
14 0.406 ± 0.14 0.457 ± 0.11 1.134 ± 0.35 0.166 ± 0.06 96.56 ± 1.19 96.11 ± 1.09 95.82 ± 0.7 89.9 ± 1.64
15 0.405 ± 0.11 0.474 ± 0.14 1.159 ± 0.21 0.146 ± 0.07 96.48 ± 1.29 96.07 ± 1.27 95.83 ± 0.85 91.46 ± 1.66
AURTC feature
1 0.496 ± 0.2 0.713 ± 0.15 2.668 ± 0.37 0.242 ± 0.05 93.73 ± 0.97 93.75 ± 1.01 93.61 ± 1.2 85.74 ± 1.94
2 0.515 ± 0.17 0.549 ± 0.18 1.63 ± 0.46 0.195 ± 0.08 93.62 ± 0.88 94.08 ± 0.48 94.35 ± 1.02 88.53 ± 2.48
3 0.477 ± 0.22 0.571 ± 0.21 2.015 ± 0.37 0.169 ± 0.07 93.82 ± 1.13 93.99 ± 0.72 94.55 ± 0.82 90.42 ± 1.75
4 0.483 ± 0.15 0.634 ± 0.14 1.784 ± 0.35 0.194 ± 0.05 93.8 ± 0.92 93.91 ± 0.82 94.61 ± 1.13 89.27 ± 2.08
5 0.479 ± 0.16 0.554 ± 0.21 1.477 ± 0.53 0.202 ± 0.08 92.83 ± 1.06 93.08 ± 0.93 94.72 ± 1.19 88.87 ± 1.9
6 0.476 ± 0.13 0.684 ± 0.22 1.532 ± 0.35 0.203 ± 0.04 93.83 ± 1.15 93.57 ± 1 94.67 ± 1.03 88.82 ± 1.86
7 0.515 ± 0.15 0.734 ± 0.18 1.669 ± 0.33 0.206 ± 0.06 93.63 ± 1.09 93.55 ± 0.83 94.66 ± 1.08 88.66 ± 1.9
8 0.48 ± 0.19 0.556 ± 0.17 1.604 ± 0.37 0.204 ± 0.05 93.81 ± 0.67 94.06 ± 0.93 94.68 ± 0.86 89.4 ± 1.91
9 0.496 ± 0.21 0.602 ± 0.23 1.574 ± 0.44 0.174 ± 0.06 93.72 ± 0.98 93.86 ± 0.63 94.12 ± 0.9 90.94 ± 2.02
10 0.552 ± 0.12 0.787 ± 0.17 1.635 ± 0.41 0.166 ± 0.06 93.45 ± 1.07 93.28 ± 0.87 94.67 ± 0.89 91.1 ± 2.14
11 0.521 ± 0.18 0.713 ± 0.22 1.54 ± 0.43 0.172 ± 0.06 93.59 ± 0.77 93.52 ± 1.04 94.71 ± 1.04 90.64 ± 2.08
12 0.522 ± 0.18 0.674 ± 0.18 1.745 ± 0.27 0.178 ± 0.07 93.58 ± 1.26 93.77 ± 0.98 94.68 ± 1.11 89.52 ± 2.14
13 0.496 ± 0.15 0.617 ± 0.19 1.625 ± 0.36 0.171 ± 0.06 93.72 ± 0.71 93.91 ± 1.01 94.7 ± 1.06 90.92 ± 1.88
14 0.481 ± 0.18 0.581 ± 0.2 1.669 ± 0.51 0.179 ± 0.06 93.8 ± 0.98 93.99 ± 0.77 94.7 ± 1.29 90.88 ± 1.48
15 0.499 ± 0.23 0.627 ± 0.19 1.842 ± 0.27 0.181 ± 0.06 93.71 ± 0.92 93.78 ± 0.48 94.66 ± 1.05 90.45 ± 1.67
8 S. Faal et al. / Measurement 148 (2019) 106936

Fig. 3. The ANN structure obtained by using the best results for predicting the pH, AC, FA and MC features.

Table 5
Results of SVM analysis with different kernel functions for predicting the physicochemical parameters of honey (mean ± standard deviation).

Kernel Parameter-extraction methods


MVRT feature AURTC feature
pH MC FA AC pH MC FA AC
Polynomial R (%) 93.03 ± 1.13 90.1 ± 1.17 93.67 ± 1.66 87.28 ± 1.27 91.94 ± 1.29 87.09 ± 1.35 91.34 ± 1.27 86.87 ± 1.33
RMSE 0.971 ± 0.18 1.265 ± 0.21 2.152 ± 0.49 0.238 ± 0.05 0.88 ± 0.24 1.376 ± 0.31 2.249 ± 0.22 0.228 ± 0.04
Normalized polynomial R (%) 94.2 ± 1.07 93.71 ± 0.82 94.29 ± 1.09 86.41 ± 0.68 92.99 ± 0.9 88.82 ± 0.81 92.06 ± 1.14 87.66 ± 1.34
RMSE 0.75 ± 0.15 0.928 ± 0.24 1.849 ± 0.44 0.242 ± 0.07 0.714 ± 0.18 1.209 ± 0.29 2.149 ± 0.26 0.215 ± 0.07
RBF R (%) 90.99 ± 0.56 82.15 ± 0.76 92.46 ± 0.66 88.85 ± 0.73 90.7 ± 0.8 81.71 ± 0.78 90.15 ± 0.69 88.04 ± 0.59
RMSE 1.111 ± 0.19 2.02 ± 0.21 2.664 ± 0.39 0.195 ± 0.06 1.056 ± 0.18 2.024 ± 0.32 2.763 ± 0.37 0.2 ± 0.06
PUK R (%) 94 ± 0.98 94.72 ± 0.94 94.8 ± 1.02 87.7 ± 0.87 92.42 ± 0.96 92.52 ± 1.18 93.05 ± 1.21 87.21 ± 1.34
RMSE 0.811 ± 0.19 0.752 ± 0.28 1.272 ± 0.4 0.226 ± 0.05 0.789 ± 0.16 0.881 ± 0.23 1.92 ± 0.28 0.215 ± 0.05

For the AURTC feature, kernels have the same treatment as accuracy (91.57 ± 1.41, 95.85 ± 1.08, 96.26 ± 0.94 and 96.64 ± 1.06)
MVRT feature. In the AURTC feature, the RBF function showed for prediction of AC, FA, MC and pH attributes, respectively. For the
the lowest RMSE (0.2 ± 0.06) and highest accuracy (88.04 ± 0.59) ANN analysis utilizing the MVRT feature, the correlations between
for predicting the AC, but for FA and MC prediction, the PUK func- the predicted and experimental AC, FA, MC and pH values for the
tion has better results, the lowest RMSE (1.920 ± 0.28, 0.881 ± 0.23) test sets are shown in Fig. 4. As seen in the figure the prediction
and highest accuracy (93.05 ± 1.21, 92.52 ± 1.18), respectively. For of the AC, FA, MC and pH values of honey samples utilizing the data
pH prediction, the Normalized polynomial kernel offers lower extracted from developed e-nose (MVRT feature) have reasonable
RMSE (0.714 ± 0.18) and higher accuracy (92.99 ± 0.9) in compar- R2 value (0.838, 0.918, 0.926 and 0.933), respectively. The results
ison with other kernels. In the previous study Huang, Liu, Zhang showed the effectiveness of the ANN model for honey properties
and Wu [34] have investigated the application of electronic nose measurement. Previous studies have shown the efficiency of
with multivariate analysis for botanical origin identification and chemo metric methods for fruit properties measurement utilizing
quality determination of honey. They have concluded that the the e-nose. Zhang, Wang, Ye and Chang [35] have investigated
quality of honey can be predicted with high accuracy utilizing the potential of using an e-nose for predicting the quality proper-
the SVM method. ties of peach. They reported that the partial least-squares regres-
sions and principal component regression models show excellent
3.3.3. Comparing SVM and ANN results ability in describing the quality indices of the peaches also the
Both SVM and ANN classifiers offer good accuracy for predicting results demonstrate that the e-nose has the potential for measur-
the physicochemical parameters of honey utilizing both MVRT and ing fruit quality properties. Similar results were obtained by Su,
AURTC features, but in comparison of both classifier the ANN clas- Zhang, Ye, Chen, Guo, Gu and Shen [36] for predicting the quality
sifier using the MVRT feature as input data has lower RMSE properties of pulp by an electronic nose. They reported that the
(0.144 ± 0.06, 1.099 ± 0.3, 0.434 ± 0.15 and 0.389 ± 0.08) and higher e-nose utilizing the discriminant factor analysis is capable to pre-
S. Faal et al. / Measurement 148 (2019) 106936 9

Fig. 4. Correlation between the measured and predicted physicochemical parameters of honey for ANN analysis using the MVRT feature (test mode).

dict the quality properties of pulp. Our study results were showed Appendix A. Supplementary data
that the developed e-nose is a reliable device for predicting the
honey ash content, free acidity, moisture content and pH as a fast, Supplementary data to this article can be found online at
cheap and non-destructive instrument that can be handled by any- https://doi.org/10.1016/j.measurement.2019.106936.
one and has the advantage of eliminating the need for specialist
operators in comparison with the other traditional methods.
References

[1] K.S. Kumar, B. Debjit, M. Chandira, Medicinal uses and health benefits of
4. Conclusion honey: an overview, J. Chem. Pharm. Res. 2 (2010) 385–395.
[2] M. Shamsipur, N. Yazdanfar, M. Ghambarian, Combination of solid-phase
extraction with dispersive liquid–liquid microextraction followed by GC–MS
This study was carried out to investigate the capabilities of an e- for determination of pesticide residues from water, milk, honey and fruit juice,
nose for non-destructive prediction of physicochemical properties Food Chem. 204 (2016) 289–297.
of honey. Two aroma features of each sensor including the maxi- [3] O. Hernández, J. Fraga, A. Jiménez, F. Jimenez, J. Arias, Characterization of
honey from the Canary Islands: determination of the mineral content by
mum value of response time for each sensor (MVRT) and the area atomic absorption spectrophotometry, Food Chem. 93 (2005) 449–458.
under the response time curve (AURTC) were extracted and after [4] S. Parizad, A. Dizadji, M.K. Habibi, S. Winter, S. Kalantari, S. Movi, C. Lorenzo
preprocessing were utilized for prediction of ash content (AC), free Tendero, G.L. Alonso, N. Moratalla-Lopez, The effects of geographical origin and
virus infection on the saffron (Crocus sativus L.) quality, Food Chem. 295
acidity (FA), moisture content (MC) and pH. The results showed (2019) 387–394.
that the developed e-nose can be a reliable technic for predicting [5] L.G.d. Oliveira, M.H.S. Kurz, M.C.M. Guimarães, M.L. Martins, O.D. Prestes, R.
the physicochemical parameters of honey. The best result for pre- Zanella, J.N.d.S. Ribeiro, F.F. Gonçalves, Development and validation of a
method for the analysis of pyrethroid residues in fish using GC-MS, Food
dicting the AC, FA, MC and pH was found for the artificial neural
Chemistry 297 (2019).
network method applied to the MVRT feature data. The results [6] N. Spano, L. Casula, A. Panzanelli, M.I. Pilo, P.C. Piu, R. Scanu, A. Tapparo, G.
demonstrated that the developed e-nose linked with the artificial Sanna, An RP-HPLC determination of 5-hydroxymethylfurfural in honey: the
case of strawberry tree honey, Talanta 68 (2006) 1390–1395.
neural network was able to predict AC, FA, MC and pH with high
[7] P.M. da Silva, C. Gauche, L.V. Gonzaga, A.C.O. Costa, R. Fett, Honey: chemical
regression coefficients. The results demonstrate that the developed composition, stability and authenticity, Food Chem. 196 (2016) 309–323.
e-nose has the potential for assessing other food products such as [8] A. Loutfi, S. Coradeschi, G.K. Mani, P. Shankar, J.B.B. Rayappan, Electronic noses
rosewater, saffron and other volatile products that their quality is for food quality: a review, J. Food Eng. 144 (2015) 103–111.
[9] Z. Haddi, A. Amari, A.O. Ali, N.E. Bari, H. Barhoumi, A. Maaref, N. Jaffrezic-
associated with their aroma. Renault, B. Bouchikhi, Discrimination and identification of geographical origin
virgin olive oil by an e-nose based on MOS sensors and pattern recognition
techniques, Procedia Eng. 25 (2011) 1137–1140.
[10] I. Concina, M. Falasconi, E. Gobbi, F. Bianchi, M. Musci, M. Mattarozzi, M. Pardo,
Acknowledgements A. Mangia, M. Careri, G. Sberveglieri, Early detection of microbial
contamination in processed tomatoes by electronic nose, Food Control 20
(2009) 873–880.
This research was jointly supported by the scholarship from Shiraz [11] M. Tohidi, M. Ghasemi-Varnamkhasti, V. Ghafarinia, M. Bonyadian, S.S.
University. Funding for HPLC analysis was provided by the Depart- Mohtasebi, Development of a metal oxide semiconductor-based artificial
ment of Food Science and Technology of Shiraz University. nose as a fast, reliable and non-expensive analytical technique for aroma
profiling of milk adulteration, Int. Dairy J. 77 (2018) 38–46.
[12] S. Kiani, S. Minaei, M. Ghasemi-Varnamkhasti, Integration of computer vision
and electronic nose as non-destructive systems for saffron adulteration
detection, Comput. Electron. Agric. 141 (2017) 46–53.
Declaration of Competing Interest [13] M. Gancarz, J. Wawrzyniak, M. Gawrysiak-Witulska, D. Wia˛cek, A. Nawrocka,
M. Tadla, R. Rusinek, Application of electronic nose with MOS sensors to
The authors declare that they have no known competing finan- prediction of rapeseed quality, Measurement 103 (2017) 227–234.
[14] S. Kiani, S. Minaei, M. Ghasemi-Varnamkhasti, Real-time aroma monitoring of
cial interests or personal relationships that could have appeared to mint (Mentha spicata L.) leaves during the drying process using electronic
influence the work reported in this paper. nose system, Measurement 124 (2018) 447–452.
10 S. Faal et al. / Measurement 148 (2019) 106936

[15] H. Guohua, W. Yuling, Y. Dandan, D. Wenwen, Z. Linshan, W. Lvye, Study of [26] L.R. Silva, A. Sousa, M. Taveira, Characterization of Portuguese honey from
peach freshness predictive method based on electronic nose, Food Control 28 Castelo Branco region according to their pollen spectrum, physicochemical
(2012) 25–32. characteristics and mineral contents, J. Food Sci. Technol. 54 (2017) 2551–
[16] X. Hong, J. Wang, G. Qi, E-nose combined with chemometrics to trace tomato- 2561.
juice quality, J. Food Eng. 149 (2015) 38–43. [27] L.R. Silva, R. Videira, A.P. Monteiro, P. Valentão, P.B. Andrade, Honey from Luso
[17] M. Xu, J. Wang, S. Gu, Rapid identification of tea quality by E-nose and region (Portugal): physicochemical characteristics and mineral contents,
computer vision combining with a synergetic data fusion strategy, J. Food Eng. Microchem. J. 93 (2009) 73–77.
241 (2019) 10–17. [28] C. Alimentarius, Revised codex standard for honey, Codex Stan 12 (2001) 1982.
[18] A. Zakaria, A.Y.M. Shakaff, M.J. Masnan, M.N. Ahmad, A.H. Adom, M.N. Jaafar, S. [29] I. Escriche, F. Tanleque-Alberto, M. Visquert, M. Oroian, Physicochemical and
A. Ghani, A.H. Abdullah, A.H.A. Aziz, L.M. Kamarudin, A biomimetic sensor for rheological characterization of honey from Mozambique, LWT-Food Sci.
the classification of honeys of different floral origin and the detection of Technol. 86 (2017) 108–115.
adulteration, Sensors 11 (2011) 7799–7822. [30] J.W. White Jr., Honey, Advances in Food Research, Elsevier, 1978, pp. 287–374.
[19] S. Benedetti, S. Mannino, A. Sabatini, G. Marcazzan, Electronic nose and neural [31] S. Saxena, S. Gautam, A. Sharma, Physical, biochemical and antioxidant
network use for the classification of honey, Apidologie 35 (2004) 397–402. properties of some Indian honeys, Food Chem. 118 (2010) 391–397.
[20] M. Ezhilan, N. Nesakumar, K. Jayanth Babu, C.S. Srinandan, J.B.B. Rayappan, An [32] F. Tornuk, S. Karaman, I. Ozturk, O.S. Toker, B. Tastemur, O. Sagdic, M. Dogan, A.
electronic nose for royal delicious apple quality assessment – a tri-layer Kayacier, Quality characterization of artisanal and retail Turkish blossom
approach, Food Res. Int. 109 (2018) 44–51. honeys: determination of physicochemical, microbiological, bioactive
[21] A. Sanaeifar, S.S. Mohtasebi, M. Ghasemi-Varnamkhasti, H. Ahmadi, properties and aroma profile, Ind. Crops Prod. 46 (2013) 124–131.
Application of MOS based electronic nose for the prediction of banana [33] N. Cristianini, J. Shawe-Taylor, An Introduction to Support Vector Machines
quality properties, Measurement 82 (2016) 105–114. and Other Kernel-Based Learning Methods, Cambridge University Press, 2000.
[22] I.H. Commission, Harmonised Methods of the European Honey Commission, [34] L. Huang, H. Liu, B. Zhang, D. Wu, Application of electronic nose with
2016. multivariate analysis and sensor selection for botanical origin identification
[23] T. Pearce, S. Schiffman, H. Nagle, J. Gardner, Handbook of Machine Olfaction: and quality determination of honey, Food Bioprocess Technol. 8 (2015) 359–
Electronic Nose Technology, Wiley-VCH Velag GmbH & Co. KGaA, Wheinheim, 370.
UK, 2003. [35] H. Zhang, J. Wang, S. Ye, M. Chang, Application of electronic nose and statistical
[24] J. Tang, Y. Tian, A multi-kernel framework with nonparallel support vector analysis to predict quality indices of peach, Food Bioprocess Technol. 5 (2012)
machine, Neurocomputing 266 (2017) 226–238. 65–72.
[25] A. Terrab, A.F. Recamales, D. Hernanz, F.J. Heredia, Characterisation of Spanish [36] M. Su, B. Zhang, Z. Ye, K. Chen, J. Guo, X. Gu, J. Shen, Pulp volatiles measured by
thyme honeys by their physicochemical characteristics and mineral contents, an electronic nose are related to harvest season, TSS concentration and TSS/TA
Food Chem. 88 (2004) 537–542. ratio among 39 peaches and nectarines, Sci. Hortic. 150 (2013) 146–153.

Вам также может понравиться