enose iki tentang
TRANSCRIPT
-
8/9/2019 enose iki tentang
1/4
The Repeatability and Discrimination Study of
Electronic Nose Features
Mazlina MamatFaculty of Engineering and Built Environment
Universiti Kebangsaan Malaysia
43600 UKM Bangi Selangor
Salina Abdul SamadFaculty of Engineering and Built Environment
Universiti Kebangsaan Malaysia
43600 UKM Bangi Selangor
Abstract — Four different features: degree of reaction, the
fractional relative response, the rate of reaction at the initial
absorption time and the accumulative total of the reactiondegree changing were extracted from the sensor response
curves of the electronic nose measurements on four different
brands cultured milk drink. The repeatability and the
discrimination ability of these four features were analyzed. The
results show that the degree of reaction exhibits the bestrepeatability, followed by the fractional relative response, the
accumulative total of the reaction degree changing and the rateof reaction at the initial absorption time, respectively. The
same observation was also recorded in the discrimination
ability of the features where the degree of reaction and the
fractional relative response show comparable performance, theaccumulative total of the reaction degree changing shows
slightly low performance while the initial absorption time failsto perform.
Keywords-electronic nose; features; repeatability;
discriminative
I. INTRODUCTION
The electronic nose is an apparatus that is used to detectand to perform classification on odorous substances. Itdetects odor by using a gas sensor array which produce aspecific pattern for specific substance. The produced patternis further analyzed by using appropriate data analysistechnique such as Principal Component Analysis (PCA),Functional Discriminant Analysis (FDA) and ArtificialNeural Network (ANN) for classification or predictionpurposes. In electronic nose, the patterns used in the dataanalysis technique are generated from the features of thesensor response curve. There are several approaches used toextract the features. The common approach is by consideringthe steady state part or the transient response of the curve.This approach was used by many researchers in variousapplications. Panigrahi et al. used the accumulative total ofthe reaction degree changing (area under the response curve),the relative humidity during absorption and desorption, theaverage temperature during measurement and the averagevoltage value of carbon dioxide sensor to classify beefsamples into groups of unspoiled and spoiled [1]. Gomez et
al. used the ratio of conductivity at 42s of the sensorresponse to monitor the storage shelf life of tomatoes [2] andmandarin oranges [3]. Labreche et al. used the optimumvalue of the resistance relative response curve to predict theshelf life of milk [4]. The relative response of the resistancewas used by Brudzewski et al. in their research todiscriminate different types of milk processing
(pasteurization or Ultra Heat Treatment) [5]. In theresearches by El Babri and colleagues, the average value ofinitial conductance, the average value of steady stateconductance, the dynamic slope of conductance duringabsorption phase and the total of the reaction degreechanging were used to assess the quality of beef, sheep meatsand Moroccan Sardines [6, 7].
The above researches prove that numerous features canbe extracted from a sensor response curve. In practice, anypoints on the sensor response curve can be treated asfeatures, however how good it can represents the wholecurve and how stable it is must be studied. This paper
presents the analyses to determine the repeatability anddiscrimination ability of four different features obtained fromthe sensor response curves. The four features are the degreeof reaction, the fractional relative response, the rate ofreaction at the initial absorption time and the accumulativetotal of the reaction degree changing.
II. APPROACH AND METHODS
A. Electronic Nose System
The E-Nose was fabricated at the Digital Signal
Processing Laboratory, University Kebangsaan Malaysia. It
consists of 5 key components: sample chamber, sensor
chamber, data acquisition system and controller unit, power
supply and a computer. The sample chamber is a cylindricalglass bottle with 40ml volume. It was used to store the
cultured milk drink during measurement. The sample
chamber was connected to the sensor chamber via two
plastic tubes attached to an air diaphragm pump to increase
flow. The sensor chamber is an airtight box with
approximately 200ml volume and consists of 14 gas sensors
(TGS813, TGS821, TGS822, TGS825, TGS826, TGS830,
TGS2180, TGS2600, TGS2602, TGS2610, TGS2611,
TGS2612, TGS2620 and TGS6812) and one temperature
sensor (LM35DZ) (Table 1). The gas sensors operate to
produce odor print of the sample while the temperature
sensor is to monitor the sensor chamber temperature during
measurement. The voltage responses of all sensors were
amplified and adjusted to 0V to 5V analog voltage,converted to digital value and transmitted to the serial port
using a data acquisition system and controller unit. A
graphic user interface program to control and to interpret the
measurement data was developed by using C++ Builder
software.
978-1-4577-0255-6/11/$26.00 ©2011 IEEE 1185 TENCON 2011
-
8/9/2019 enose iki tentang
2/4
TABLE I. LIST OF SENSORS USED IN THE ELECTRONIC NOSE SYSTEM
Sensor Target gas
TGS813 Combustible Gases (methane, propane, butane)
TGS821 Hydrogen
TGS822 Organic Solvent Vapors (ethanol)
TGS825 Hydrogen Sulfide
TGS826 Ammonia
TGS830 Chlorofluorocarbons
TGS2180 Water vapor
TGS2600 Air Contaminants (hydrogen, carbon monoxide)
TGS2602 Air Contaminants (VOCs and odorous gases )TGS2610 LP Gas and its component gases
TGS2611 Methane
TGS2612 Methane and LP Gases
TGS2620 Alcohol and Solvent Vapors
TGS6812 Hydrogen, Methane and LP Gas
LM35DZ Temperature sensor
B. Electronic Nose Measurement and Feature Selection
The measurement starts by performing the base linecorrection to the sensors by purging the sensors withatmospheric air for 200 s. This is to ensure that the sensorsare completely free from possible contaminated odors fromprevious measurement. Then the sampling chamber whichcontains the sample was attached to the sensor chamber and
the odor was sucked into the airtight sensor chamber for200s. Next, the sensor chamber was cleaned again foranother 200 s by purging the sensor chamber with ambientair. The ambient air is used without any pretreatment andwas proven to sufficiently clean the sensors from previousmeasurement. The time required for cleaning and sampling is10 minutes (600 s = 200s + 200s + 200s) and during theprocess, the voltage reading of the sensors were acquired andsaved.
From the sensor response curve obtained in the
measurement, four different features were extracted. The
features are:
• The degree of reaction (d ),
max mind V V = − (1)
• The relative response (r ),
max min
min
( )V V r
V
−= (2)
• The rate of reaction at the initial absorption time (rt ),
250 201( )
50
t t V V rt = =−
= (3)
• The accumulative total of the reaction degreechanging (i),
0
t
t i d =
∫ (4)
The accumulative total of the reaction degree changing isequivalent to the area below the response curve. This areawas computed by using the trapezoidal rule. The sensorresponse curve and the corresponding parameters used tocompute the four features are displayed in Fig. 1.
Figure 1. The sensor response curve.
C. Samples
The samples measured in the experiment were 4
different brands of cultured milk drinks of original flavor
purchased from a local supermarket. A total of 12 bottles of
Vitagen, 10 bottles of Solivite, 10 bottles of Nutrigen and 8
bottles of Yakult were used in the experiment. To maintain
the freshness of the cultured milk drinks, they were stored
unopened in 4oC freezer and were placed in ambienttemperature about 30 minutes prior to experiment.
D. Performance criteria
The repeatability of features can be determined by
computing the Relative Standard Deviation (RSD) of each
feature. Feature with less RSD value indicates higher
repeatability. The small RSD values indicate that the sensors
show relatively good precision hence confirm the
repeatability of the measurements. The RSD of sensor i isgiven by the following equation:
i
i
i RSD
σ
µ =
(5)
where σ = standard deviation, and µ = average value of the
feature. Usually the RSD value is presented in the
percentage form.
PCA is a common pattern recognition algorithm used to
analyze data obtained from electronic nose system [8,9,10].
In particular, PCA is used to reduce the complexity of data
by computing a new, much smaller set of uncorrelated
variables which best represent the original data. This is done
by projecting the high dimensional data set in a dimensional
reduced space based on the uncorrelated and orthogonal
eigenvectors of the covariance matrix computed from the e-
nose features. These eigenvectors were called principalcomponents of the features and were arranged in sequence
where the first principal component was the one with the
greatest amount of variance, followed by the second greatest
and so on. The plot of the original data in the new space
defined by the first few principal components will give
visual interpretation on how the original data are scattered.
1186
-
8/9/2019 enose iki tentang
3/4
In particular, the plot will shows features that have small
variation appear together while the features with large
variation appear distant. Therefore PCA is able to expose
some clusters of the data naturally.
III. RESULTS AND DISCUSSION
A. The response of sensors
This analysis was conducted to examine the response of
the 14 sensors to the cultured milk drink odor.In this analysis, one measurement was conducted to each
brand of cultured milk drink. The degree of reactions for the
14 sensors in the measurements of the four brands were
recorded and presented in Fig. 2. The chart shows that 8
sensors (TGS822, TGS813, TGS821, TGS2602, TGS826,
TGS2620, TGS825, and TGS2600) give significant
response to the cultured milk drink odor while the other 6sensors (TGS830, TGS2180, TGS6812, TGS2610,
TGS2612 and TGS2611) give little or no response. It can be
noted that the TGS826 followed by TGS822 and TGS825
give strong response while the other 5 sensors give weaker
but yet still significant response. To observe the pattern
associated with each cultured milk drink brand, the degree
of reactions obtained for the four brands were plotted inradar form presented in Fig. 3. The radar plot shows that the
four brands have quite similar patterns with different
magnitude for each sensor. The Vitagen and Solivite brands
emanate strongest odor, followed by Nutrigen and Yakult
brands. The 6 sensors which show little or no response to
the cultured milk odor were excluded from the next
analyses.
Figure 2. The degree of reaction of the 14 sensors.
TABLE 2. THE RSD VALUES OF THE 8 SENSORS
BrandFeatures
d r rt i
Vitagen 8.8 11.8 109.7 9.1
Solivite 5.5 6.7 20.3 10.7
Nutrigen 11.8 16.2 28.7 16.7
Average 8.7 11.7 52.9 12.2
B. The repeatability of features
The analysis to determine which features show the best
repeatability was conducted on three cultured milk brands ie
Solivite, Vitagen and Nutrigen. In this analysis, 10ml of
milk were used and 3 consecutive measurements were
performed using the same sample obtained from the three
milk brands. The RSD of the 8 sensors and their average
were obtained for each feature. The results were presented
in Table 2. The RSD values indicate that the degree of
reaction was the most repeating feature among the four
features analyzed. This can be observed by the small RSD
score by this feature for the three milk brands which are 8.8,
5.5 and 11.8 for Vitagen, Solivite and Nutrigen,
respectively, with the average of 8.7. The relative response
and the accumulative of total reaction changing show good
repeatability with the average value of 11.7 and 12.2,
respectively. Among the four features, the rate of reaction
measured during the first 50 seconds of absorption show
worst repeatability with the average value obtained for the
three cultured milk drink brands is 52.9.
C. The discrimination of features
The PCA is used to determine the ability of the fourfeatures to discriminate between 4 different brands of
cultured milk. In this analysis, the PCA was performed on
the whole data set and the score values obtained were
plotted in Fig. 4 for degree of reaction, Fig. 5 for relative
response, Fig. 6 for rate of reaction at initial absorption time
and Fig. 7 for accumulative total of the reaction degree
changing. These score plots show that the four brands are
obviously separated by the degree of reaction feature and
the relative response feature while adequately separated by
the accumulative total of the reaction degree changing
feature. The score plot for the rate of reaction at initial
absorption time indicates that this feature is unable to
discriminate the four brands effectively.
Figure 3. The degree of reaction for the 4 cultured milk drink brands.
1187
-
8/9/2019 enose iki tentang
4/4
-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
Component 1 (92.7%)
C o m p o n e n t 2 ( 6 . 1
% )
Figure 4. The PCA plot using the degree of reaction.
-100 0 100 200 300 400 500-8
-6
-4
-2
0
2
4
6
Component 1 (99.9%)
C o m p o n e n t 2 ( 0 . 1
% )
Figure 6. The PCA plot using the rate of reaction at the initial absorption
time.
IV. CONCLUSION
Various features can be extracted from a sensor response
curve obtained in the electronic nose measurement. The
quality of these features in terms of their repeatability and
discrimination ability to provide a standard reading must bestudied before being used in any classification or prediction
problems. This paper presented four different features
obtained from a sensor response curve. The results showed
that the degree of reaction is the best feature to represent theodor of cultured milk drink. This feature is the most
repeatable and the best feature to discriminate the four
brands of cultured milk drink. Besides that, the results show
that the sensors exhibit different reaction rate at the initial
absorption time resulting to the instable reading of the
reaction at the initial point of absorption.
REFERENCES
[1] S. Panigrahi, S. Balasubramaniam, H. Gu, C. M. Logue, and M.Marchello, “Design and development of a metal oxide basedelectronic nose for spoilage classification of beef,” Sensors andActuators B, vol. 119, pp. 2-14, June 2006.
[2] A. H. Gomez, J. Wang, G. Hu and A. G. Pereira, “Monitoring storageshelf life of tomato using electronic nose technique,” Journal of FoodEngineering, vol. 85, pp. 625-631, September 2007.
-1.5 -1 -0.5 0 0.5 1 1.5-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
Component 1 (89.6%)
C o m p o n e n t 2 ( 8 . 6
3 % )
Figure 5. The PCA plot using the relative reaction.
-400 -300 -200 -100 0 100 200 300-300
-250
-200
-150
-100
-50
0
50
100
150
Component 1 (86.6%)
C o m p o n e n t 2 ( 1 1 . 9
% )
Figure 7. The PCA plot using the accumulative total of the reaction degreechanging.
[3] A. H. Gomez, J. Wang, G. Hu and A. G. Pereira, “Discrimination ofsorage shelf life for mandarin by electronic nose technique,” LWT,vol. 40, pp 681-689, March 2006.
[4] S. Labreche, S. Bazzo, S. Cade, and E. Chanie, “Shelf lifedetermination by electronic nose: application to milk,” Sensors andActuators B, vol. 106, pp. 199-206, March 2005.
[5] K. Brudzewski, S. Osowski, and T. Markiewicz, “ Classification ofmilk by means of an electronic nose and SVM neural network,”Sensors and Actuators B, vol. 98, pp. 291-298, 2004.
[6] N. El Barbri, E. Llobet, N. El Bari, X. Correig, and B. Bouchiki,Electronic Nose based on metal oxide semiconductor sensors as analternative for the spoilage clasification of red meat,” Sensors, vol. 8,pp. 142-156, January 2008.
[7] A. Amari, N. El Barbri, E. Llobet, N. El Bari, X. Correig, and B.Bouchiki, “Monitoring the freshness of moroccon sardines with aneural network based electronic nose,” Sensors, vol. 6, pp. 1209-1223, October 2006.
[8] M. Peris, and L. E. Gilabert, “A 21 st century technique for foodcontrol: electronic noses. Analytica Chimica Acta,” vol. 638, pp. 1-15, 2009.
[9] K. Brudzewski, and J. Ulaczyk, “An effective method for analysis ofdynamic electronic nose responses,” Sensors and Actuators B:Chemical, vol. 140, pp. 43-50, 2009.
[10] S. Zhang, C. Xie, D. Zeng, Q. Zhang, H. Li, and Z. Bi, “A featureextraction method and a sampling system for fast recognition offlammable liquids with a portable E-nose,” Sensors and Actuators B,vol. 124, pp. 437-443, 2007.
1188