CC BY
0O PEN ACCESS
SSnSDlS
ISSN 1424-8220 www m dpicom /jpuincLl/sen^gDis
Article
Classification of A garw ood O iL Using an EHectronic Noss
W ahyu Hidayat *,ALiYson M d. Shakaff M ohd NoorAhmad and AbdulHamid Adom
Sen^sorTechnoLogy and ApplicationsResearch Cluster, Universiti. M alaysia Perls UnM AP), 01000 Kangar, Peris, M alaysia; E-M ails: aliyeon@ uninapedumy A YM S); mohdnoors uninapedumy M NA .); abdham id@ uninapedumy AHA.)
* Autthortowhom correspondenceshould be addre^ed; E-M ail: wahyuh@ hotmailcom ; Tel.: +60-4-9798897; Fax: +60-4-9775302.
Received: 1 March 2010; in revised form : 14 April 2010 /Accepted: 19 April 2010 / Published: 6 M ay 2010
Abstract: Presently, the quaüy assistance of agarw ood oil is perform ed by sensory panels which has sdgnificantdraw backs in tern s of objectivity and repeatability, in this paper, it is show n how an electronic nose (e-nose) may be successfully utilised for the classification of agarwood oil. H ierarchical Cluster Analysis HCA) and Principal Component Analysis PC A ), were used to classify different types of oil. The H CA produced a dendrogram show ing the separation of e-nose data into three different groups of oils, The PCA scatter plot revealed a distinct separation between the three groups, An Artificial Neural Network ANN) was used fora betterprediction of unknow n sam pies,
K eyw ords: agarw ood oil; e-nose; HCA; PCA ; ANN; din ensionality reduction
1. Introduction
Agarwood is the well-known name for a resinous heartwood from 'wounded/infected' Aquilaria trees, a tropical ioresfc product which has a high value ininternational trading. There are increasing dem ands foragarw ood-based products for use in m edicine, perfum e, and incense. Agarw ood is traded in the form of product derivatives such as wood chips, powder, and oil. The wholesale price for high quality agarwood oils is around US$30,000-US$50,000 per liter [1], depending on the oil quality, which is based upon the fragrance strength and longevity, resin content, geographical origin, and oil purity [2].
Traditionally, agarwood grading has been perform ed by trained hum an graders (sensory panels). However, the m ethod has disadvantages in term s of objectivity and nepmtability [3] .Inadditbn, a hum an nose cannot tolerate a high num ber of sam pies because it fatigues rapidly w ith increasing num ber of sam pies. in this paper, it is show n how an electronic nose (e-nose) may be used to resolve these iteues. A commercial e-nose, the Cyranose 320 Smith Detection, USA) was used tto collect the OTell (fragrance) data (herein termed as 'otellprint) which was then processed on a personal computer using different- pattern recognition metthods: Hierarchical Cluster Analysis (HCA), Principal ComponentAnalysis (PCA) and Artificial Ne iral Network ANN).
2. Experim ental
An experiment aimed to produce a data set with robust consistency was conducted using a Cyranose 320 (see Figure 1). The important features of the Cyranose 320 are provided in Table 1 [4]. The acquired raw data was then proce^ed and interpreted into meaningful informaton. The sampies were obtained from three groups of agarwood oils orginating from Laos, Johor M alaysia) and Terengganu M alaysia),w hich were labelled as G12, G22, and G32, respectively. Two ^L sample from each locality was dilated using 500 mL glycerol as a solvent. The samples were placed in a 40 mL stoppered vial connected to a Cyranose 320 via an inert tubing and heated to 60 °C in a heater block:. Each experimentwas run for30 m in.
Figure 1. The Cyranose 320 [4].
Table 1. Important features of the Cyranose 320 [4].
Sensors 32 polym ercarbon black com posies
Operating Tem perature 0 to 40 °C D2 to 104 °F)
Response Time 10 sec
Sampling Pump Low : 50 mL/n in, M edium : 120 mL/n in,
High: 180 mL/n in.
Communication RS-232 @ 9,600 to 57,600 bps
Algorithm s PCA , KNN , K-means, CDA
Figure 2 illustrates an airtight recycle system for continuous shifting. This approach couldavoid volatile concentration loss and also pressure loss at the glass vial which w ould affect the experim entt An automatic valve as well as a small pump was provided in the Cyrnose 320 to control the system flow . I addition, the use of an extra valve m echanism outside the Cyanose 320 w as im provissd to facilitate sensor cleansing. In Figure 2, the red arrows (axlid-line) indicate a sampling cycle. An internal valve was ^itched to the sample inlet X). The volatile was then sucked into the e-nose through inlet X) and was retained for 20 s in the sensor chamber before being removed through outlet (Z). At the sametme, the externalvalve allowed volatile from Z to fill the glass vial. Atthe end of 10 cycles, the internal valve allowed the nitrogen gas (N2) to purge the sensors via inlet Y) and rem oved out to the atm osphere through port (Z). The purge cycle is illustrated by the blue arrow s (dashed-line) in Figure 2. The experim ents were controlled by the Cyranose 320 according to the œt-up param eters as show n in Table 2.
Figure 2. Experim ental œtup for the clarification of agarw ood oil.
Computer
Table 2. Cyrnose 320 param eter setup for sam pling agarw ood oil.
Run tim e Pum p speed
Baseline purge tim e 10 sec 120 mL/n in
Sam pling tim e Draw 1 20 sec 180 mL/n in
Purge tim e 1st air intake purge 2nd sam pie gas purge 5 sec 30 sec 180 mL/n in 180 mL/n in
Digital filtering On
Substrate heatertem perature 42 °C
Training repeat count 10
3. Results and Discussion
31. Sm ellprint
The agarwood oiil volatiles are adsorbed on the sensor's surfaces and cause a change in its resistance. The response of the sensor is defined by using fractional baseline m anipulaton [5] :
where ÀRs is the resdstance change of sensor s;, Rsn is the output resistance and RS/0 is the baseline output. The subscript index s is the sensor num ber used in the Cyranose (s = 1.. .32) and n is an index forthe numberof data (n = 1...N).
As an example, Figure 3 show s responses from seven sensors of the Cyanose 320. The data was taken from one sam pling cycle of a G12 experim entt The figure also illistates the base line purge tm e, sam pling tim e and purge tim e.
Figure 3. M easurm entstaken from seven of the sensors fforone sampling cycle.
The average of values evaluated by Equation 1) is plotted as show n in Figure 4, and corresponds to the sm elprhts of the three different agarwood oils. Sensors w ith high responses are analyzed by comparing their peaks and profiles [6]. Sensor numbers 6, 31, 5, 23, and 28 (in the order of dim inishing responses) have higher responses com pared to the restwhen exposed to the volatles of the different grade of oils. However, the analysis of ot ellprints becom es m ore difficult when there is an increase in the number of samples having overlapping profiles. This issue can be solved using graphical methods based on statistical theories [7], and this was adopted and presented in the next section.
Figure 4. Sm
3 2. StatisticalAna]ysds
There are m any statistical-based m ethods for processing e-nose data. This paper presents the implam entation of the Hjerarchical Cluster Analysis (HCA) and Principal ComponentAnalysis PCA) to
3 21. Hierarchical ClusterAnalysis HCA)
The aim of perform ig Hierarchical cluster analysiis HCA) is to separate data inte» specific groups by considering sim ilarity crtEron, a distance m etric such as Euclidean distance, as folbw s:
di =(Z (r - r:k
where K is tthe number of variables (in this case K is equal to 32 tthatis the number of sensors in the Cyranose), while i and j are the indices forgroups of sam p]es. H ence, a param eterto m easure the level of sim ilarity, Sij, is defiined as [8]:
Sij =1" d^Anax^} 3)
The com putational process of S±j using M A TLA B gives a dendrogram as show n in Figure 5. The figure proved the capabilty of HCA to diffprentiate betw een G12, G 22, and G 33.
Figure 5. A dendrogram for the three-object data œtfrom each 10 samples of G12, G 22, and G 32.
3 2 2. Principal C om ponent Analysis (PCA )
Principal com ponentanalysis PCA ) is an unsuperviœd statistical m etthod that generates a new set of variables, called principal com ponents. Each principal com ponent is a linear com binationof the original-variables U defined by [9] :
PC pn = «p^n + «p^n + • " + «p,sfs,n (4)
where PCpn is the notation forthe p-th orderprincipal com ponent forthe overall n num ber of data and is tern edas scores. Coefficients transform ations (ap/s), referred as loadings, are obtainedby taking elem ents of the eigenvectors from the covariant of the original data. The eigenvalue represents the variance associated w ith each principal com ponent. By using M ATLAB software, the two principal components {PC1/n, PC2n} are obtainedandhave the two greatest variances: 88 096% and 11202% ortotal cum ulative variance of 99 298% ). The results of the PCA analysis are shown in Figure 6. The œores of the three groups of oils are plotted for principal component 2 (PC2) verais principal com ponent 1 (PC1). The diœrim inationbetw een the different types of oils can be clearlyseen from the figure.
Figure 6. Principal com ponents score plot proves the capabilityof e-nose to classifythe
\ 12, G22, and G32.
33. ArtificialNeural Network ANN)
3 31. Result from 32 Sensors as Input
The previous two statistical approaches, HCA and PC A , successfully showed their capabilities to distinguish different types of agarw ood oils. Both are typically used for exploratory data analysis to see how the m ulivariate data is clustered and to arer the linear separability of the odour clasps. However, in cases where prediction is required (eg., when implem enting an auttom ated clarifier), the ANN is the m ore appropriate tool [7]. This section presents the use of ANN as an alternative choice to sxlve the classification problem in this work.
In thisexperiment,the backpropagation ANN with the Levenberg-M arquardt training algorithm was applied. The training used all the 32 sensors as inputs, 20 neurons in the single hidden layer, and three neurons at the output layer. The activatinfunctions used are sigm oidandidentiyfunctions at the hidden and output layer, respectively. Figure 7 illustrates the structure of the ANN and how the input and outputdata were organted and indexed..
Figure 7. The architecture of three layers ANN with Levenberg-M arquardt algorithm applied fcrtraining and identification G12, G 22, and G 32.
600 401 200 1
I G32 G22 G12 , I-■ I-1 I-1 I-1
Data for training Input Layer Hidden Layer Ouput Layer Data for Target
The same experimental procedure was carried out for the trining as well as the testing data.. The 200 raw data points were collected by experim ents for each of the oil types G12, G 22, and G 32. This resulted in 600 data points for use in training and validation. After performing baseline manpulatbnand auto-scaling, the data were organted inone m atrix to be fed as input The testing data was collected on a different day, by sniffing the odour of the oils in nine vials that were aligned fortesting only. The totalof 1800 data points was used as the testing data.
As shown in Table 3, the ANN performed very well in discriminating the three types of oils with 5.713345 x 10 mean square error M SE) and 100% prediction performance. The prediction performance was defined as: (numberof corrctclareifration/h.um berof totaldata) x 100% .
Table 3. ANN output for32 sensors.
Sensor Target ANN Output Averaged)
No. M SE Gradient Sample--Accuracy
selection T1 T2 T3 O1 O2 O3
1 All 5.7133 x 10"8 0 3536 G12 1 0 0 0 9999 0 0001 0 0000 100%
G 22 0 1 0 0.0003 0 9997 0 0000 100% G 32 0 0 1 0 0005 0 0002 0 9997 100%
332. Results from Selected Sensors
The use of too m any sensors m ay increase noise, redundant inform atton and provides no real benefit whereas minimizing the number of sensors can result in the loss of som.e useful input inform aton [10,11]. Thus, optm izatonfor a specific application can be achieved by observing the sensors which provide high contribution to the ^stem and elim inating the low er ones.
in this w ork, PCA is used to reduce high din ensbnally data andto im prove ANN training [12]. Table 4 is the list of loadings for PC 1 that w as sorted in descending order. The sum m atonof the correlation coefficient from the m atrix response data for each sensor is also provided in Table 4 as a com parisonw iththe PCA results. H is evident that the higher loading values of PC 1 for all sensors correspond to the ler correlated msors.
From Table 4, the data from the five least correlated sensors (sensornum ber23, 31, 1, 2, and 4) are elected as input forthe ANN training. The result shows an improvementwhere the ANN has a lower M SE, 8 20279 x 10 9 and 100% succesflilprediction of unknown data.. The tim e gpentfortraining and ientification also decrease. For specific application, in this case to disarm inate the three different types of agarwood oils G12, G22, and G32, the five elected sensors (from the 32 total sensors in the Cyranose) is an effective choice in terms of ^peedof detection and high accuracy. Attempts to further reduce the numberof sensorswas not suaaerfuL Table 5 compares the resulsof the ANN training for the case of using reduced num berof sensors.
Table 4. The value of total correlation coefficientand loadings of PC 1 for seven sensors.
No Sensor num ber The sum m ation of correlation coefficient Sensor num ber Loadings for PC 1
1 23 5.4960 23 -0 03041
2 31 16.016 31 -0 09815
3 1 19.334 1 -012595
4 2 24.291 2 -015748
5 4 25.343 4 -016391
6 9 25.714 9 -016607
7 5 26 599 5 -017615
Table 5. The comparison performance ofANN using selected sensors.
No. Sensor M SE G radient Sam ple ■ Target ANN Output Averaged) ■A ccuracy
election T1 T2 T3 01 02 03
1 S23,S31, 8 2038 x 10"9 0.1074 G12 1 0 0 0 9983 0.0045 0 0070 100%
S1,S2, G 22 0 1 0 0 0001 0.9999 0.0070 100%
S4 G 32 0 0 1 0 0001 0.0000 0 9999 100%
2 S23 and 8 9927 x 10"8 0 00246 G12 1 0 0 0 9803 0 0125 0 0070 100%
S31 G 22 0 1 0 0 0010 0.9999 0.0009 100%
G 32 0 0 1 0 6203 0.0000 0.3797 37.18%
4. Conclusions
Classificationof agarwoodoils using an e-nose is able to provide rapid and accurate results. The
data from the Cyranose 320 were proces^d using in-house developed aftware in M ATLAB and was
able to identify three different. types of agarw ood oils G12, G 22, and G 32. Hierarchical cluster analysis
HCA) and principal component analysis PCA) were suaaerful in separating the samples into
differentgroups or clusters. ANN was also successfully applied to predictunknow n agarw ood sam p]es. The optim um num ber of sensors for this application has been determ ined by PC A analysis, w hich subsequently m inim ize the num ber of ANN input variables. The current research in our laboratories is to verify the purity and grading of the oil based on quantitative analysis using the e-nose.
A cknow ledgem ents
The authors gratefully acknow ledge the financial support from the M inistry of Science, Technology and Environm ent, M alaysia through Grant No. 9005 -00007. A1d, to Universiti M alaysia Perls for the financiala^istance given to W ahyu Hidayat. The authors w ish to thankNor A zah M A from Forest Research institute M alaysia, M R Awang from M alaysian Institute for Nuclear Technology Research and M R Kamarudin for their collaboration., supply of samples, assistance and useful discu^ions.
References
1. Persoon., G A . Growing 'the wood of the Gods': Agarw ood production in Southeast Asia.. in Smallholder Tree Growing for Rural 245 Development and Environmental Services; Snelder, D J., Lasco, RD. Eds.; Springer: Dordrecht, The Netherlands, 2008; pp. 245-262.
2. Barden, A.; Anak, N A .; M ulUken., T.; Song, M . Hearth of the Matter: Agarwood Use and Trade and CITES Implementation for Aquilaria M alaccensis; Traffic Network Report 1998-2000; TRA FFÜC international: Cam bridge, UK, 2000; pp. 1-51.
3. Keller, PE. M im icking biology: applications of cognitive ^stem s to electronic noses. mProoeedings of 3EEE International Symposium on Tnt-elligenr Conlrnyrht-elligenr System s and Semiotics, Cambridge, M A ,USA ,Septem ber1999.
4. Cyranose 320 User M anual; Sm ith Detection: Pasadena, CA, USA, 2004.
5. Gardner, JW ; Bartl^tt, P N. Electronic Noses: Principles andApplicationis; OxiordUniiversity Press: New York, NY, USA, 1999.
6. Luo, D ; Ho^eini, H G ; Stuart, JR. ApplicationLof ANN w ith extracted param eters from an electronic nose in cigarette brand üentificationL. Sens. Actuat. B: Chem . 2004, 99, 253-257.
7. Scott, S M ; Jam es, D ; A li.Z . Review : D ata analysis ibr electronic no^ ^stem s. M icrochim . Acta. 2007, 156, 183-207.
8. H ines, E L ; lobet. EL; Gardner, JW . Electronic noses: a review of signal proce^ing techniques. in Proceedings of IEEE Circuits Device System , Sale Lake City, UT, USA , December 1999; Volume 146, No. 6.
9. Jolliffe, IT . Principal Component Analysis, 2nd ed; Spriger-Verlag: New York, NY , USA , 2002.
10. Penza, M ; Ca^ano, G. Application ofprincioal com ponentanalysis and artificial neural networks to recognize the individual VOCs of methahol2-propahol in a binary mixture by SAW muHti-sensorarray.Sens. Actuat. B: Chem . 2003, 89, 269-284.
11. Gaidnei, JW .; BoiHot, P.; Hines, E L. Enhancing electiDn;ic nose pei-fbim ance by sensd selection using a new integei-based genetic algorithm appioach. Sens. Actuat. B : Chem . 200S, 106, 114-121.
12. Dutta, R .,• Das, A . Stockas, N G . M oigan, D . Stochastic resonance-based eLectronic nose: A novel way to classify bacteria. Sens. Actuat. B: Chem . 2006, 11S, 17-27.
© 2010 by the authois; licensee M DPI, Basel, Switzeiand. This article is an open-access aiticle distributed undei the terms and conditions of the Creative Commons Attribiitin license (http ://ciœtiveGbm m onsbig/llicenLses/by/ 0/).
Copyright of Sensors (14248220) is the property of MDPI Publishing and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.
