Citation
Predicting spice mixture composition

Material Information

Title:
Predicting spice mixture composition comparing electronic nose, gas chromatography, and sensory methods
Creator:
Zhang, Haoxian
Publication Date:
Language:
English
Physical Description:
xiii, 124 leaves : ill. ; 29 cm.

Subjects

Subjects / Keywords:
Basil ( jstor )
Chemicals ( jstor )
Cinnamon ( jstor )
Electronics ( jstor )
Garlic ( jstor )
Neural networks ( jstor )
Nose ( jstor )
Odors ( jstor )
Sensors ( jstor )
Spices ( jstor )
Agricultural and Biological Engineering thesis, Ph.D ( lcsh )
Dissertations, Academic -- Agricultural and Biological Engineering -- UF ( lcsh )
Genre:
bibliography ( marcgt )
theses ( marcgt )
non-fiction ( marcgt )

Notes

Thesis:
Thesis (Ph.D.)--University of Florida, 2003.
Bibliography:
Includes bibliographical references.
General Note:
Printout.
General Note:
Vita.
Statement of Responsibility:
by Haoxian Zhang.

Record Information

Source Institution:
University of Florida
Rights Management:
The University of Florida George A. Smathers Libraries respect the intellectual property rights of others and do not claim any copyright interest in this item. This item may be protected by copyright but is made available here under a claim of fair use (17 U.S.C. §107) for non-profit research and educational purposes. Users of this work have responsibility for determining copyright status prior to reusing, publishing or reproducing this item for purposes other than what is allowed by fair use or other copyright exemptions. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder. The Smathers Libraries would like to learn more about this item and invite individuals or organizations to contact the RDS coordinator (ufdissertations@uflib.ufl.edu) with any additional information they can provide.
Resource Identifier:
030457542 ( ALEPH )
81464934 ( OCLC )

Downloads

This item has the following downloads:


Full Text











PREDICTING SPICE MIXTURE COMPOSITION: COMPARING ELECTRONIC
NOSE, GAS CHROMATOGRAPHY, AND SENSORY METHODS














By

HAOXIAN ZHANG


A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY

UNIVERSITY OF FLORIDA


2003































Copyright 2003 by

HAOXIAN ZHANG































To my dearly loved husband, Hailong, with whom enjoyable and difficult times were shared throughout this journey, and to my parents, Chuanmao and Jiyun, for their love and support throughout my life.














ACKNOWLEDGMENTS

I would like to express my sincere thanks and appreciation to my advisor, Dr. Murat 0. Balaban, for his invaluable advice, encouragement, support and guidance throughout my graduate studies at the University of Florida and for giving me the opportunity to study the interesting subject of this research.

I would also like to thank my committee members, Drs. Kenneth M. Portier, Jose C. Principe, Charles A. Sims and Arthur A. Teixeira, for their help, suggestions, and words for encouraging this research. My accomplishments could not have been achieved without their support. Special thanks go to Dr. Portier and Dr. Principe for their time and patience to teach me in the topics of multivariate statistics and artificial neural networks, respectively.

Finally, I am thankful to all my friends at the University of Florida, especially those at Dr. Balaban and Dr. Sims' lab, for their helpful suggestions, discussions and friendship. Also my appreciation goes to the panelists who helped in the sensory studies.


iv
















TABLE OF CONTENTS
Page

ACKN OW LED GM EN TS ............................................................................................. iv

LIST OF TABLES........................................................................................................... viii

LIST OF FIGURES ........................................................................................................ x

ABSTRA CT...................................................................................................................... xii

CHAPTER

I IN TRODU CTION .................................................................................................... 1

2 LITERATURE REV IEW ......................................................................................... 5

Spices............................................................................................................................5
Electronic N ose...................................................................................................... 6
Introduction ...................................................................................................... 6
Applications in Food Area................................................................................ 7
Equipm ent....................................................................................................... 8
Data Analysis M ethods ........................................................................................11
Gas Chrom atography ............................................................................................. 13
Introduction .................................................................................................... 13
Equipm ent....................................................................................................... 13
Sam pling M ethods of V olatiles....................................................................... 16
Application to Spice M ixture Analysis ........................................................... 20
Sensory Evaluation ................................................................................................ 21
Introduction .................................................................................................... 21
M easurem ent of Sensory Thresholds ............................................................. 22
Discrim ination Tests....................................................................................... 24
Difference Threshold - Psychophysical Theory .............................................26
Difference Thresholds of M ixtures ............................................................... 26
N eural N etworks.................................................................................................... 27
Introduction .................................................................................................... 27
Multilayer Perceptrons (MLPs) Trained by Back-Propagation (BP)..............28
Tim e-Delay N eural N etwork (TDNN) ........................................................... 30
Sofinax................................................................................................................33


v









3 O BJECTIV E ......................................................................- .. ....................34

Electronic Nose Analysis .....................................34
G C A nalysis.............................................................................. ...... ...........34
Sensory A nalysis ............................................................................................34

4 M A TER IA LS A N D M ETH O D S............................................................................ 35

M aterials .....................................................................................................................35
Experim ental D esigns and M ethods....................................................................... 36
Mixture Experimental Design of Three Components .................................... 36
Electronic N ose ............................................................................................... 38
G as Chrom atography....................................................................................... 41
Sensory Thresholds .................................................................................... .. 47
D ata A nalysis..............................................................................................................51
Electronic N ose ............................................................................................... 51
M LP......................................................................................................... 51
PCA -M LP................................................................................................ 52
TDNN ....................................................................................................... 53
G as Chrom atography ....................................................................................... 53
Single volatile m ethod.............................................................................. 54
Five-volatiles m ethod.............................................................................. 55
Sensory Thresholds ......................................................................................... 56


5 RESU LT AN D D ISSCU SION ................................................................................ 58

Electronic N ose....................................................................................................... 58
Raw D ata ......................................................................................................... 58
D ata A nalysis Results U sing M LP .................................................................. 59
D ata A nalysis Results U sing PCA -M LP ......................................................... 60
D ata A nalysis Results U sing TDN N .............................................................. 61
D iscussion....................................................................................................... 63
G as Chrom atography .............................................................................................. 67
V olatile Com ponents of Spices ....................................................................... 67
Single V olatile M ethod................................................................................... 72
Five-volatiles M ethod..................................................................................... 72
D iscussion....................................................................................................... 73
Sensory Thresholds................................................................................................ 79
Comparison between E-nose and GC/Sensory Methods....................................... 81











vi









6 SUMMARY, CONCLUSIONS AND SUGGESTIONS FOR FUTURE STUDY ....85 APPENDIX

A INSTANTANEOUS DATA FOR E-NOSE ANALYSIS .................................... 88

B TIME SERIES DATA FOR E-NOSE ANALYSIS............................................. 96

C RAW DATA FOR GC ANALYSIS ..........................................................................98

D RAW DATA FROM SENSORY ANALYSIS........................................................103

LIST O F REFEREN CES.................................................................................................115

BIOGRAPHICAL SKETCH ...........................................................................................124


vii















LIST OF TABLES


Table page

2-1. Classification of test methods in sensory evaluation...........................................21

4-1. The chemical names and their corresponding CAS# and KI (for DB-5) of major
volatile compounds of basil, cinnamon and garlic ............................................. 36

4-2. The weight fractions of design points for training/cross-validation....................40

4-3. The weight fractions of spices for testing........................................................... 40

4-4. The changes of the percentage of extraction of the total volatiles from the spice
mixtures with equal proportions of basil, cinnamon and garlic versus distillation
tim e using SD E .................................................................................................... 44

4-5. The retention times of n-paraffin hydrocarbons (C7-C20)..................................47

4-6. The mixing fractions of the six pairs of spice mixtures compared by panelists.......50 5-1. The experimental and predicted mass fractions of spice mixtures predicted by the
optimal performance of MLP............................................................................. 59

5-2. The proportion and cumulative proportion of total variance explained by principal
com ponent I through 10...................................................................................... 61

5-3. The experimental and predicted mass fractions of spice mixtures predicted by the
optimal performance of PCA-MLP.................................................................... 62

5-4. The experimental and predicted mass fractions of spice mixtures predicted by the
optimal performance of TDNN...........................................................................63

5-5. The average relative amounts of all volatiles identified in each of the three spices
and their corresponding retention times............................................................. 68

5-6. The five most abundant unique volatiles identified in the basil, cinnamon and
garlic, and the volatiles' corresponding chemical identities ...............................70

5-7. The average fraction (in percentage) of the volatiles in their corresponding pure
spice extracts. ...........................................................................................................71


viii









5-9. The K values of the most abundant garlic volatile at different weights of garlic in
the corresponding spice mixture ........................................................................ 76

5-10. The K values of the five most abundant unique basil volatiles at different weights of
basil in the corresponding spice mixture.............................................................76

5-11. The K values of the five most abundant unique cinnamon volatiles at different
weights of cinnamon in the corresponding spice mixture....................................78

5-12. The K values of the five most abundant unique garlic volatiles at different weights
of garlic in the corresponding spice mixture.......................................................78

5-13. Comparison between the prediction error of electronic nose methods and the
sensory threshold of human subjects.................................................................. 82

5-14. Comparison between the prediction performance of electronic nose and gas
chrom atography m ethod....................................................................................... 84


xi
















LIST OF FIGURES


Figure page

2-1. Cross section of a Fused Silica Open Tubular Column....................................... 15

2-2. Multilayer Perceptrons (MLPs) with one hidden-layer....................................... 30

2-3. Tapped delay line with p delay units....................................................................31

2-4. Paradigm of Time-delay Neural Network with one hidden-layer ........................32

4-1. The simplex factor space with three components................................................ 37

4-2. The three basic design methods in the simplex space with three components.........38

4-3. The distribution of design points in the simplex space ........................................39

4-4. The apparatus used for simultaneous distillation extraction. .............................. 43

4-5. The comparison points for sensory threshold testing ...........................................49

4-6. The single volatile method for GC data analysis..................................................55

4-7. The five-volatile method for GC data analysis....................................................56

5-1. A sample of time series data obtained..................................................................58

5-2. The optimal topology of MLP............................................................................. 60

5-3. The optimal topology of PCA-MLP.................................................................... 63

5-4. The optimal topology of TDNN ...........................................................................65

5-5. The distribution of the prediction points, including those from e-nose MLP, MLPPCA and TDNN analysis, in the simplex space.................................................. 66

5-6. The distribution of the prediction points, including those from GC single volatile
and five-volatile methods, in the simplex space .................................................74

5-7. The K values of the most abundant basil volatile at different weights of basil in the
corresponding spice mixture ............................................................................... 75


x









5-8. The experimental and predicted mass fractions of spice mixtures predicted by the
single volatile m ethod ........................................................................................ 72

5-9. The experimental and predicted mass fractions of spice mixtures predicted by the
five-volatile m ethod ............................................................................................. 72

5-10. The results of the triangle tests for each of the six pairs of comparisons............79

5-11. The number of correct responses necessary for the triangular tests to establish a
significant difference between the two samples under comparison when the total
num ber of the tests is 50....................................................................................... 80

5-12. The comparison of panelists' performance between the tests carried first in that day
and those carried second ...................................................................................... 81

5-13. Comparing the accuracy and the efficiency of GC and electronic nose methods ....83 A-1. The instantaneous data obtained from e-nose experiments for training ...............89

A-2. The instantaneous data obtained from e-nose experiments for testing.................95

C-1. The raw GC data from each of the three pure spice extracts - first injection.....99 C-2. The raw GC data from each of the three pure spice extracts - second injection.....100 C-3. The raw GC data from the extracts from each of the five spice mixtures - first
injection ..................................................................................................................10 1

C-4. The raw GC data from the extracts from each of the five spice mixtures - second
injection ..................................................................................................................102

D-1. The triangle test results from the comparisons of the mixtures IT and 1T-1.........104

D-2. The triangle test results from the comparisons of the mixtures 3T and 3T-1.........106

D-3. The triangle test results from the comparisons of the mixtures 4T and 4T-1.........108

D-4. The triangle test results from the comparisons of the mixtures 4T and 4T-2.........110

D-5. The triangle test results from the comparisons of the mixtures 4T and 4T-3.........112

D-6. The triangle test results from the comparisons of the mixtures 5T and 5T-I.........114


ix














Abstract of Dissertation Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy

PREDICTING SPICE MIXTURE COMPOSITION: COMPARING ELECTRONIC
NOSE, GAS CHROMATOGRAPHY, AND SENSORY METHODS By

Haoxian Zhang

August 2003

Chair: Murat 0. Balaban
Major Department: Agricultural and Biological Engineering

Electronic noses (e-nose) are instruments that can quickly detect odors at low cost, and their potential applications are very diverse. Limited work has been done in investigating e-noses' quantitative ability and no work has been reported on the application of e-noses to predict mixture compositions.

The objective of this project was to develop a quantitative procedure that could quickly predict the compositions of a ternary spice mixture by using an e-nose for measurement and multivariate statistics/neural networks (NN) for data analysis. The relative accuracy and efficiency of the developed e-nose methods were determined by comparing them to those resulting from gas chromatography (GC) and sensory methods.

Three ground spices (basil, cinnamon and garlic) were mixed in different

compositions and presented to an e-nose. Nineteen training blends were used to build a predictive model, the performance of which was tested by five other blends. Three NN structures were used for predictive model building (multilayer perceptron (MLP), MLP


xii








using principal components analysis as preprocess and time-delay NN). For GC analysis, the volatile components of spice mixtures were collected by simultaneous distillationextraction and quantified by GC. Mixture compositions were predicted based on the amounts of unique volatiles of each spice. The testing blends applied in GC analysis were the same as those in e-nose experiments. For sensory analysis, triangle tests were performed by 50 panelists to estimate the difference thresholds of spice mixtures.

The best NN model built from e-nose data predicted the compositions of testing spice mixtures with an error less than 0.06. A difference of 0.06 between two spice mixtures was determined through sensory analysis to be lower than human sensory thresholds. The GC method provided a more accurate but less efficient prediction. Its experimental time required for each unknown sample was 8 hours, instead of 50 minutes for the e-nose method.

The procedure developed in this study can predict the compositions of a ternary spice mixture with an acceptable accuracy and significantly improved efficiency. The procedure will be valuable in quality monitoring or process control, in which efficiency is essential.


xiii













CHAPTER 1
INTRODUCTION

Electronic noses (e-noses) are instruments developed in the last 10 years to mimic the sense of smell. Consisting of olfactory sensors and multivariate signal analysis methods, they are able to quickly detect and distinguish odors at low cost. The potential applications of electronic noses are very diverse, ranging from military applications, clinical diagnosis, and environmental monitoring, to applications in food, flavor and biotechnology industries. Currently, the biggest use of the electronic nose is in the food and fragrance related areas.

Until now, electronic nose studies were focused on investigating qualitative aspects of samples, e.g., detecting or classifying odor patterns, such as "fresh" versus "spoiled." More work in exploiting their quantitative ability can extend the horizon of e-nose applications. No work was found in applying e-noses to predict mixture compositions. In the food and fragrance industries, many products are formed by mixing two or more ingredients with different olfactory attributes. A simple, fast mixture composition prediction tool for such products would be valuable for routine quality or process control, as well as product matching or re-formulation.

There are additional constraints in mixture experiments: the sum of all fractions must add up to one, and the fraction of each component must be non-negative. These bring specific considerations for the optimal experimental design. One of the objectives of this study was to develop an experimental procedure that can quantify mixture compositions using an electronic nose as the measurement tool.


I






2


The signal analysis (data analysis) methods for e-noses were designed for pattern classification or recognition. Even in the few studies dealing with quantitative attributes of samples (Alpha M.O.S., 1997; Dittmann et aL., 2000), the samples with different quantitative attributes were treated as patterns during data analysis. These methods perform well for qualitative analysis, but may not work well for quantitative requirements. Therefore, it is necessary to develop appropriate quantitative data analysis methods for e-noses. The development of these methods was another objective of this study.

Multilayer perceptron trained by back-propagation (MLP), a simple and popular neural network topology, is a method for quantitative data analysis. Time delay neural network (TDNN), a dynamic network with embedded local memory, has demonstrated to improve an e-nose's pattern recognition ability (Zhang et al., 2003). Therefore, TDNN was selected as another possible method, with a potential to perform better than MLP. Other data analysis methods include incorporating principal component analysis (PCA), a multivariate statistical method that can preserve information in a less dimensional format, with the neural network structures described above.

Spice mixtures were selected as the model system in this study because spices have high volatile content easily detectable by an e-nose; they are stable during storage making a large experiment feasible; they are valuable for their olfactory attributes, and they are regularly used as mixtures in the industry.

For odor analysis, there are two traditional competitors of e-noses. One is gas

chromatography (GC), an instrumental method that can separate and then identify volatile chemical compounds. Another is sensory analysis, which uses the senses of human






3


subjects to measure the organoleptic characteristics of samples. The GC methods GC can provide detailed information about the volatile chemicals composing odors, and the variability of the GC methods is small. However, the experimental procedures of GC, especially the procedure of sample preparation, are involved and time-consuming. On the other hand, human sensory panels are powerful in assessing qualitative attributes of odors, and they are the ultimate arbiters in quality evaluation and product match. However, sensory methods also have limitations, including high expense and subjectivity.

The assessment of how well a newly developed e-nose procedure performs in a mixture composition prediction should be based on the comparison of the performance between the e-nose and the two traditional methods. Theoretically, e-nose methods should predict the mixture composition more efficiently than GC methods (in terms of time) with some loss of prediction accuracy (the difference between prediction and real). Human subjects, whose quantitative ability is quite limited, are not expected to rival the prediction accuracy of electronic noses. A reasonable estimation of the acceptable prediction accuracy for quality evaluation or product matching can be determined by qualitative sensory methods since human subjects are the ultimate arbiters in these areas. Under the acceptable prediction accuracy levels, the difference between the real spice mixture and the predicted one should not be perceived by human subjects.

The objective of this study was to develop both the experimental and data analysis methods of an e-nose to quickly predict the mixing compositions of ternary spice mixtures. The performance of the developed e-nose method(s) was determined in terms of accuracy and efficiency by comparing them to those from GC and sensory studies






4


using the same ternary spice mixtures. It was expected that the e-nose method could save significant time in data collection and result in reasonable prediction accuracy.













CHAPTER 2
LITERATURE REVIEW

Spices

According to the American Spice Trade Association's (ASTA) official definition, spices are "any dried plant product used primarily for seasoning" (1999). Many parts of plants are used as spices, such as seeds, leaves, berries, bark, kernels, arils, stems, stalks, rhizomes, roots, flowers, bulbs, fruits and flower buds.

Spices are often used in their dried form to assure year-round availability, ease of processing, longer shelf life and lower cost. The essential oil of a spice can be extracted by steam distillation. The extracted essential oil may exhibit the characteristic flavor and aroma properties of the actual spice, but may or may not have exactly the same taste. Essential oils do not generally carry the "biting" principles of the spice and, therefore, may taste almost bland (Hirasa and Takemasa, 1998; Uhl, 2000).

Spices are valued for the flavor, aroma, pungency and color they impart to food, not for their nutrient content. Therefore, the quality of spices must be estimated on the basis of their sensory characteristics and intensity.

Spices and their extracts should be stored in closed containers under cool and dry conditions. Recommended storage temperature is 200C with 50% relative humidity. Light-sensitive materials such as parsley, chives and other green herbs should be protected against direct exposure to sunlight and fluorescent light (Giese, 1994). Storage under refrigeration temperatures can slow microbial growth but may negatively affect sensory quality.


5






6


Electronic Nose

Introduction

The concept of electronic noses (e-nose) first appeared in the literature in the early 1980s (Persaud and Dodd, 1982). A well-accepted definition of the electronic nose is "an instrument which comprises an array of electronic chemical sensors with partial specificity and an appropriate pattern recognition system capable of recognizing simple or complex odors" (page 18, Gardner and Bartlett, 1994). The commercial instruments were available at the beginning of 1990s. Since then, the electronic nose has generated a widespread interest in the food, chemical, biotech and pharmaceutical industries. Currently, the biggest market for electronic noses is in the food area. Numerous studies have shown its potential to be used as a quality control or process-monitoring tool.

Generally, an electronic nose is composed of a sampling system, a sensor array, a data acquisition system, and a signal-processing algorithm. Mimicking the human nose, the operation of electronic nose begins with "sniffing": collecting and conveying the volatile components of the sample to the sensor array. Sensor "states" are altered through chemical or physical interaction between volatile components and sensors, resulting in electronic signals. These are captured by a data acquisition system, and a "cleaning" procedure is then applied to restore initial conditions in both sensors and sampling system. The electronic signals are further analyzed by pattern recognition algorithms or other data analysis techniques. Electronic noses are different from traditional gas sensors in two aspects: partially selective sensors that are all broadly tuned, and multidimensional data that should be analyzed using multivariate techniques.

Electronic noses can be used to evaluate foods when the food has volatile

compounds that change with the characteristics under investigation, either qualitatively or





7


quantitatively. The concentrations of volatile compounds in headspace should also be high enough to be detected by electronic noses. When they are too low to be detected at room temperature, heating the sample can release more volatile components into the headspace and may help the analysis.

Electronic noses differ from human noses in both their sensors' types/numbers and signal processing methods, resulting in unmatched detection thresholds and odor discrimination capabilities (Doleman and Lewis, 2001). In other words, what an electronic nose smells is not the same as what a human nose smells (Burl et aL., 2001). Therefore, the electronic nose is not a primary analytical technique. An electronic nose has to be trained to fulfill its capability of odor identification.

The data used for training can be collected in two ways. One is by correlating an electronic nose's responses with other primary analytical methods, such as sensory evaluation, chromatography and wet chemistry analyses. The other is by collecting an electronic nose's response toward the "known" samples. The "known" samples can either be prepared or be obtained from a supplier. Applications in Food Area

Substantial research has been done to apply electronic nose technology in the food area. Electronic nose technology has been applied for fruit ripeness determination (Benady et aL., 1995; Simon et aL., 1996; Llobet et al., 1999; Maul et al., 1999; Brezmes et al., 2000), fermentation process monitoring (Pearce et al., 1993; Ekl6v et al., 1998; Bachinger et al., 1998; Pinheiro et aL., 2002), spoilage detection (Schweizer-Berberich et aL., 1994; Blixt and Borch, 1999; Muhl et aL., 2000; Evans et aL., 2000; Korel and Balaban, 2002; Park et aL., 2002), instrument assessment of sensory attributes (AnnorFrempong et aL., 1998; Shen et aL., 2001; Korel et aL., 2001; Korel and Balaban, 2002;





8


Garcia-Gonzalez and Aparicio, 2002; Bleibaum et al., 2002), detection of packaging odors (Holmberg et al. 1995; Poling et al., 1997; Deventer and Mallikarjunan, 2002; Heinio and Ahvenainen, 2002), and aroma profile control (Gretsh et al., 1998; Hirschfelder et al., 2000; Stella et al., 2000).

Until now, most e-nose studies were aimed at detection of different odor patterns among samples, such as "fresh" versus "spoiled." There are a few studies dealing with quantitative attributes of samples. Dittmann et al. (2000) reported using an e-nose in detecting different doses of garlic flavorings. In an application note from Alpha-MOS, an e-nose was applied to discriminate binary hops blends with 3 different mix ratios (1997). However, in these studies the samples with different quantitative attributes were treated as patterns during the data analysis. This means that the developed procedure could only be used to determine whether the sample had this quantitative attribute or not, and the actual level of the quantitative attribute could not be determined. For the purpose of quality control or process monitoring, it would be valuable to know not only the qualitative aspects of samples but also their quantitative attributes. Equipment

Dozens of companies are now designing and selling electronic noses. The aspects to be considered when selecting an electronic nose instrument include its sampling system, sensor types/numbers, and data acquisition system.

The aroma of a food is a complex mixture of volatile compounds. Often the difference between a "good" or "bad" odor associated with a food is in the relative amounts of the volatile compounds. It is impossible to have sensors specific to each chemical compound. The "broadly tuned" sensors of electronic noses are designed to solve this problem. Except the non-specificity requirement, ideally each sensor's






9


response should be orthogonal to the others. The less the correlation between different sensors, the more information about sample properties is provided by the sensor array's responses. To evaluate a food whose quality deteriorates fast, the sensor's rapid response and recovery time is also important. Other requirements of the electronic nose sensors include reproducibility to a given odor and insensibility to changes in temperature, humidity and flow rate.

At present, three types of gas sensors--metal-oxide, polymer, and surface acoustic wave gas sensors--are widely used in commercial instruments (Snopok and Kruglenko, 2002). The significant advantage of metal-oxide sensors (MOS) is their low sensitivity toward humidity and low drift over time. Compared with MOS, the polymer sensors have the advantages of (1) responding to a broad range of organic vapors, (2) operating at room temperature, and (3) rapid response and recovery time. Surface acoustic wave gas sensors are of two main types: the bulk acoustic wave device (BAW), also referred to as the quartz crystal microbalance (QCM), and the surface acoustic wave device (SAW). QCM has a linear response to concentration, compared with the power law of metaloxide and Langmuir response for conducting polymers. The disadvantage of BAW and SAW is their high sensitivity to disturbances such as temperature and humidity fluctuations (Gardner and Bartlett, 1999).

There are new types of chemical sensors under development for electronic noses. One example is the so-called Smell-SeeingT (Rakow and Suslick, 2000) sensor, which detects odors using the colorimetric response from a library of immobilized vaporsensing dyes. These sensors are insensitive to humidity and provide visual identification of odors. Yet, for the purpose of objective data analysis, the color graphs generated from





10


these sensors may need complex preprocessing to generate suitable multidimensional data.

It is arguable whether a single sensing device that produces an array of

measurements of odors can be regarded as an electronic nose sensor array. There were two such "single sensing devices" that had been put under the umbrella of electronic nose sensor arrays: mass spectrometry based sensors (MS-sensor) and gas chromatograph based SAW sensors (GC/SAW). In the MS-sensor system, volatile components are introduced into mass spectrometer without separation and then selected fragment ions are treated as response (Dittmann et al., 2000). As a mature technology, mass spectrometry provides the benefit of reproducibility and standard calibration methods. Yet the cost of the instrument may compromise its benefit. For the GC/SAW system, volatile components are separated by a fast GC column and detected by SAW sensor; the responses of the single sensor at different times are viewed as a multidimensional response. The problem with this sensor is that the single sensor used may not respond to some of the important volatiles.

The sampling system is designed to provide a stable and reproducible sensor

reading environment so that all factors capable of influencing sensor responses are kept under control (Falcitelli et al., 2002). Generally the sampling system of an electronic nose has two separate chambers: a sample chamber and a sensor chamber. The temperature and/or humidity of both chambers are controlled. The static or dynamic headspace of samples is conveyed to the sensor chamber through gas flow. Inert gas is then applied to both chambers to clean any possible odor memory of previous samples or possible contamination. Some of the hand-held electronic noses, such as the product from Cyrano





II


Sciences, do not have a sample chamber. It is necessary in that case to design a sample chamber to ensure repeatable measurements. Non-absorbent and inert materials should be selected to build both chambers to prevent leftover odors from previous sample or other contaminations (Falcitelli et al., 2002).

The fluid dynamics of sampling system should also be properly designed to avoid any stagnant regions and the cleaning procedure of using inert gas to remove odor memory should have flexibility. For foods with strong volatile contents, a longer cleaning time or fast cleaning gas flow rate is expected.

The way to pre-process the time-dependent analog sensor signal is also an

important design parameter for an electronic nose instrument. The relative and fractional change of baseline is the most common output reading of electronic nose sensors. Steady state or static rather than transient or dynamic are the descriptors commonly used to define the sensors' responses.

Most commercial electronic noses provide their pattern recognition software

package (Snopok and Kruglenko, 2002). However, independent software, such as SAS*, STATISTICA*, S-Plus*, SPSS*, MATLAB* or other Neural Network packages, are also popular in electronic nose data analysis.

Data Analysis Methods

The underlying relationship between the responses of an electronic nose and the characteristics of the sample is determined by two basic approaches - statistical multivariate analysis or artificial neural networks. Currently, the data analysis methods for electronic noses are solely used to classify or recognize patterns. The commonly used statistical multivariate methods include principal component analysis (PCA), discriminant function analysis (DFA), cluster analysis (CA), partial least squares (PLS)





12


and canonical correlation analysis (CCA). For electronic nose data analysis, the most commonly used neural network structure is multi-layer perceptron (MLP) trained by back-propagation. Other neural network structures that have been applied to electronic nose data include neuro-fuzzy system (Ping and Jun, 1996), self-organizing maps (SOM) (Marco et aL. 1998), MLP with genetic algorithms (Kermani et al., 1999), adaptive resonance theory (ART) (Distante et aL., 2000) and radial basis function (RBF) (Evans et aL., 2000).

Most statistical multivariate methods are based on a linear approach while neural networks are non-linear. In cases in which data correlations are non-linear, neural networks may perform better than conventional multivariate methods.

According to the goal of the study, the data analysis methods applied for electronic noses can be divided into two categories: exploratory methods and predictive methods. Exploratory methods are used to detect whether there are systematic patterns in the investigated data set, while predictive methods are designed to find the prediction model to describe the data structures when there is priori knowledge about their existence. Among the commonly used e-nose data analysis methods, principal component analysis, cluster analysis and unsupervised neural networks, such as self organizing maps, belong to the category of the exploratory methods. Discriminant function analysis, partial least square, canonical correlation analysis and all supervised neural networks, including multilayer perceptron, are predictive methods.

As one of the most widely applied multivariate techniques, principal component analysis can also be used as a "data-compressing" tool for data pre-processing in neural networks (Principe et aL., 1999). By reducing the dimensionality of information, principal





13


component analysis lowers the number of inputs to neural networks, resulting in fewer parameters necessary in neural networks. Like any other predictive structure, the generalization ability of a neural network will be enhanced with fewer parameters. Hence, pre-processing of data by principal component analysis improves the generalization ability of neural networks. This method has been applied to classify different brand of coffees (Pardo et al., 2000).

Gas Chromatography

Introduction

Gas chromatography (GC) is an instrumental method for separation and

identification of volatile chemical compounds. The compounds of interest must first be removed from the matrix (solid or liquid phase) and isolated from any interference through the process of sampling. The extracted sample is then introduced into a heated injector, carried through a separating column by an inert gas, and detected as a series of peaks on a recorder when components leave the column. Each component of the sample reaches the detector at a different time and produces a signal at a characteristic time called a retention time. The area under a peak is related to the amount of that component present in the sample. The retention time of a compound differs from column to column and is subject to the settings of the GC operation parameters. Kovats retention indices

(KI), which reflect the retention of a given compound relative to those of the two bracketing n-paraffin hydrocarbons, are a constant for a chemical compound and can be used to identify the compound when the standard information is available. Equipment

GCs are composed of four basic units: an injector, an analytical column, a detector and a signal recorder.






14


The role of the injector is to rapidly vaporize the sample and thoroughly mix the volatiles so they can be swept onto the column by the carrier gas. The most commonly used injector is the programmed temperature vaporizing (PTV) injector that can be operated in two modes: split and splitless. While the temperature of the injector is programmed to increase, each component in the sample is vaporized and moved to the column. In split injection, a defined fraction of the sample vapor enters the column, with the remainder leaving the injection through a vent. In splitless injection, the split vent flow is blocked during the injection period such that all sample vapors enter the column (Jennings et al., 1997).

The analytical column is where the analytes are separated. The separation relies on the distribution of analytes between two phases: the stationary phase on the inside of the column and the mobile phase of the carrier gas. Today, the Fused Silica Open Tubular (FSOT) column is viewed as "state of the art." Its cross section is shown in Figure 2-1.

The factors to be considered in column selection include the stationary phase, film thickness of the stationary phase, column diameter and column length. Selection of stationary phase is based on sample polarity; for complex samples, the selected column should best reflect the overall polarity of the sample. In the area of food flavor analysis, the most commonly used column is the DB-5 or equivalent (apolar column with 95% polydimethylpolysiloxane and 5% Phenyl). Typically, Kovats retention indices for the polydimethylsiloxane column are a constant within a window of +/- 0.1%. Increasing film thickness and diameter of the column will allow for a greater sample capacity with wider peaks and lower resolutions. It is generally advisable to use small internal diameter (0.2-





15


0.32 mm) columns with thin films (0.2-0.35 pim). For a general-purpose column, 30 m is usually the most useful length (Jennings et al., 1997).

Polyimide coating
- Fused silica tube
Chemically bonded stationary phase



Figure 2-1. Cross section of a Fused Silica Open Tubular Column

As American Society of Testing and Materials (ASTM) have defined that detectors are devices that identify the presence of the components as they elute from the column (1996). There are many types of detectors available, each with its particular utility. Some are universal detectors such as the mass spectrometer (MS) or the flame ionization detector (FID). Some are selective in that they respond more to certain classes of compounds. Two examples of selective detectors are the flame photometric detector (FPD) used for the analysis of organophosphate pesticides, and the electron capture detector (ECD) used mainly for the analysis of chlorinated compounds (Scott and Perry, 1998). A third, newer detector is the atomic emission detector (AED), which can be specific to particular elements such as carbon, chlorine, bromine, iodine, tin, mercury, sulfur, nitrogen, phosphorous as well as others common to pesticides and herbicides, and thus is much less subject to interferences.

The signal recorder is used to record the signal detected by detectors. In most cases, the signal recorder is either a computer or a printer or a combination of both.

Chemically inert gas such as nitrogen, helium and hydrogen can be used as carrier gas. Hydrogen, allowing much higher flows with little loss of resolution, is the best among the three for GC applications (ASTM, 2000).





16


Sampling Methods of Volatiles

For the investigation of volatiles, the sample preparation is the most critical and

most complex step in the entire analytical process. At the present, the two most common procedures are headspace methods and extraction methods.

The following extraction methods have been reported in the literature: steam

distillation followed by solvent extraction, direct solvent extraction, simultaneous steam distillation/extraction (SDE), co-distillation and supercritical fluid extraction (SFE) (Parliament, 1997; Sides et al., 2000).

Comparing with direct solvent extraction, steam distillation followed by solvent extraction has the advantage of separating the volatiles from the nonvolatiles. Steam distillation works best for compounds that are slightly volatile and water insoluble. Steam for distillation can be generated internally or externally. The usage of external steam (indirect steam distillation) results in an extraction with less decomposition of the sample since the sample is not heated directly. If sample decomposition remains a concern, then the steam distillation may be operated under vacuum.

A direct solvent extraction is regularly performed with a Soxhlet extractor. A dried sample such as a spice or a grain can be ground finely and placed in a Soxhlet thimble and extracted by an organic solvent such as diethyl ether or methylene chloride. After a number of cycles, the solvents are then combined and concentrated. Nonvolatile organic materials such as lipids and pigments will also be concentrated. The sample may be analyzed directly or after removal of the solvent. If the sample contained large amounts of lipids as in the coffee or chocolate, subsequent steam distillation may be needed to remove the nonvolatiles.





17


Simultaneous steam distillation/extraction (SDE) apparatus was first described by Likens and Nickerson (1964). The apparatus can ensure continuously recycling for both the steam distillate and an immiscible organic solvent. SDE technique provides a single operation that can remove and highly concentrate the volatiles. In the operation, just a small volume of solvent is required, which reduces the problem of artifact buildup, as solvents are concentrated. The apparatus can be operated under reduced pressure to reduce thermal decomposition but the operation under vacuum is quite complex (Maignial et al., 1992). A number of refinements have been made to the basic SDE apparatus and several versions are commercially available (Chaintreau, 2001). Many solvents have been employed in SDE. Hexane was an excellent solvent except for lowerboiling water-soluble compounds, where diethyl ether was considerably better (Schultz et al., 1977). Use of methylene chloride has been recommended in a modified LikensNickerson extractor (Aug-Yeung and MacLeod, 1981). Currently, most researchers appear to be using pentane-diethyl ether mixtures.

A relatively new technique is co-distillation. In this technique, a solvent such as

diethyl ether, pentane, or methylene chloride is dispersed in the sample and the sample is distilled rapidly at 200*C until all the solvent and a small amount of water have passed over (Misharina et al., 1994). The advantages of co-distillation are that isolates are generated without a boiled note, and it takes only 15-20 minutes for a distillation.

The recovery of volatiles by SFE is comparable with that achieved by direct solvent extraction using Soxhlet apparatus (Ropkins and Taylor, 1996). SFE performed well in extracts tightly bound and encapsulated volatiles, and the solvent is easy to remove. However, the equipment of the SFE system is complex (consisting of a source fluid, a





18


pump, a sample vessel, a restrictor and a trapping device) and optimization of extraction conditions such as temperature and pressure is necessary (Messer et al., 1998; Burford, 1998; Diaz-Maroto et al., 2002).

Headspace sampling techniques are frequently divided into two broad categories: static headspace and dynamic headspace. In the headspace analysis, volatile analytes from a solid or liquid matrix are sampled by investigation of the atmosphere adjacent to the sample (Wampler, 1997; Rouseff and Cadwallader, 2001). Solid-phase microextraction (SPME) is a relatively new technique that can be used for headspace sampling (Harmon, 1997).

Static headspace sampling is performed as follows: the food sample is placed into a headspace vial, sealed and then allowed to stand for a period of time to establish equilibrium at that temperature. The vial may be warmed to enhance vaporization of volatile. A small sample (usually about Iml) of the atmosphere around the sample is injected directly into the GC column. The amount of gas that may be injected into a gas chromatography is limited by the capacity of injection port, the column, and consideration of the increase in the pressure and flow in the injection port caused by a gas phase injection. The disadvantage of static headspace methods is that they may lack sensitivity due to low concentration of volatiles in sample headspaces (Kolb and Ettre, 1997).

Dynamic headspace involves moving the analytes away from the sample headspace and concentrating in some kind of "trap." Instead of allowing the sample volatiles to come to equilibrium between the sample matrix and the surrounding headspace, the atmosphere around sample matrix is constantly swept away by a flow of carrier gas,





19


taking the volatiles with it. This prevents the establishment of an equilibrium state and then allows more release of volatiles into the headspace. Therefore, dynamic headspace offers increased sensitivity compared to static headspace. "Purge and trap" is another term used in literature for dynamic headspace technique. Regularly, purge and trap refers to the technique applied to a liquid sample, while "dynamic headspace" is used when the sample material is a solid.

The trapping system for dynamic headspace can be either sorbent or cryogenic trapping. Most sorbent materials are porous polymers similar to materials used to fill packed GC columns for gas analysis. Tenax@ is the most widely used, general purpose sorbent for dynamic headspace techniques. Both liquid nitrogen and solid carbon dioxide can be applied for cryogenic trapping. Since the presence of water on a capillary GC column can pose a serious analytical problem, it is important to remove water that is carried away from the sample and collected in the trap before transferring the trapped organics to the chromatograph. Tenax@ is hydrophobic. It is usually enough to pass a source of dry carrier gas through the trap for a minute or two to vent the water from the trap without disturbing the organics.

SPME was developed by Pawliszyn's research group at the University of Waterloo in the late 1980s. It is a solventless technique that incorporates extraction, concentration, and sample introduction. The SPME devices are syringe-like with an outer septumpiercing needle. A plunger houses a fused silica fiber coated with a stationary phase. The fiber can be inserted into the sample matrix (aqueous samples) or the sample headspace. After concentration of analytes on the fiber, the syringe assembly is inserted into the injection port of a gas chromatograph where the analytes are thermally desorbed from the





20


fiber and cold-trapped on the head of the capillary column (Pawliszyn, 1999; Pawliszyn, 2000). Fibers coated with nonpolar polydimethylsiloxane (similar to OV-101) and the more polar polyacrylate are commercially available. For most volatile flavor analyses, a fiber having a 100-p m coating of polydimethylsiloxane is preferred. In general, the fibers coated with thicker films will require a somewhat longer time to achieve equilibrium but might provide higher sensitivity due to the greater mass of the analytes that can be absorbed. For analysis of high-boiling-point components, fibers with a 7-micron thickness of polydimethylsiloxane usually work best. The major advantages of SPME include fast, solvent free, easy-to-automate and sensitive to high-boiling-point components. The key to getting accurate and reproducible results with SPME is to be sure to perform sampling in exactly the same way each time. The actual position of the fiber in the headspace is also important (Cai et al., 2001; Marsili, 2001).

One should be aware that none of these techniques of volatile sampling produce an isolate that quantitatively represents the composition of the starting material. Jennings et al. (1977) has compared various sample preparation techniques, including dynamic headspace and simultaneous distillation-extraction. Their conclusion was that the extracts collected by distillation-extraction most nearly agreed with the original sample. Application to Spice Mixture Analysis

The identification and quantification of volatile components of spices have been well-investigated using GC methods and their combination with mass spectrometry. However, most of published articles were based on the analysis of an individual spice and there were a few published articles of identification/quantification of spices from a mixture. Cheng et al. (1997) reported a qualitative and quantitative spice mixture analysis





21


method. The spice mixtures were extracted by simultaneous steam distillation/extraction

and then injected into GC. The recognition and quantification of spices was accomplished

via numerical methods based on a database consisting of 355 different spices. The

accuracy of the method was not defined since the spices in the prepared mixture and the

spices in the database were not identical.

Sensory Evaluation

Introduction

Sensory evaluation has been defined as "a scientific discipline used to evoke,

measure, analyze, and interpret reactions to stimuli perceived through the senses" (page

3, ASTM, 2001). As summarized in Table 2.1, the commonly used sensory tests can be

divided into three categories according to their goals and their criteria for the selection of

panelists (Lawless and Heymann, 1998; Meilgaard et al., 1999; ASTM, 2001).

Table 2-1. Classification of test methods in sensory evaluation
Categories Question of Test Methods Panelist Characteristics
Interest
Discrimination Is there any Triangle Screened for sensory
detectable Duo-trio acuity, oriented to test
difference in "A"- not "A" method, sometimes
products? Paired comparison trained
Rating
Threshold
Descriptive How do products Flavor profile Screened for sensory
differ in specific Texture profile acuity and motivation,
sensory QDA@ T trained or highly trained
characteristics? Spectrum"
Time-Intensity
Free-choice

Affective How well are Ranking Screened for product use,
products liked or Hedonic untrained
which products
are preferred?






22


The sensory perception of stimulus is an active and selective process, which is influenced by both physiological and psychological factors. Those factors were well discussed by Meilgaard et al. (1999). There is also standard practice for serving protocol to avoid bias generated by these factors (ASTM, 1997a). Measurement of Sensory Thresholds

Threshold is a measure of human sensitivity to a given stimulus. Four types of thresholds exist, namely absolute threshold, recognition threshold, difference threshold and terminal threshold, each described below (Meilgaard et al., 1999; ASTM, 2001):

Absolute threshold: also called detection threshold, is the minimum intensity of stimulation capable of eliciting a response at a probability of 50%.

Recognition threshold: the minimum intensity of stimulation at which the stimulus can be recognized and identified with a specific probability, most frequently 0.50.

Difference threshold: the minimum of difference required between two stimuli that will elicit a perceived difference with a specific probability, most frequently 0.50.

Terminal threshold: the intensity of stimulation above which increase in intensity cannot be detected.

In the early days of psychophysics, thresholds were commonly measured by the

"method of limits" or the "method of constant stimuli." In the method of limits, stimulus intensity would be raised and then lowered sequentially to find the average point at which the observer's response changed from negative to positive or from positive to negative. In the method of constant stimuli, a number of stimuli values are selected to "bracket" the assumed threshold. These stimuli are then presented many times to the subjects in a randomized order. For each of the stimuli, the percentage of correct response is calculated and the threshold is interpolated based on the predetermined criterion. The





23


problem associated with the classical methods is that the perception of a response depends on the observer's criteria and willingness to guess. An objective proof of detection can be achieved by incorporating the forced-choice element (subjects are forced to make guesses whenever they are in doubt) into the classic methods (Moulton et al., 1975; Baird, 1997). A typical example of such methods is the ascending series of 3-AFC tests for absolute or recognition threshold detection, which has been described in the ASTM standard practice E679-91 and E1432-91. The test procedure works as follow: First, based on the estimated threshold range, the test samples whose concentrations "bracket" the threshold are prepared. These samples should have over three to four concentration and each concentration differed by a factor of 2 or 3. The test samples are then presented to the test panel that should have more than 25 panelists in an ascending 3AFC series (Brown et aL., 1978; ASTM, 1991a; ASTM 1991b). It should be noted that in forced-choice methods, part of the correct responses are not elicited by perceived sample characteristics, but identified by guessing. Therefore, for threshold determination using forced-choice methods, the criterion that defines a threshold is actually the probability of correct response excluding the guessing rate. This probability is referred as "percent correct above chance" in the literature.

Group thresholds can be derived by averaging the individual thresholds or from group "percent correct above chance" at each stimulus level. In the latter approach, the tests results from forced-choice method the can be converted to percent correct at a stimulus level by a defined formula (Moulton et aL., 1975; ASTM, 199 1b; Antinone et aL., 1994).






24


Most of the forced-choice tests used in the threshold determination are based on discrimination methods. Threshold testing can be viewed as a special case of discrimination (simple difference) testing. Since some discrimination methods present a more difficult task to the participants in the test, the value of a threshold is not fixed, but is method-dependent and will rise with the difficulty of the task.

In recent years, the Signal Detection Theory (SDT) has become popular among

psychophysicists to determine thresholds. In SDT methods, the point of interest is not the thresholds, but "the size of the psychological difference between the two stimuli", which has the name of d'. The advantage of SDT is that the decision process of subjects becomes more explicit and can be modeled statistically (Meilgaard et al., 1999). However, SDT procedures are more time-consuming, and Frijters and his co-workers' studies have shown that there is a 1:1 relationship between d' and the classical threshold for forced-choice methods (Frijters, 1980a; Frijters et al., 1980b). For these reasons, both ASTM (1991a and 1991b) and ISO (1999) are still using the method of limits as their practice standard for threshold measurements. Discrimination Tests

There are two groups of discrimination tests: overall discrimination tests and

attribute discrimination tests. Overall discrimination tests are designed to fmd whether a sensory difference exists between samples, while attribute discrimination tests are designed to answer the question "how does attribute X differ between samples?" Three most commonly used discrimination tests are triangle, duo-trio and n-alternative forcedchoice procedure (n-AFC) (Meilgaard et al., 1999).

In the triangle test, three samples are presented simultaneously to the panelists; two samples are from the same formulation and one is different. Each panelist has to indicate






25


which sample is the odd one. Generally, 20 to 40 panelists are necessary for a triangle test, while 50 to 100 panelists are required if the study seeks to demonstrate similarity. The panelists should be familiar with the triangle test and with the product tested (ASTM, 1997b; Meilgaard et al., 1999). Triangle tests, along with duo-trio, are overall discrimination tests.

N-AFC methods, also called directional difference tests, are attribute discrimination tests. In n-alternative forced-choice procedure, the panelists are asked to choose the object with the most (or the least) of an attribute from among n objects, n-I of which have identical independent distributions of the sensory attribute. The danger with the method is that other sensory changes may occur when one attribute is modified and these may obscure the attribute in question. The panelists should be trained to be familiar with the attribute in question (Lawless and Heymann, 1998; Meilgaard et al., 1999).

Power of a discrimination test is defined as 1- P, where P is the probability of

committing a Type II error (not rejecting the null hypothesis when the null hypothesis is not true). Triangle and duo-trio tests are much less powerful than n-AFC tests. This is due to the difficulty of the tests being performed. In n-AFC methods, subjects need only determine the maximum or minimum of an attribute, while in triangle and duo-trio subjects are assumed to compare distances between pairs of percepts (Ennis, 1990).

In case there is a known sensory attribute that differs in samples, n-AFC is more efficient and powerful than the overall discrimination test like triangle and duo-trio. However, when the sensory attribute(s) differing in the sample are unknown or there is a comparison for multidimensional stimuli, an overall discrimination test may be used since it is not necessary to state the sensory attribute or percept involved in making the






26


discrimination judgment for this type of tests (Ennis, 1990; Lawless and Heymann, 1998).

Difference Threshold - Psychophysical Theory

Difference thresholds are also referred to as just noticeable differences (JNDs) in the literature. There is at least one psychophysical law directly associated with difference thresholds: Weber's Law states that difference thresholds increase in proportion to the background intensity, e.g.,

AS = kS (2.1)

where AS is a difference threshold, S is the intensity of a background stimulus and k is a Weber fraction. The value of k reflects the discriminability of closely spaced stimuli: the higher k, the lower the sensitivity. A typical Weber fraction for taste (salt) is 0.14, while that of smell is 0.24. Weber's Law is not always true, but it is good as a baseline to compare performance and as a rule-of-thumb. Weber's Law often fails near absolute threshold. A modified version of Weber's law is as follows: AS = k(S + c) (2.2)

where c is a constant, usually small that represents a baseline level of stimulus that must be surpassed (Baird, 1997).

Difference Thresholds of Mixtures

For chemical senses (taste and smell), the perception behavior of a mixture can be very complicated (Gregson, 1984; Fritjers, 1987; Gregson, 1992). No study was found in the difference thresholds detection for mixtures. A related study found was by comparing the ratings of non-trained panelists, the assessment of the model that described overall similarities between binary odor mixtures (Gregson, 1984).






27


Neural Networks

Introduction

An artificial neural network, inspired by the way in which the brain works,

resembles the brain in two respects: (1) Knowledge is acquired by the network through a learning process and (2) Interneuron connection strengths, known as synaptic weights, are used to store the acquired knowledge (Haykin, 1999). A neural network has the ability to learn and therefore generalize. In other words, it can build up an implicit model based directly on real-life data and then uses the implicit model to predict the unknown cases.

There are two main elements in a neural network: synapses (connections) and

neurons (nodes). The way in which the neurons of a network connect with each other is called its topology or architecture. There are three fundamental classes of network topologies: single-layer feedforward, multilayer feedforward and recurrent network. The details of network topology include the number of layers, the number of nodes in each layer and full or part connection of the network. A learning algorithm refers to the welldefined rules of how to adjust the synaptic weights and bias of neural networks according to real-life data.

The design of a neural network may proceed as follows: First is the learning step an appropriate architecture is selected and real-life data are used to train the network by a suitable learning algorithm. Second is the generalization step - the recognition performance of the trained network is tested with data not seen before.

Network size, e.g., the number of parameters adjusted by using the learning algorithm, affects the learning and generation ability of the network. The larger the network, the better the fitting to the learning data provided that enough data are available






28


to train the network. Too large a network, however, cannot generalize well because the fit is too specific to the training-set data. So, the neural network should be sufficiently large to solve the problem, but not larger (Principe et al., 1999).

It is ideal to randomize the order of presentation of training examples from one

epoch to the next. This randomization tends to make the search in weight space stochastic (involving chance or probability) over the learning cycles (Haykin, 1999). Multilayer Perceptrons (MLPs) Trained by Back-Propagation (BP)

The perceptron, consisting of a single neuron with adjustable synaptic weights and bias, is the simplest form of a neural network. It can be used for the classification of patterns that are linearly separable (i.e., patterns that lie on opposite sides of a hyperplane) (Haykin, 1999). Multilayer Perceptrons (MLPs) extend the perceptron with hidden layers. Typically, a MLP is constituted with an input layer, one or more hidden layers of neurons, and an output layer, as shown in Figure 2-2. The neurons in both hidden layers and output layers have computational ability. The input signal propagates through the network on a layer-by-layer basis and in a forward direction. The computing power of the MLP lies in its hidden neurons that act as feature detector.

Multilayer Perceptrons are regularly trained in a supervised manner with a highly popular algorithm known as "back-propagation." Basically, back-propagation consists of two passes through the different layers of the network: a forward pass of signals and a backward pass of the local error. This algorithm can be applied independent of the topology of the network and the input dimensionality. It adjusts the synaptic weights and biases in five steps:

1. Initialization. When there is no prior information available (regularly the

case), randomize the synaptic weights and biases. These synaptic weights






29


or biases should form a uniform distribution whose mean is zero and

whose variance is chosen to make the standard deviation of the induced

local fields of the neurons lie at the transition between the linear and

saturated parts of the sigmoid activation function.

2. Presentation of Training Examples. Present the network with an epoch

(batch) of training set. The sequence of forward and backward

computations, as described under 3 and 4 respectively, is performed for

each example in the set.

3. Forward Computation. The signals of an example in the training set are

applied to the input neurons of the network, and its effect propagates

through the network layer by layer. A set of outputs is then produced as the

actual response of the network. The synaptic weights / biases are all fixed

during the forward computation.

4. Backward Computation. The actual response of the network is subtracted

from a desired (target) response to produce the error signal. The error is

propagated and scaled back by the chain rule. The synaptic weights are adjusted to make the actual response of the network move closer to the

desired response in a statistical sense.

5. Iteration. Iterate the forward and backward computations under 3 and 4 by

presenting new epochs of training examples to the network until the

stopping criterion is met.

MLPs are universal approximators. They can be used for both pattern classification and function approximation. The difference between using an MLP for function





30


approximation and for classification is that for function approximation the output neurons are linear, while for classification the output neurons must be nonlinear. It has been mentioned that the feature detection ability lies in the hidden layer. Provided enough neurons are available in the hidden layer, any continuous function can be approximated by the MLP topology. Two-hidden-layers are necessary when the functions or patterns to be modeled are discontinuous. There are rare applications that need more than two hidden layers (Principe et al., 1999).

Input Hidden Output
layer layer layer
xl









g2( wx1 +br ), i = 1,2,..m
1-1


Figure 2-2. Multilayer Perceptrons (MLPs) with one hidden-layer Time-Delay Neural Network (TDNN)

In temporal processes such as speech signals or fluctuations in stock market prices, the measurements from the world are functions of time. For a neural network response to the temporal structure of information-bearing signals, it must be given memory. Memory may be divided into "short-term" and "long-term" memory. Long-term memory is stored in the synaptic weights of the network through supervised learning. Short-term memory, on the other hand, can be incorporated into the structure of a neural network through the





31


use of time delays, which can be implemented at the synaptic level inside the network or at the input layer of the network.

Memory structures are sensitive to the sequence of information presentation. By embedding memory into the structure of a static network such as an ordinary multilayer perceptron, the output of the network becomes a function of time. This approach for building a nonlinear dynamical system provides a clear separation of responsibilities: the static network accounts for nonlinearity, and the memory accounts for temporal effects.

Figure 2-3 shows a diagram of the simplest and most commonly used form of

short-term memory called tapped delay line memory. The delay unit, denoted by z-1, is a linear system that delays the input signal by one time unit. The tapped delay line is built from a cascade of these delay units.


Input Unit I Unit 2 Unit p

x(n)


x(-1) x(n-2)+) x(np)

Figure 2-3. Tapped delay line with p delay units

The Time-delay neural network (TDNN) is a multilayer feedforward network with embedded local memory (tapped delay line) in both input and hidden layers. As shown in Figure 2-4, all the taps of the tapped delay lines are connected to the neurons of the next layer. A TDNN was first described by Waibel et al. (1989) and was devised to capture explicitly the time symmetry in the recognition of phonemes. In this study, a TDNN with two hidden layers was used to recognize three isolated words: "bee", "dee", and "gee" and achieved an average recognition rate of 98.5%.






32


Xp(ll)




Memory


Z-1

Z



Z




Figure 2-4. Paradigm of Time-delay Neural Network with one hidden-layer

It should be noted that for TDNNs, not only the inputs are time signals, but also the desired responses.

The problems of designing TDNN topologies are the same as for the MLP, with the addition of choosing the size of the tapped delay line (also called the memory layer). The size of memory layer depends on the number of past samples that are needed to describe the time structure of the inputs. If the size is too small, there may not be enough data from the past and the performance of neural network will suffer. On the other hand, if the






33


size is too big, there will be many more weights in the network, resulting in slower training. In the TDNN, the size of memory layer has to be physically modified.

Since the tapped delay line at the input does not have any free parameters, in the TDNN static back-propagation algorithm still can be used to train the network (Principe et al., 1999).

Softmax

It is required that each output is between 0 and 1 and all outputs sum to 1 in order to be able to interpret a MLP's outputs as posteriori probabilities (Bihop, 1995). This can be implemented by utilizing an output neuron with softmax activation function:

= exp(net,)
Y Zexp(net(

where the denominator sums over all network outputs. The softmax function is similar to the sigmoid neurons, except that the outputs are scaled by the total activation at the output layer. The softmax is effectively a competitive structure that forces the sum of the outputs of the neuron to be 1.













CHAPTER 3
OBJECTIVE

The overall objective of this project was to develop a quantitative method to

quickly predict the mixing compositions of a ternary spice mixture by using electronic nose as the measurement tool, and neural network/multivariate statistical methods for data analysis, and to compare the testing results from the electronic nose with that from GC and sensory analyses. The specific objectives of the project were: Electronic Nose Analysis

To obtain sufficient e-nose response data on various spice mixtures to perform data analysis and to develop statistical/neural network data analysis procedures. The developed procedures are capable of predicting the mixture compositions of unknown spice mixtures (testing samples) with acceptable accuracy. GC Analysis

To identify and quantify the volatile compounds in spice mixtures, and to determine the accuracy and efficiency of GC methods in predicting spice mixture compositions, and to compare e-nose and GC's prediction accuracies and the time/effort necessary to obtain these accuracies.

Sensory Analysis

To assess difference thresholds of the spice mixtures, and to determine whether the e-nose can predict the mixture compositions with an error below the perception threshold of human subjects.


34













CHAPTER 4
MATERIALS AND METHODS

Materials

Three ground spices: basil (Ocimum basilicum), cinnamon (Cinnamomum cassia) and garlic (Allium sativum) were selected to formulate the spice mixtures. This selection was based on the criteria: 1) the major volatile components of the three spices should be different from each other and should be easily separated by GC, and thus facilitates the mixture analysis by GC, 2) the odor of the three spices should be significantly different from each other and should be pleasant to smell, which facilitates the sensory analysis using human subjects. As shown in Table 4-1, the major volatile components of basil, cinnamon and garlic were different. Their Kovats retention indices are also different enough to allow acceptable separation by a DB-5 column (Wijesekera, 1978; Yu et al., 1989; Yu et al., 1993; Vernin et aL., 1994; Yu et aL., 1994; Adams, 1995; Kim et aL., 1995; Marotti et aL., 1996; Miller et aL., 1996; Lachowicz et al., 1997; Antonelli et al., 1998; Uhl, 2000; Acree and Am, 2001). Preliminary studies with human subjects showed that the three spices smelled significantly differently and had pleasant odors whether in pure form or in the formulated mixtures.

Basil was purchased from Cibolo Junction Food & Spice Company (Albuquerque, NM). Cinnamon and garlic were purchased from Frontier Natural Products (Norway, IA). Each spice was sealed in a large glass container and stored under refrigeration (30C) for further use. To eliminate the effects of different humidity on e-nose sensors, the three spices were individually equilibrated with saturated potassium carbonate solution at room


35





36


temperature (21 0C to 220C) for 20 hours to reach the same relative humidity (43.2%) before any analysis. For GC and sensory analyses, these spices with adjusted humidity were used to ensure a fair comparison among the three methods. Table 4-1. The chemical names and their corresponding CAS# and KI (for DB-5) of
major volatile compounds of basil, cinnamon and garlic
Spice Name CAS# Chemical Name KI (DB-5)
Basil 78-70-6 Linalool 1100
1Basil 1140-67-0 Estragole IN/A
Basil 97-53-0 JEugenol 11356
Basil 470-82-6 1,8-cineole 11030
Cinnamon 14371-10-9 Itrans-Cinnamaldehyde 11266
Cinnamon 103-54-8 Cinnamyl acetate 11443
Cinnamon 140-10-3 Itrans-Cinnamic acid 11438
Cinnamon 100-52-7 jBenzaldehyde 1968
Garlic 2179-57-9 IDiallyl disulfide 11320
Garlic 592-88-1 Diallyl sulfide 11085
Garlic IN/A Methyl allyl disulfide 11149

Experimental Designs and Methods Mixture Experimental Design of Three Components

With three components, the simplex space of the mixtures is an equilateral triangle. The fractions of three components can be either weight or mole based, and are usually denoted by x1, x2, and x3, respectively. The compositions of any mixture with three components 1, 2 and 3 are represented as (xi, x2, x3). As shown in Figure 4-1 part A, the vertices of the triangle represent the single-component mixtures and the internal points of the triangle represent mixtures in which none of the three components are absent. The way to identify the fractions of a mixture in the simplex space was illustrated in Figure 41 part B. Suppose 0 was an arbitrary point in the simplex space. To get the value of x1, line EF was drawn parallel to the triangle boundary line BC. AH is a line perpendicular to






37


BC. The value of xj is equal to the ratio of the length of segment GH to that of AH. The value x2 and x3 can be determined in a similar way.

component 3 component 3 B)
(0,0,1) A





(0.5, 0, 0.5) (0, 0.5,0.5)

0 0
(1/3,1/3,1/3) E - ------- - --- F
(xI,x2,xs) |G

(1,0,0) (03,0.5, 0) (01,1,0) B H C
component 1 component 2 component 1 component 2

Figure 4-1. The simplex factor space with three components. A) The vertices of the
triangle represent the single-component mixtures; the points in the boundary line are the mixture of two components; and the central point represent equal fractions of the three components. B) For an arbitrary point 0 in the simplex space, the value of x, is equal to the ratio of the length of segment GH to that of AH while line EF was parallel to line BC and AH is a line perpendicular to
BC.

There are three basic mixture design methods: simplex-lattice, simplex-centroid and axial design, as shown in the parts (A), (B) and (C) of Figure 4-2. Each dot in the figure indicates a design point. The simplex-lattice and component simplex-centroid designs are boundary designs in that most of the design points, except the overall centroid, are on the boundaries of the simplex space. Axial designs, on the other hand, are designs where most of the points are positioned inside the simplex space. Some designs could be a combination of the three methods, as shown in part (D) of Figure 4-2 (Cornell, 1990).






38


A)


6


C1)



0

- I
I A
S
- =
- I
- I
.5 4 S.
- I


B)












D)



*


Figure 4-2. The three basic design methods in the simplex space with three components:
(A){3, 3} simplex-lattice; (B) three components simplex-centroid, (C) axial
design, and a combination of the three methods: (D) simplex-centroid
augmented with 3 interior points.

Electronic Nose

To obtain sufficient data for prediction model building, the training points should have a satisfactory distribution throughout the experimental region and provide an internal estimate of the error variance. On the other hand, the testing points that test the performance of the model built should be randomly distributed in the experimental region. The selected spice mixtures for training/cross-validation and for testing are illustrated in Figure 4-3. The spice mixtures used for training/cross-validation are symbolized by dots and numbered from 1 to 19; those for testing are symbolized by small triangles and denoted as IT through 5T. The fractions (based on weight) of spices






39


corresponding to each design point were listed in Table 4-2 (training) and Table 4-3 (testing). The weight fraction of basil, cinnamon and garlic were represented by xj, x2 and

x3, respectively. In the testing spice mixtures, the weight fractions x, , X2, x3 were generated using the random function in Microsoft Excel using the following equations:

x, = RAND( )

2 =RAND( )x (1 - x,)

x3 =1-x1 -x2


GARLIC
3

16

1T
13
5 * 6



9 8
7
18 7 19
4T 2TA
5T
3T I t * A
A 11 10 12
14 15
1 2
BASIL 17 4 CINNAMON


Figure 4-3. The distribution of design points in the simplex space, with the dots,
numbered from 1 to 19 representing the spice mixtures used for training/crossvalidation and the small triangles, denoted as 1T through 5T, representing
those for testing.






40


Table 4-2. The weight fractions of design points for training/cross-validation Mix # x, (Basil) x2 (Cinnamon) x3 (Garlic)
1 1.0 0.0 0.0
2 0.0 1.0 0.0
3 0.0 0.0 1.0
4 0.333 0.667 0
5 0.333 0.0 0.667
6 0 0.333 0.667
7 0.333 0.333 0.333
8 0.167 0.417 0.417
9 0.417 0.167 0.417
10 0.417 0.417 0.167
11 0.667 0.167 0.167
12 0.167 0.667 0.167
13 0.167 0.167 0.667
14 0.833 0.083 0.083
15 0.083 0.833 0.083
16 0.083 0.083 0.833
17 0.667 0.333 0
18 0.667 0.0 0.333
19 0 0.667 0.333

Table 4-3. The weight fractions of spices for testing Mix # x, (Basil) x2 (Cinnamon) x3 (Garlic)
IT 0.218 0.015 0.767
2T 0.114 0.610 0.277
3T 0.829 0.047 0.124
4T 0.413 0.366 0.221
5T 0.052 0.777 0.171

Spice mixtures were formulated in the weight fractions listed in Tables 4-2 and 4-3.

Each experimental sample contained 10-gram formulated spice mixture, which was

stored in a 70 ml weighting bottle (I.D.x H = 40 x 80 mm, Fisher Scientific, Fair Lawn,

NJ). Eight samples were prepared for each training/cross-validation point and five

samples for each testing point. The prepared samples were presented to an e-nose (e-Nose

4000, Neotronics, Gainesville, GA) with 12 conducting polymer sensors (sensor types:

483, 478, 464, 463, 462, 461, 460, 459, 458, 401, 298 and 297) that had been calibrated

using 75% v/v propylene glycol water solution as recommended by the manufacturer






41


(100% propylene glycol purchased from Fisher Scientific, Fair Lawn, NJ). The order of sample presentation was randomized. For each measurement, the sample chamber was purged by compressed dry air (BOC GASES, Murray Hill, NJ) for 4 minutes and the sensor chamber was purged 2 minutes to eliminate any foreign odor from environment or any residue odor from previous experiments. The flow rate of the compressed air (BOC GASES, Murray Hill, NJ) used for both chambers was 800 cm3/min. The sensors' responses were recorded every second up to 4 minutes. All the electronic nose experiments were run at room temperature (230C to 250C). Gas Chromatography

Simultaneous distillation-extraction (SDE) was selected to collect the volatile fraction of the spice samples for the following reasons:

1) SDE was the only method found in the literature to deal with spice

mixtures (Cheng et al., 1997).

2) Preliminary studies have shown that static headspace, Solid-Phase

Microextraction (SPME) and direct solvent extraction methods were not suitable for the spice mixtures. The static headspace method did not have

enough sensitivity for all three spices. For the SPME method, the fibers were overwhelmed by the volatile compounds in cinnamon, resulting in lack of sensitivity towards the volatile compounds from garlic and basil.

Significant amount of pigments, especially these from basil, were

concentrated in the extracts by the direct solvent extraction method,

making the GC tests difficult.

3) Both dynamic headspace and supercritical fluid extraction (SFE) are

complex in operations and equipment. Dynamic headspace methods may






42


have the same problem as SPME methods, while SFE may have the same

problem as the direct solvent extraction method.

For the SDE method, the assumption was that the volatiles were proportionally extracted and concentrated for GC analysis (Chien 1985; Lawrence and Shu, 1993; Cheng et al., 1997). Based on the above assumption and the objective of GC analysis (to compare the accuracy of the GC to predict the spice mixing fractions with that of the enose), the GC analyses were only carried out for the pure spices and the testing mixtures shown in Table 4-3.

The solvent employed for SDE analysis was the pentane-diethyl ether (1:1

mixture), which was lighter than water. Both the n-pentane and diethyl ether were HPLC grade (Fisher Scientific, Fair Lawn, NJ).

The apparatus (#523010, Kimble/Kontes, Vineland, NJ) used for SDE was similar to that described by Likens and Nickerson (1964) and was illustrated in Figure 4-4. The vapor of the solvent entered into the condensation chamber (composed of three cooling jackets/condensers running ice-cold water) through arm C, while the steam and the volatiles from spices entered into the chamber through arm D. The volatiles were extracted by the solvent at the large condensation surface, and most of the volatiles were transferred to the solvent phase and returned to solvent flask through arm A. The water phase with limited volatile components had to pass through the solvent trap in arm A, where more volatiles were transferred into the solvent phase. The water phase then returned to the water/spice flask through arm B. This apparatus provided continuous fresh solvent for volatile extraction. By carefully controlling the temperatures of both solvent and water/spice sides, the volatiles were transferred and concentrated in the solvent flask.







43


4

















~-/ )













i~F < hr


C


/


.4-


Recycle


\ I1Pen Pe



A B

ter Phe' B


A B




5ml solvent
In 100ml flask


Water bath at 500C 5g spice mixtures +
100ml distilled water in 500ml flask Hot plate - Stirrer




Figure 4-4. The apparatus used for simultaneous distillation extraction.


D






44


As shown in the Table 4-4, preliminary results from spice mixtures with equal

proportions of basil, cinnamon and garlic showed that more than 94% and 98% of total volatiles were extracted after 1.5 and 2 hours, respectively. Table 4-4. The changes of the percentage of extraction of the total volatiles from the
spice mixtures with equal proportions of basil, cinnamon and garlic versus
distillation time using SDE
Extraction Time (hours) Percentage of extraction (%)
1 90.9
1.5 94.5
2 98.5
2.5 98.8
3 99.6
3.5 99.9

In this study, the SDE procedure was carried as follows:

- The spice sample (5 grams) was put into a 500ml round flask and mixed with

100ml distilled water;

- Solvent (pentane-diethyl ether 1:1, 5ml) was added to the solvent flask;

- Distilled water (1 7ml) and solvent (IIml) were added into arm A and arm B,

respectively;

- Temperature of water bath was brought to 500C and the temperature of water

bath was then controlled at 50 0.50C;

- After 3 minutes, heating of the water/spice flask was begun. The heating and

stirring rate were carefully controlled to keep the flask in boiling condition

(1000C);

- Started to count distillation time after the condensation of the steam/volatile

showed in the condensation chamber. This typically took 15 minutes after

previous step;

- Operated in the conditions as described above for 2 hours;





45


- Stopped the heat for the water/spice flask;

- Turned off the heat for water bath after 20 minutes. By using ice, the

temperature of water bath was then lowered to 0-50C;

- Waited additional 10 minutes;

- Immediately transferred all the extracts in the solvent flask to a calibrated vial

and adjusted the volume to 5ml;

- Added 5pl decane (Sigma-Aldrich , St. Louis, MO) into the extracts as a

standard;

- The vial was capped and stored in refrigerator until use.

The extracts were introduced by a Ig1p syringe (MICROLITER* #7101, Hamilton Co., Reno, NV) into the injection port of the GC (Shimadzu GC 14A, Norcross, GA). The separation of the volatiles was carried out on a DB-5 column (length: 30m; I.D.:

0.25mm; film thickness: 0.25pm; J&W Scientific, Folsom, CA) for a total of 30 minutes. The separated volatiles were detected by a flame ionization detector (FID). The temperature profile used for column was as follows: after 3 minutes holding at 300C, the temperature was increased from 300C to 1500C at 1 00C per minute, and then increased from 1500C to 2250C at 50C per minute. The temperature of the injection port was set as 2500C, while that of the FID detector was set at 2600C. The carrier gas used was hydrogen that generated by Hydrogen Generator 9200 (Packard Instrument Co., Meriden CT) and the pressure for the carrier gas was set at 0.5kg/cm2 . For the FID in GC, the air (BOC GASES, Murray Hill, NJ) pressure was set at 0.5kg/cm2, while the pressure for hydrogen was set at I.0kg/cm2





46


Three replicates were performed for each pure spice, and the extract from each of the three replicates was injected into GC twice (3 replicates x 2 injections for each pure spice). Two replicates were performed for each of the spice mixtures listed in Table 4-3. The extract from the spice mixtures was also injected into GC twice (2 replicates x 2 injections for each mixture).

For each of the three pure spices, the retention times for five biggest unique peaks were recorded. The peak areas under these retention times (a total of 5 x 3 retention times) were recorded for all injections, including those from the pure spices and those from the spice mixtures. For each injection, the peak area of the added standard (decane) was also recorded.

Volatile components in the extracts were identified by either comparing their

calculated Kovats retention indices (RI) to those from literatures or by adding the known chemical standards to the extracts. If the added chemical standard increases the peak area of a volatile, that volatile can be assumed to be identical to the chemical standard. All the chemical standards were purchased either from Acros Organic (Morris Plains, NJ) or Sigma-Aldrich (St. Louis, MO). For a certain volatile compounds, the Kovats retention index were calculated by formula 4-1:

C +(rt - rtc, )
RI =100 x (4.1)
(rtC, -rtC)

where RI is the Kovats retention index, rtc, and rtc2 are the retention times of two consecutive n-paraffin hydrocarbons that bracket the retention time (rt) of the investigated compound, with rtcj less than rt and rtc2 larger than rt. The value of "C" is






47


determined by counting the number of carbons in the molecule of the n-paraffin whose

retention time is rtcl.

The retention times of the n-paraffin hydrocarbons in the GC set were obtained by

injecting the C7-C20 mixtures (Supelco, Bellefonte, PA) into the GC. Those retention

times are listed in Table 4-5.

Table 4-5. The retention times of n-paraffin hydrocarbons (C7-C20) Standards Run 1 Run 2 Average
RI (minutes) RI (minutes) RI (minutes)
C7 4.002 3.991 3.997
C8 6.205 6.196 6.201
C9 8.498 8.493 8.496
CIO 10.626 10.62 10.623
Cli 12.506 12.503 12.505
C12 14.247 14.244 14.246
C13 15.898 15.895 15.897
C14 17.628 17.624 17.626
C15 19.474 19.469 19.472
C16 21.426 21.422 21.424
C17 23.445 23.446 23.446
C18 25.487 25.49 25.489
C19 27.514 27.517 27.516
C20 29.499 29.504 29.502

Sensory Thresholds

As indicated in the literature review, sensory thresholds can be determined by

combining the method of constant stimuli with the forced-choice discrimination methods.

Among the discrimination methods, n-AFC, especially 2-AFC and 3-AFC are most

commonly applied methods for sensory threshold detection (Baird, 1997). The benefits of

n-AFCs over the overall discrimination tests such as triangle tests and duo-trio is that

they are more powerful and generally more sensitive due to their simplicity for subjects'

performance, resulting in a lower level of the estimation threshold (Ennis, 1990; Lawless

and Heymann, 1998). Yet the power and the sensitivity of the n-AFCs are gained through






48


the directional comparison: the attribute that differs in the two samples under comparison is well defined and the panelists are well trained to compare the intensity of the attribute between samples. The benefits of n-AFCs are compromised in our case since the spice mixtures under comparison were different in the intensities of three attributes: the odor of basil, the odor of cinnamon and the odor of garlic. It is also difficult, if not impossible, to carry the n-AFCs methods in this study since they require intensive panel training to familiarize them with the three attributes. Triangular tests, which have a slightly higher power than duo-trio tests, were then selected in this study to assess the difference thresholds of spice mixtures.

As mentioned in Chapter 2, the perception behavior for an odor mixture can be very complicated. It is difficult to define the difference/similarity between two spice mixture with different compositions. For simplicity, here the difference of any two mixtures was defined as the absolute average fraction difference, e.g., for mixture {x11, X12, x13} and {x21, x22, x23}, the difference is calculated by formula 4.2.


difference = lXII 21112 2211 3 2 x100% (4.2)
3

To test similarity, triangle tests require at least 50 panelists (Meilgaard et al., 1999). The number of panelists required in each discrimination test limited the total number of the triangle tests performed. In order to effectively assess the range of difference thresholds over the whole simplex space, the mixture samples under comparison should be representative over that whole space. Additionally, it was reasonable to asses the difference thresholds around the e-nose testing points since the objective of the sensory tests was to determine if the e-nose could predict the mixing fractions with an error less than human perception and the e-nose prediction errors were based on the testing points.






49


Based on the criteria mentioned above, the comparison points were selected. These points are illustrated in Figure 4-5. The spice mixtures used for testing in e-nose experiment were IT through 5T, as listed in Table 4-3, and 1T-1, 3T-1, 5T-1, 4T-1, 4T-2 and 4T-3 were the points being compared to their corresponding sensory testing points. The difference between each of the two points under comparison was 6%, which had been shown to be difficult to discriminate in the preliminary study using 8 panelists. The mixing fractions of all the comparison points in Figure 4-5 are listed in Table 4-6.

Garlic


0.06 Difference from
. 1T-1 Testing Points
1T /

Testing Points of
e-nose experiments

o Sensory Comparison Points

4T-2

, 4T-3 \ 5T- 1
3T- 1 4 4T
5T ~
3T 4T- 1

Basil Cinnamon

Figure 4-5. The comparison points for sensory threshold testing, with IT through 5T
were the spice mixtures used for testing in e-nose experiment and IT-1, 3T-1,
5T-1, 4T-1, 4T-2 and 4T-3 were 0.06 absolute average fraction difference
from corresponding testing points

Spice mixtures were formulated in the weight fractions listed in Table 4-6. Each experimental sample contained I-gram formulated spice mixture, which was stored in a 22ml amber glass vial with aluminum-lined screw cap (Supelco, Bellefonte, PA). For






50


each of the vials, the transparent part was well covered by aluminum foil to avoid giving any clue to the panelists about the spice compositions based on their visual characters. Table 4-6. The mixing fractions of the six pairs of spice mixtures compared by panelists
Testing Compared X1 X2 X3
Points with
4T 0.413 0.366 0.221 4T-1 0.413 0.456 0.131
4T 0.413 0.366 0.221 4T-2 0.323 0.366 0.311
4T 0.413 0.366 0.221 4T-3 0.503 0.276 0.221
1T 0.218 0.015 0.767 IT-I 0.128 0.105 0.767
3T 0.829 0.047 0.124 3T-1 0.739 0.092 0.169
5T 0.052 0.777 0.171 5T-1 0.097 0.687 0.216

For the triangle tests, panelists were instructed to open the cap just before sniffing the sample and then close the cap once finished. The samples should be sniffed in an order of from left to right. Panelists was not allowed to reopen the cap to sniff when not sure which sample was the odd one to make sure that panelists sniffed only the saturated headspace of spice mixtures. No training was given and panelists were informed that they were evaluating spice mixtures of basil, cinnamon and garlic, and were presented samples that differed only in compositions of these three spices. The interval between two successive triangle tests were set as 2 minutes and panelists were asked to take deep breaths during the interval to reduce carry-over and adaptation effects. In order to avoid general fatigue of the subjects, there were only two sections of triangular tests carried per day. Panelists were healthy adults selected from campus ranging in age from 20 to 53 and had no self-reported problem in their sense of smell. A total of 50 panelists performed the triangle tests on each of three consecutive days. In the first day, the tests were performed to compare the pair of mixture IT and IT-I and that of 3T and 3T-1. The pair of 5T and 5T-1 and the pair of 4T and 4T-1 were compared in the following day. The comparison between 4T and 4T-2, and between 4T and 4T-3 were carried out in the third day.






51


There are a total of six possible sequences to present the odd sample (AAB, ABA, BAA, BBA, BAB and ABB). Compusensor@ (Compusense Inc., Guelph, Ontario, Canada) was used to assign the six sequences to the two samples under comparison. Compusensor@ also was used to generated random numbers to represent each sample and helped in collecting and recording data.

Data Analysis

Electronic Nose

Three neural network methods were used to analyze data obtained from the e-nose experiment: Multilayer Perceptron (MLP), MLP with principal component scores as inputs (PCA-MLP) and Time-delay Neural Networks (TDNN). These are described in detail below.

MLP

Twelve inputs and three outputs were used in MLP. Each input corresponded to one of the twelve e-nose sensors' response. The sensors' responses used here were the instantaneous sensor responses at 4 minutes. Each of the three outputs indicated the mass fraction of a spice (basil, cinnamon or garlic) in the mixture. The values of desired outputs were set as the experimental fractions of the corresponding spice mixture, as those listed in Tables 4-2 and 4-3. The activation function of the output layer was a softmax, which can restrict each output to a value between 0 and 1 and makes them sum up to 1. By using softmax activation function in the output layer, the internal constraint of the mixture experiments (the mass fractions of three spices must add up to one) was implemented.

The training/cross-validation data set was partitioned into a training set and a crossvalidation set: 2 replicates from each spice mixture were randomly selected to build the






52


cross-validation set, and the remaining 6 replicates were used as the training set. The order of replicates in the training set was randomized before presenting to a neural network to avoid training bias.

The selection of the suitable MLP topology (the number of hidden layers and the number of neurons in each hidden layer) and parameters (weights and biases) was based on the performance of that topology/parameters toward the testing data. The smaller the error of the testing results generated from a set of MLP topology/parameters, the better that topology-parameter combination. The selection of topology/parameters proceeded as follows: First, a network topology was selected. This network was trained at least 10 times with random initial conditions based on how variable the testing results were, and the best testing result was recorded for that topology. Then, the network topology was changed, and the same training method was applied. This iterative method was continued until performance on testing data could not be improved. By the above procedure, the optimal topology-parameter combination of MLP was determined and the corresponding testing result was recorded as the performance of the network.

In this study, the testing result of a neural network topology/parameters was

obtained by averaging the five replicates of e-nose's responses corresponding to each testing point and then using these average responses as the inputs of that neural network. The error of a testing result was measured by mean square error (MSE). PCA-MLP

This method was similar to that of MLP except that PCA (principal component

analysis) was first applied to instantaneous sensor responses to reduce dimentionality of the original data. This resulted in a reduced number of inputs into the neural network. Then, principal component scores instead of original sensor responses were used as






53


inputs to the neural network, and the number of inputs was equal to the number of principal components selected.

TDNN

This method was different from that of MLP in the following aspects:

- For the inputs, the time series sensor responses from 0 to 4 minutes instead

of the instantaneous sensor responses at 4 minutes were used.

- There were embedded local memories (tapped delay line) in both input and

hidden layers, which made the NN sensitive to the sequence of information

in inputs.

- The three outputs that indicated the mass fraction of basil, cinnamon and

garlic respectively were set as the experimental fractions of the

corresponding spice in the mixture, regardless of the time at which the

sensor responses were collected.

- Since the outputs of the NN were also time series data, the predicted

mixture compositions for each spice were based on the average of the

corresponding outputs over time.

The randomized selecting and ordering of replicates was accomplished in Microsoft Excel spreadsheets, using either the random/sort function or customized VBA codes. PCA was performed using SAS@ (SAS Institute Inc., Cary, NC), and neural networks were implemented using NeuroSolutions@ (NeuroDimension Inc., Gainesville, FL). Gas Chromatography

Since the concentration of decane (the added standard) in every extract was

controlled (1Il/ml), the relative amount of a volatile in an extract could be determined by dividing the corresponding peak area to the peak area of decane in the same injection.






54


This adjustment could also eliminate the variability associated with injection and other GC operations. The GC analysis was then based on the relative amounts of volatiles within extracts.

Two approaches were used to analyze the GC data. One was based on the relative amounts of the most abundant unique volatile of each spice, while another was based on the five most abundant volatiles of each spice. Both of these two methods assumed that the relative amount of an extracted volatile was proportional to the amount of the volatile in the spice mixture. The calculation of both methods was constrained by the fact that the fractions of three spices in a mix add up to 1. The two methods were referred later as "single volatile method" and "five-volatiles method." Single volatile method

As illustrated in the Figure 4-6, this method started with the identification of the

most abundant unique volatile in each of the three spices. These volatiles were labeled by their corresponding retention time tb, t, and tg for basil, cinnamon and garlic, respectively. For all of the 3 replicates x 2 injections carried for the pure basil samples, the relative volatile amounts at tb were averaged and recorded as ABP. The same procedures were performed for the pure cinnamon and garlic samples and resulted in the value of Acp and AGP. For each of the testing spice mixtures, the relative volatile amounts at tb, t, and tg that were obtained from the 2 replicates x 2 injections were averaged and recorded as AB, Ac and AG, respectively. The mixing fractions of basil, cinnamon and garlic, symbolized byX, B and X respectively, were calculated using the equations listed at the right of Figure 4-6. It should be noted that the mixing fractions of basil, cinnamon and garlic should equal to AB/A Bp, AC /ACp and AG/AGP respectively based on the assumption that







55


the relative amount of a volatile was proportional to the amount of the volatile in the spice mixture. These estimated mixing fractions were divided by the common factor (the sum of the estimated fractions) to satisfy the constraint that all three fractions should add up to 1.


From Training Points (Pure spice):

Relative amount of the most (3 replicate x 2 injection)
abundant volatile in Basil Average 3 A BP at retention time tb

Relative amount of the most (3 replicate x 2 injection)
abundant volatile in Cinnamon Average 0 A cp at retention time t.

Relative amount of the most (3 replicate x 2 injection)
abundant volatie in Garlic Average 3- AGp at retention time t9



From Testing points (Spice Mixture):

Relative amount of (2 replicate x 2 injection)
volatile at t b Average a A

Relative amount of (2 replicate x 2 injection)
volatile at t, Average Ac

Relative amount of (2 replicate x 2 injection)
volatile at t Average AG3


BP
B B
+ +
A~p ACp A0p

Ac
C Ac?




ABP -CP -GP
+C +


Figure 4-6. The single volatile method for GC data analysis.X B If. and (;

represented the estimated mixture compositions for basil, cinnamon and
garlic, respectively


Five-volatiles method

As shown in Figure 4-7, this method was similar to the single volatile method


except that the sum of the relative amounts of five most abundant unique volatiles were used to estimate the factions instead of the relative amount of the single most abundant unique volatile from each spice.


I/







56


From Training Points (Pure spice):

Sum of relative amounts (3 replicate x 2 injection) A BP at retention time of five most abundant Average tbl, t1 t* tb* and ts
volatiles in BasilAvrg bt2 3.tant5

Sum of relative amounts (3 replicate x 2 injection) A cP at retention time
of five most abundant C ,
volatiles in Cinnamon Average tci. t,2 tc3, tc4, and tc5

Sum of relative amounts (3 replicate x 2 injection) A Gp a retention time
of five most abundant
volatiles in Garlic Average to. tg, tS3 tot, and tg5



From Testing points (Spice Mixture):

Sum of volatile relative amount (2 replicate x 2 injection)
at tbie b tb, stand tb Average As

Sum of volatile relative amount (2 replicate x 2 injection)
at te, t, t., te4, and tC5 Average A C

Sum of volatile relative amount (2 replicate x 2 injection)
at t g, t t9*tL, and tg5 Average AG


I


AB
A8 =
4B + AC + g ABP ACp A0p

AG


ABP, A01 A01,

A0

AB + AC +A
--'BP ACP Aop


Figure 4-7. The five-volatiles method for GC data analysis. XB aC d XG

represented the estimated mixture compositions for basil, cinnamon and
garlic, respectively


Sensory Thresholds


The estimation of difference thresholds was based on the criterion that the


probability of correct responses should be 50% correct above chance when the difference of the two samples is at the threshold. The observed probability of correct responses from the triangle tests were converted to percent correct above chance by using formula 4-3.



P. = Pos cheX 100 (4.3)
100 - P e


where P(corr) equal to the percentage of correct response above the chance to a stimulus, P(obs) equal to the percentage of correct response given by subjects to that stimulus and P(chance) equal to the percentage of correct response expected on the basis of chance alone.






57


For triangle tests, the percentage of correct responses observed should be 66.7% to reach "50% correct above chance" since the guessing chance of the tests is 33.3%. In other words, if the triangle tests show that at a level of difference, the observed percentage of correct responses was lower than 66.7%, that level of difference could not be perceived by human subjects at a 0.50 probability.

















CHAPTER 5
RESULT AND DISSCUSION

Electronic Nose

Raw Data

Appendix A lists the instantaneous e-nose data obtained for the purpose of training

and testing. The time series data were too large to be listed. The original data are

available in a CD and can be obtained from Dr. Murat 0. Balaban at the University of

Florida. Appendix B shows the organization and the path of these raw time series data

files in the CD. A sample of the time series data is illustrated in Figure 5-1. The data were

obtained as a replicate of mixture #12 (basil: cinnamon: garlic = 0.167:0.667:0.167).

Fractions of each spice
Responses from 12 Sensors within the mixture


S1 S2 S3 ... S12 Basil Cinnamon Garlic
0 0 0 ... 0 0.167 0.667 0.167
0 0.01 0.01 ... 0 0.167 0.667 0.167
0 0.04 0.04 ... 0 0.167 0.667 0.167
0 0.1 0.09 ... 0.02 0.167 0.667 0.167
0 0.2 0.17 ... 0.04 0.167 0.667 0.167
0 0.33 0.29 ... 0.08 0.167 0.667 0.167
A total of 240 ines 0 0.5 0.43 ... 0.14 0.167 0.667 0.167
of data over a period 0 0 7 0.6 ... 0.21 0.167 0.667 0.167
of4 minutes per 0 0.92 0.79 ... 0.29 0.167 0.667 0.167
replicate per mixture. 0 1.12 0.97 ... 0.39 0.167 0.667 0.167
0 1.32 1.15 ... 0.49 0.167 0.667 0.167
0.01 1.49 1.33 ... 0.59 0.167 0.667 0.167
0.01 1.64 1.48 ... 0.69 0.167 0.667 0.167
0.01 1.77 1.62 ... 0.78 0.167 0.667 0.167

1.05 6.72 6.06 ... 3.55 0.167 0.667 0.167
1.05 6.72 6.06 ... 3.55 0.167 0.667 0.167


Figure 5-1. A sample of time series data obtained


58






59


Data Analysis Results Using MLP

As mentioned in Chapters 1 and 4, MLP was applied to analyze the e-nose

instantaneous data. One hidden-layer was selected for the MLP since two-hidden-layers are necessary only when the functions to be modeled are discontinuous and here sensors' responses toward mixtures were expected to be continuous.

The number of neurons in the hidden layer was first set at 3, and the smallest

prediction error of this topology on the testing set was 0.0084 (MSE). When the number of neurons increased to 8, the smallest prediction error obtained from this topology was

0.0059 (MSE). With the number of neurons in the hidden layer being 4 and 5, the smallest prediction errors obtained were 0.0051 and 0.0050 in MSE, respectively.

Based on above results and the criteria of network size selection (the neural

network should be sufficiently large to solve the problem, but not larger), the optimal topology of MLP was determined to be one hidden layer with 12 inputs/3 outputs and 4 hidden neurons (Figure 5-2). The activation function of the hidden layer was sigmoid and that of output layer was softmax. This optimal MLP predicted the mixture compositions with an error of 0.0051 (MSE). Table 5-1 lists the prediction values for each spice at each testing point.

Table 5-1. The experimental and predicted mass fractions of spice mixtures predicted by
the optimal performance of MLP
Testing Experiental Weight Fraction Predicted Weight Fraction Average Absolute Points Basil Cinnamon Garlic Basil Cinnamon Garlic Prediction error IT 0.218 0.015 0.767 0.226 0.099 0.675 0.061
2T 0.114 0.61 0.277 0.116 0.625 0.259 0.012
3T 0.829 0.047 0.124 0.789 0.03 0.180 0.038
4T 0.413 0.366 0.221 0.349 0.407 0.244 0.043
5T 0.052 0.777 0.171 0.049 0.926 0.024 0.100
Mean Square Error (MSE): 0.0051






60


Sensor 1 Sensor 2 Sensor 3 Sensor 4

Sensor5 f f
Sensor 6
Z f Cinnamon
Sensor 7
Z f Garlic
Sensor 8 Sensor 9

Sensor10

Sensor 11
Sensor12


Figure 5-2. The optimal topology of MLP Data Analysis Results Using PCA-MLP

Having the potential to perform better than MLP, PCA-MLP structure was also

applied to analyze the instantaneous e-nose data. It is important for this method to select suitable principal component scores as the inputs to a neural network. Khattree and Naik (2000) indicated that the amount of information contained in a multivariate data set can be measured in terms of total variance. Therefore, the selection of principal components is based on the amount of variance that each principal component preserves. The proportion and cumulative proportion of total variance explained by principal component

1 through 10 were sequentially listed in Table 5-2. It could be found that first two principal components explain more than 99% of total variance. Based on this, only the first two principal component scores were selected as neural network inputs.






61


Table 5-2. The proportion and cumulative proportion of total variance explained by
principal component 1 through 10
Principal Proportion of Total Cumulative Proportion of
Component Variance Explained Total Variance Explained
1 0.9850 0.9850
2 0.0066 0.9916
3 0.0032 0.9948
4 0.0017 0.9964
5 0.0013 0.9978
6 0.0007 0.9985
7 0.0006 0.9991
8 0.0004 0.9995
9 0.0002 0.9997
10 0.0001 0.9998

Due to the same considerations as those in the MLP method, one hidden layer was selected. Similar to the training procedure in the MLP method, the number of neurons in the hidden layer was first set at 3 and resulted in 0.0053 (MSE) as the smallest prediction error of this topology. Then, the number of neurons was increased to 8, and resulted in

0.0039 MSE as the smallest prediction error. The smallest prediction errors obtained were

0.0041 and 0.0043 in MSE for the networks with 4 and 5 neurons in the hidden layer, respectively. Therefore, the optimal topology of PCA-MLP was determined to be 2 inputs/3 outputs and 1 hidden layer with 4 hidden neurons (Figure 5-3), and the prediction values for each spice at each testing point are listed in Table 5-3. The activation function of the hidden layer was sigmoid and that of output layer was softmax. This optimal topology predicted the mixture compositions with an error of 0.0041 (MSE). Data Analysis Results Using TDNN

TDNN was applied here to analyze the e-nose time series data. One hidden layer was used here for the same reasons as in MLP and PCA-MLP. It was experimentally found that the TDNN's performance (measured by MSE resulted from the testing set)






62


was improved by averaging every four seconds' responses of e-nose sensors. This may be because the averaging procedure removed part of the noise from the time series data. The best performance of TDNN was obtained by using the time series sensors' responses collected between 2 and 4 minutes. This indicated that the time related information was more abundant in the e-nose time series data form 2 to 4 minutes than that contained in the time series data from 0 to 2 minutes. The amounts of time information in the data from 0 to 2 minutes was small enough so that it could not counter the negative effects on the TDNN performance resulting from the increasing number of parameters, which was necessary to handle data in a longer time window. Table 5-3. The experimental and predicted mass fractions of spice mixtures predicted by
the optimal performance of PCA-MLP
Testing Experimental Weight Fraction Predicted Weight Fraction Average Absolute Points Basil Cinnamon Garlic Basil Cinnamon Garlic Prediction error IT 0.218 0.015 0.767 0.162 0.05 0.787 0.056
2T 0.114 0.61 0.277 0.08 0.629 0.291 0.034
3T 0.829 0.047 0.124 0.841 0.035 0.124 0.012
4T 0.413 0.366 0.221 0.28 0.498 0.222 0.133
ST 0.052 0.777 0.171 0.04 0.869 0.092 0.012
Mean Square Error (MSE): 0.0041

For the one-hidden layer TDNN, except for the weights and biases that can be automatically searched by back-propagation algorithm, there were several other parameters needed to be physically adjusted: the number of neurons in hidden layer, and the depth (number of the taps) and resolution (each tap delays the signal how many units) of each memory structure. To search for the best combination of the above parameters, just one of these parameters was adjusted in every step until the best performance was found. It was found that the best TDNN performance predicted the mixture compositions with an error of 0.0035 (MSE). The prediction values for each spice at each testing point






63


are listed in Table 5-4. The optimal architecture of TDNN, as shown in Figure 5-4, were described as follow:

- 12 inputs/3 outputs and 1 hidden layer with 10 neurons;

- The activation function of the hidden layer was tanh.

- The activation function of the output layer was softmax.

- The number of taps in the input layer was 6 with 1 tap delay.

- The number of taps in hidden layer was 3 with 1 tap delay.

Table 5-4. The experimental and predicted mass fractions of spice mixtures predicted by
the optimal performance of TDNN
Testing Experimental Weight Fraction Predicted Weight Fraction Average Absolute Points Basil Cinnamon Garlic Basil Cinnamon Garlic Prediction error IT 0.218 0.015 0.767 0.192 0.129 0.679 0.076
2T 0.114 0.61 0.277 0.092 0.585 0.323 0.031
3T 0.829 0.047 0.124 0.789 0.048 0.181 0.033
4T 0.413 0.366 0.221 0.323 0.418 0.258 0.06
ST 0.052 0.777 0.171 0.052 0.839 0.109 0.041
Mean Square Error (MSE): 0.0035


p f
Z Basil
PC1I D . f
Cinnamon
PC2
I f Garlic
I_ f


Figure 5-3. The optimal topology of PCA-MLP Discussion

Based on the prediction MSE values, it could be concluded that in this study PCAMLP performed better than MLP, and TDNN performed better than PCA-MLP in predicting mixture compositions. Just like any other modeling structure, the generalization ability of a MLP model will deteriorate with increasing number of






64


parameters, given the same amount of inputs / outputs information (Principe et al., 1999). PCA extracted most of the inputs' information (over 99%) in a reduced dimensional format (from 12 to 2 dimensions). Since the number of parameters necessary in a MLP structure was directly correlated with the dimensions of its inputs, the preprocessing of input information by PCA significantly decreased the number of parameters necessary in the MLP structure. That resulted in improved generalization ability and a better performance of the PCA-MLP structure over MLP. It was not surprising that TDNN provided the smallest prediction error since it captured the additional information about time related differences in sensors' responses. This was not available in the instantaneous data used in MLP and PCA-MLP.

Figure 5-5 shows the distribution of the prediction points, including those from

MLP, MLP-PCA and TDNN, in the simplex space. It could be noted that the predictions from all the three methods skewed toward cinnamon compared with the experimental spice mixture points. In other words, all these methods tended to predict higher cinnamon fractions within a mixture. Considering that cinnamon provided the strongest odor impact among the three spices, this skewed prediction might be due to the following two reasons: first, the e-nose sensors might be overwhelmed by the cinnamon volatiles during the "sniffing" process of e-nose operations; second, there might be cinnamon odor residues in the sampling or sensor chamber despite the e-nose cleaning procedure.

To predict unknown mixture compositions, the training points for prediction model building shall cover the whole mixture space and should be evenly distributed within that space. It means that the model developed based on the training data can't be used to predict an unknown mixture whose compositions are not within the training mixture








65


space. Evenly distributed training points will ensure a fair estimation of the parameters of the prediction model. Sensor 1


Z-1
Z-1
Z-1
Z--1




Sensor 2

Z-1 1 f Basil












Figure 54 Th opialtployofT1
Z-1 Z- 7 Cinnamon

2-1 f Gadlic





Sensor 12
Z-1

Z-1

Z-
Z-1




Figure 5-4. The optimal topology of TDNN


There are two main factors which influence how many training points will be


necessary for prediction model building: the number of components being mixed and the complexity of the sensors' response to the mixture. The more components being mixed and the more complex the sensors' response, the more training points will be necessary. The number of samples necessary for each training point depends on the variability of the e-nose response to an identical mixture composition, the differences among the mixtures with different compositions, and the target prediction accuracy of the mixture compositions.






66


Garlic

* Experimental Points (Desired)

Predicted Points from MLP
* using sensors' responses
as Inputs

Predicted Points from MLP
using PCA scores as Inputs

Predicted Points from TDNN
using time series data



U
*ILE



Basil Cinnamon

Figure 5-5. The distribution of the prediction points, including those from e-nose MLP,
MLP-PCA and TDNN analysis, in the simplex space

For a ternary mixture system, the prediction of compositions is actually a twodimensional problem due to the built-in constraint of compositional data. Ideally, to solve this two dimensional problem, the e-nose response data should show the following structure: the first two principal components should explain more than 90% of total variance yet the second principal component should account for a significant proportion of total variability, say, more than 20%. Accordingly, if a mixture with four components is under investigation, the first three principal components are expected to explain more than 90% of total variance yet the third principal component accounts for a significant proportion of total variance. The e-nose response data obtained in this study did not have these ideal properties. Among the selected first two principal components used in PCAMLP, the first principal component explained more than 98% of total variance (Table 5-






67


2), resulting in the second principal component having a high noise to signal ratio. The ideal structure of the data will not only improve the prediction ability of the PCA-MLP structure, but also other prediction models applied, which include both the multivariate statistical and neural network methods. A proper sensor array design/selection will make the sensors' responses approach ideal conditions.

It should be noted that neural networks have the advantage over the traditional

statistical methods only when the investigated data structures are highly complex. In case the investigated data structures are simple, alternative statistical methods such as principal component regression or partial least squares may perform better than neural networks in a quantitative e-nose study.

Gas Chromatography

Volatile Components of Spices

Table 5-5 shows the average relative amounts of all the volatiles identified in each of the three spices and their corresponding retention times. The selection of the five most abundant unique volatiles in each spice was based on the following criteria: first, in order to be identified as a unique volatile of a certain spice, the ratio of the volatile relative amount in that particular spice to those in the two other spices should be equal or larger than 15; second, the relative amount of the volatile should be as large as possible in that particular spice.

The retention times of the five most abundant unique volatiles in each of the three spices were recorded and listed in the Table 5-6. In this table, the star signs under the columns heading with spice names indicated the membership of these volatiles. The time window of each of these volatiles identified by GC-FID was its corresponding retention time plus/minus 0.05 minutes. Table 5-6 also lists the calculated retention index of all the






68


volatiles and the literature values of some of the volatiles (Yu et al., 1989; Yu et al.,

1993; Vernin et al., 1994; Yu et al., 1994; Adams, 1995; Kim et al., 1995; Marotti et al.,

1996; Miller et al., 1996; Lachowicz et al., 1997; Antonelli et al., 1998; Acree and Am,

2001).

Table 5-5. The average relative amounts of all volatiles identified in each of the three
spices and their corresponding retention times

Volatile # Retention Time Average volatile relative amounts
(minutes) Basil Cinnamon Garlic
1 7.537 0.002 0 0.034
2 9.795 0.006 0.227 0
3 11.295 0.041 0 0
4 11.858 0 0 0.062
5 12.178 0 0 0.406
6 12.527 0.641 0 0
7 13.285 0 0 0.232
8 14.281 0.547 0 0
9 14.672 0 0.364 0.014
10 15.518 0.043 25.174 0.010
11 16.075 0.072 0.064 0.685
12 16.981 0.292 0 0
13 17.463 0.478 0.989 0
14 18.474 0.242 0 0
15 18.586 0.021 0.332 0
16 19.732 0.003 0.331 0
17 20.178 0.039 0.759 0
18 22.562 0.184 0 0
Sum of relative volatile amounts in 2.610 28.240 1.442
each of the three spices:

The five most abundant unique volatiles in the basil emerged in the minutes 11.295,

12.527, 14.281, 16.981 and 18.474, and were labeled as volatile 2, 5, 7, 11 and 12

respectively. As discussed in Chapter 4, there were two methods to determine the

chemical identity of a certain volatile. One was to compare the calculated retention index

with the value from literature. Another was to add a known chemical standard into the

sample, which resulted in an increased peak whose identity was the same as that of the






69


added chemical standard. The first method was called the retention index method in this chapter, while the second was called the chemical standard method. Volatiles 2, 5 and 11 were identified as 1,8-cineole, linalool and eugenol by both the retention index and the chemical standard methods. Volatile 7 was identified as estangole by the chemical standard method. Volatile 12 was unidentified.

The five most abundant unique volatiles in the cinnamon emerged at 14.672,

15.518, 18.586, 19.732 and 20.178 minutes, and were labeled as volatile 8, 9, 13, 14 and 15, respectively. Volatiles 8, 9 and 13 were identified as geraniol, trans-cinnamaldehyde and cinnamyl acetate by both the retention index and the chemical standard methods. Volatiles 14 and 15 were unidentified.

The five most abundant unique volatiles in the garlic emerged at 7.537, 11.858, 12.178, 13.285 and 16.075 minutes and were labeled as volatile 1, 3, 4, 6 and 10, respectively. Volatiles 4 and 10 were identified as diallyl sufide, and diallyl disulfide by both the retention index and the chemical standard methods. Volatile 6 was identified as methyl allyl disufide by the retention index method. Volatiles 1 and 3 were unidentified.

The raw data obtained from both the pure spices and the spice mixtures are listed in Appendix C.









Table 5-6. The five most abundant unique volatiles identified in the basil, cinnamon and garlic, and the volatiles' corresponding
chemical identities
Volatile Retention Time Basil Cinnamon Garlic Calculated CAS# Chemical Retention Index
No. (minutes) Retention Index Name (Literature)
1 7.537 * 858
2 11.295 * 1036 470-82-6 1,8-cineole 1030
3 11.858 * 1066
4 12.178 * 1083 592-88-1 diallyl sufide 1085
5 12.527 * 1101 78-70-6 Linalool 1100
6 13.285 * 1145 methyl allyl disufide 1149
7 14.281 * 1202 140-67-0 Estragole N/A
8 14.672 * 1226 106-24-1 geraniol 1255
9 15.518 * 1277 14371-10-9 trans-Cinnamaldehyde 1266
10 16.075 * 1310 2179-57-9 diallyl disufide 1320
11 16.981 * 1363 97-53-0 Eugenol 1356
12 18.474 * 1446
13 18.586 * 1452 103-54-8 Cinnamyl acetate 1443
14 19.732 * 1513
15 20.178 * 1536

1 . CAS# refers the CAS Registration Number. CAS (Chemical Abstracts Service) registry is the largest chemical substance
identification system in existence. When there is a new substance shown in the literature, it is assigned a unique CAS Registry
Number. CAS Registry Numbers are used to identify substances without the ambiguity of chemical nomenclature (Anonymous,
2003).


0






71


Table 5-7 listed the average fraction (in percentage) of the volatiles (those

identified in Table 5-6) in their corresponding pure spice extracts. It could be found that

linalool was the most abundant component in the basil's extracts, accounting for 24.6%

of the extracted total volatiles. Trans-cinnamaldehyde and diallyl disulfide were the most

abundant components in the cinnamon and garlic's extracts, respectively. Transcinnamaldehyde accounted almost 90% of the total extracted volatiles from cinnamon.

Diallyl disulfide accounted for nearly 50% of the total extracted volatiles from garlic.

Therefore, the unique volatiles selected for the single volatile method were linalool,

trans-cinnamaldehyde and diallyl disulfide for basil, cinnamon and garlic, respectively.

Table 5-7. The average fraction (in percentage) of the volatiles in their corresponding
pure spice extracts.

Volatile # Retention Time % in basil % in cinnamon % in garlic
(minutes) extracts extracts extracts
1 7.537 0.07 0 2.36
2 11.295 1.56 0 0
3 11.858 0 0 4.28
4 12.178 0 0 28.18
5 12.527 24.59 0 0
6 13.285 0 0 16.08
7 14.281 20.97 0 0
8 14.672 0 1.31 0.98
9 15.518 1.64 89.11 0.70
10 16.075 2.75 0.23 47.42
11 16.981 11.20 0 0
12 18.474 9.25 0 0
13 18.586 0.80 1.18 0
14 19.732 0.10 1.17 0
15 20.178 1.49 2.67 0
Sum of fractions of five most abundant 67.57 95.44 98.31
volatiles in the spice extracts:






72


Single Volatile Method

The single volatile method predicted the mixture compositions with an error of

0.000755 (MSE). The prediction values for each spice at each testing point are listed in

Table 5-8.

Table 5-8. The experimental and predicted mass fractions of spice mixtures predicted by
the single volatile method
Testing Experimental Weight Fraction Predicted Weight Fraction Average Absolute Points Basil Cinnamon Garlic Basil Cinnamon Garlic Prediction error IT 0.218 0.015 0.767 0.233 0.016 0.715 0.023
2T 0.114 0.61 0.277 0.124 0.644 0.232 0.030
3T 0.829 0.047 0.124 0.816 0.063 0.121 0.011
4T 0.413 0.366 0.221 0.441 0.386 0.172 0.032
ST 0.052 0.777 0.171 0.064 0.779 0.157 0.009
Mean Square Error (MSE): 0.000755

Five-volatiles Method

The five-volatiles method predicted the mixture compositions with an error of

0.00156 (MSE). The prediction values for each spice at each testing point are listed in

Table 5-9.

Table 5-9. The experimental and predicted mass fractions of spice mixtures predicted by
the five-volatile method
Testing Experimental Weight Fraction Predicted Weight Fraction Average Absolute Points Basil Cinnamon Garlic Basil Cinnamon Garlic Prediction error IT 0.218 0.015 0.767 0.224 0.018 0.758 0.006
2T 0.114 0.61 0.277 0.113 0.682 0.204 0.049
3T 0.829 0.047 0.124 0.835 0.069 0.096 0.019
4T 0.413 0.366 0.221 0.434 0.407 0.159 0.041
ST 0.052 0.777 0.171 0.060 0.817 0.123 0.032
Mean Square Error (MSE): 0.00156






73


Discussion

Figure 5-6 shows the distribution of the prediction points, including those from

single volatile and five-volatiles methods, in the simplex space. Similar to those from enose analysis, the predictions resulting from the GC methods tended to predict more cinnamon and less garlic in a mixture. The residues in the extraction-distillation apparatus might account for the prediction skew. The data in Table 5-5 demonstrated the existence of the residues. It was found from the table that there were minor amounts of the unique components from cinnamon and garlic in the other two spices. These residues were difficult to be avoided since the complex shape of the extraction-distillation apparatus made it very hard to be cleaned. Table 5-5 also showed that the volatile concentration in cinnamon extracts was much higher than that in basil and garlic (cinnamon:basil:garlic 55:5:3). The abundance of cinnamon volatiles within the extraction system also added the possibility of absorbed cinnamon volatile residues.

It had been mentioned in the Chapter 4 that the two GC methods were based on the assumption that the relative amount of a volatile in extracts was proportional to the amount of the volatile in the spice mixture. To test this assumption in the single volatile method, the K value was defined here as the ratio of the relative amount of a volatile to the weight of the spice from which the volatile was extracted. If the above assumption held, the K values should be constant for basil, for cinnamon and for garlic regardless of the weight of the corresponding spice in the mixture extracted. The K values of the most abundant volatile in basil, cinnamon and garlic at different weights of the corresponding spice were calculated individually and are illustrated in Figures 5-7, 5-8 and 5-9, respectively. It should be noted that since the total weight of the spice sample for extraction was fixed (5 grams), the trend of the K values versus the corresponding spice






74


weights was the same as that of the K values versus the mixing fraction of the corresponding spice.


Garlic

* Experimen


, Predicted single vola + Predicted five-volatile






0 +
M+ it
++


tal Points (Desired)


Points from tile method Points from method


Basil Cinnamon

Figure 5-6. The distribution of the prediction points, including those from GC single
volatile and five-volatile methods, in the simplex space

It could be found from Figure 5-7 and Figure 5-8 that the assumption of constant K values was valid for the most abundant volatile in basil and in cinnamon except at very low weights. As discussed before, minor residues existed in the extraction-distillation apparatus. When the absolute amount of a volatile decreased, the ratio of the concentration of residue volatile to that from extraction increased, resulting in increased K value. For the most abundant volatile in garlic, the assumption held with exception at the garlic weight around 1 gram.






75


0
0
(A


0.17 0.16 0.15 0.14

0.13

0.12 n 14


J. I


0 1 2 3 4 5 6
The weights of basil in the mixture (gram)


Figure 5-7. The K values of the most abundant basil volatile at different weights of basil
in the corresponding spice mixture


C0

wE

> C


8

7

6

5

4


0 1 2 3 4 5 6
The weights of cinnamon in the mixture (gram)


Figure 5-8. The K values of the most abundant cinnamon volatile at different weights of cinnamon in the corresponding spice mixture


*




* + -


_ _ *





*-

















+.-- .

-.


0.17 0.15 0.13 0.11 0.09 0.07 0.05


The weights of garlic in the mixture (gram)


Figure 5-9. The K values of the most abundant garlic volatile at different weights of
garlic in the corresponding spice mixture


0 0.5 cc
o 0.45 i$ 0.4

0.35
0

S0.3 S0.25


0


i


3:. .4


1


2


3


5


6


4


The weights of basil in the mixture (gram)


Figure 5-10. The K values of the five most abundant unique basil volatiles at different
weights of basil in the corresponding spice mixture


76


'

0 U)


$r


1


*


0


2


3


4


5


6


-




-






77


To test the same assumption of the five-volatiles method, the K value was then defined as the ratio of the relative amount of a group of volatiles, here the five most abundant unique volatiles of the corresponding spice, to the weight of that spice. The K values of basil, cinnamon and garlic were calculated and illustrated in Figures 5-10, 5-11 and 5-12, respectively.

Similar to that of the single volatile method, Figures 5-10 and 5-11 show that the assumption of constant K values held for basil and for cinnamon except at very low weights. However, for garlic there was an apparent trend of increasing K values with the weight of garlic within the spice mixture (Figure 5-12). This may due to the absorption of garlic volatiles into the matrices of the other two spices. With the increasing amount of the other two spices, more garlic volatiles were absorbed, and thus less garlic volatiles were extracted by solvents. This resulted in increased K values with the garlic weight. By comparing Figure 5-12 to Figure 5-9, it could be found that the increasing trend was not apparent for the most abundant volatile in garlic (diallyl disulfide). It may be speculated that the absorption rates of the other volatiles in garlic were higher than that of diallyl disulfide.

Theoretically, the five-volatiles method should perform better than the single volatile method since the calculation was multivariate based instead of univariate. However, based on the resulted prediction MSE, the single volatile method provided better prediction than the five-volatiles method. The MSE of the five-volatilse method was twice as big as that of the single volatile method. As discussed above, the assumption of a constant K value was the calculation base of both methods. For garlic, this






78


assumption roughly held for the single volatile method but was not valid for the fivevolatiles method. This might explain the better performance of the single volatile method.


0
0
(U
0
0

0 "C
4.' 'a
0
0
S


9

8 7-


5

4

3


0


1


2


3


4


5


6


The weights of cinnamon in the mixture (gram)


Figure 5-11. The K values of the five most abundant unique cinnamon volatiles at
different weights of cinnamon in the corresponding spice mixture


_0
0




(U


0.35 0.3 0.25 0.2

0.15 0.1


0


- -$ - - - y= 0.0243x + 0.165 R2 = 0.7839


1


2


3


4


5


6


The weights of garlic in the mixture (gram)


Figure 5-12. The K values of the five most abundant unique garlic volatiles at different
weights of garlic in the corresponding spice mixture






79


Sensory Thresholds

The raw data obtained from the triangle tests, which included the order of samples presented and the random code assigned to each sample and the panelists' responses, are listed in Appendix D. Forty-two panelists came on three consecutive days to perform the triangle tests. Five other panelists performed the tests in two days and the rest of the fourteen panelists performed the tests for one day.

Table 5-10 lists the number of panelists correctly recognizing the odd sample for each of the six pairs of comparisons. The correct rates and the correct above chance rates were also calculated and listed in the table. It could be found that except for the comparison between IT and 1T-1, the correct rates were at or lower than 33.3%, which is the guessing rate of a triangle test. Table 5-11 lists the number of correct responses necessary for the triangular tests to establish a significant difference between the two samples under comparison when the total number of the tests was 50 (Meilgaard et al., 1999). Based on these numbers, it could be concluded that except for IT and IT-1, the five pairs of spice mixtures were perceived as similar by the panelists. Table 5-10. The results of the triangle tests for each of the six pairs of comparisons
Compared Between Correct Response Correct Rate Correct above chance (Out of 50 tests)
4T 4T-1 17 34% 1%
4T 4T-2 16 32% -2%
4T 4T-3 12 24% -14%
1T iT-i 30 60% 40%
3T 3T-1 12 24% -14%
5T 5T-i 14 28% -8%

The difference threshold is the minimum difference required between two stimuli that will elicit a perceived difference with a specific probability. As mentioned in Chapter 2, the specific probability was regularly set at 50%. Table 5-10 shows that even for the






80


comparison between IT and IT-1, the correct-above-chance was lower than 50%. Therefore, it could be concluded that the 6% difference as defined in Chapter 4 was below the difference threshold and was not big enough to elicit a perceived difference with a probability of 50%.

Table 5-11. The number of correct responses necessary for the triangular tests to establish
a significant difference between the two samples under comparison when the
total number of the tests is 50
Significance Level 10% 5% 1% 0.1%
No. of Correct Responses 22 23 26 28

It should be noted that the number of correct responses for the comparison between 1T and IT-I was almost twice as big as that from the rest of the five comparisons. Figure 4-6 illustrates the distribution of the points under comparison in the simplex space. It was found that the difference between 1T and other testing points (3T through 5T) was that IT was very close to the boundary of the simplex space, which indicated that the content of one of the three spices (cinnamon) in the mixture was very limited. Therefore, the task of discriminating between IT and IT-I was similar to the task of comparing a twocomponent mixture with a three-component mixture. This was a much easier task compared with discriminating two three-component mixtures. The significant performance difference between the pair of 1T and IT-I and others could also be explained by Weber's Law. Weber's Law states that the difference thresholds increase in proportion to the background intensity, e.g., the perception is more sensitive at low background intensity. Here the background intensity could be taken as the cinnamon fraction within the mixture. When the cinnamon fraction was low enough, the difference between the spice mixtures was easy to discriminate.






81


As mentioned in Chapter 4, on each of three consecutive days there were two

sessions of triangle tests. Table 5-12 compares the panelists' performance between the tests carried first in that day and those carried second. The performance on the pair of 1T and IT-I was not considered in this comparison due to the unique characteristics of this comparison as discussed above. It could be found that there was no significant difference of panelists' performance (correct rate) between the tests carried first and those carried second. This indicates that the 2 minutes interval between two successive triangle tests was enough to avoid sensory adaptation or other negative effects that might compromise the discrimination ability of panelists.

Table 5-12. The comparison of panelists' performance between the tests carried first in
that day and those carried second
Compared Between Correct Response Correct Rate Order of the tests carried in (Out of 50 tests) the corresponding day
IT IT-1 30 60% 1
5T 5T-I 14 28% 1
4T 4T-2 16 32% 1
3T 3T-1 12 24% 2
4T 4T-1 17 34% 2
4T 4T-3 12 24% 2

Comparison between E-nose and GC/Sensory Methods

Figure 5-13 illustrates the prediction accuracy of the three neural network methods by comparing them with the estimated sensory thresholds. It could be found that except for a few points, the predictions from all three NN methods were within a range of 6% difference from the real values. Since the 6% difference had been demonstrated through the sensory analysis to be lower than the sensory threshold of human subjects, especially in the central area of the simplex space, the prediction accuracies from all of the three enose data analysis methods were acceptable.






82


Garlic

.6% Difference from Desired Points (below the sensory thresholds)

1T 1T-1 * Desired Points

0 Sensory Comparison Points

Predicted Points from MLP using senors' reponses as inputs

+ Predicted Pointd from MLP using PCA scores as inputs
4T-2 Predicted Pointed from TDNN using
2T e+ time-series data
4T-3 4T. T-1
3T-1 K
5T
3T 4T-1

Basil Cinnamon

Figure 5-13. Comparison between the prediction error of electronic nose methods and the
sensory threshold of human subjects

Table 5-13 compares the prediction accuracy and efficiency of the electronic nose and GC methods applied in this study. The efficiency was measured by the time to prepare and analyze an unknown mixture leading to the prediction of its compositions. The longer the time needed, the less the efficiency. As mentioned in Chapter 4, for the GC methods there were 2 replicates x 2 injections required to investigate the composition of a sample of spice mixture. The sample preparation time (the extraction-distillation procedure) for each replicate was 3 hours and for each injection the operation time of GC was 30 minutes. Therefore, the total time needed for preparing/analyzing a mixture for composition prediction was 3 x 2 + 0.5 x 4 = 8 hours. This estimation of time did not consider the GC cool down time, which was around 20 minutes and was necessary between injections. For e-nose methods, the time needed for preparing the sample was negligible. Since five replicates were obtained for each mixture composition prediction,






83


and the operation time for a replicate was 10 minutes, the total time needed for e-nose method was a few minutes more than 50 minutes, e.g., less than one hour. It was found that although the GC methods predicted the mixture compositions more accurately, their efficiencies were much lower than that of e-nose methods. Table 5-13. Comparing the accuracy and the efficiency of GC and electronic nose
methods
Accuracy Efficiency
(Prediction MSE) - (Exp. Operation Time, hours) GC Single volatile method 0.000755 More than eight
Five-volatile method 0.00156
MLP 0.0051
E-nose PCA-MLP 0.0041 Less than one
TDNN 0.0035

Figure 5-14 illustrates the best prediction performances resulting from electronic nose analysis and that from gas chromatography methods. It should be noted that the GC method predicted the mixture compositions more accurately when the point representing the desired mixture was close to the boundary of the simplex space, while the e-nose method generated a more uniform error distribution. For the mixtures who represented points far away from the simplex space boundary (2T and 4T in the Figure 6-2), the amounts of the prediction errors were similar for the GC and e-nose methods. In other words, when there are no dominant component(s) in a mixture, the e-nose method may predict the mixture's compositions as well as the GC methods.

In addition to efficiency, the e-nose methods have two other advantages over the GC methods: the non-destructive sampling procedure making the on-line or near on-line quality/processing monitoring feasible, and the solvent-free analysis that reduces the concerns of inhalation of harmful chemicals during experiments.






84


Garlic





1T t


0




A

2T. 4T A

3T 3T


6% Difference from Desired Points
(below the sensory thresholds)



Desired Points

Sensory Comparison Points Predicted Pointed from TDNN using e-nose time-series data Predicted Pointed from GC analysis (Single volatile method)


Cinnamon


Figure 5-14. Comparison between the prediction performance of electronic nose and gas
chromatography method


Basil














CHAPTER 6
SUMMARY, CONCLUSIONS AND SUGGESTIONS FOR FUTURE STUDY This study investigated both the experimental and the data analysis methods in

applying an electronic nose to predict the mixture composition of a ternary spice mixture. The developed experimental method provided sufficient data to perform further data analysis. Three different neural network structures: multilayer perceptron (MLP), MLP using principal components analysis as preprocess and time-delay neural network (TDNN) were applied to e-nose data analysis. TDNN achieved the best prediction on testing data. The relative accuracy and efficiency of the developed methods were determined by comparing them to those from the two traditional methods: GC and sensory analysis. In GC methods, the volatile components of the spice mixtures were extracted and then quantified by GC. The mixture compositions of the testing blends were predicted based on the amounts of the extracted volatiles from each spice. The testing blends used in GC analysis were the same as those in e-nose experiments. The sensory analysis was performed to estimate the difference thresholds of the ternary spice mixtures.

Except for a few testing points, the three neural network models built from e-nose data predicted the compositions of the testing mixtures with an error less than 6%. It was concluded from the sensory analysis that the 6% difference was far below the difference threshold and not big enough to elicit a perceived difference when there was no dominant component(s) in the ternary mixture. The GC methods provided a more accurate but less


85






86


efficient prediction. The experimental time required for GC methods was 8 hours, instead of 50 minutes for e-nose methods.

In summary, the objectives of this study have been achieved. Both the appropriate experimental and data analysis methods were developed for an e-nose to quantitatively predict the composition of a spice mixture. The combination of these methods also provided significantly improved efficiency in composition prediction compared with the GC method yet resulted with acceptable prediction accuracy.

Using the developed procedure, the mixture composition can be predicted in a near on-line speed (50 minutes for an unknown sample), which makes the procedure valuable in quality monitoring or process control. The e-nose methods can also be used as a fast screening tool for product matching or re-formulation to indicate the lack or redundancy of the ingredient(s) of interest.

There were three quantitative data analysis methods developed in this study and all showed acceptable prediction accuracy. These methods could be used in other quantitative analyses with an electronic nose as the measurement tool.

This study has successfully developed the procedure to predict the composition of a ternary mixture. It would be interesting to investigate the e-nose's prediction ability toward the mixture with more components (equal or more than four). As discussed in Chapter 5, the data structures of the e-nose sensors responses are not ideal. Development and selection of suitable combinations of e-nose sensors will provide better prediction for mixture compositions. The GC data analysis methods used in this study were based on unique volatiles from each spice component. It will also be interesting to investigate the data analysis method for GC in case that the mixture under study without unique major






87


volatile(s). As that has been done in this study for the developed e-nose data analysis methods, sensory evaluation methods can be used to evaluate the relative accuracy of the developed GC data analysis methods. The sensory perception behavior towards mixtures is complex. Further investigation on this topic will provide valuable information for product matching or re-formulation.




Full Text

PAGE 1

PREDICTING SPICE MIXTURE COMPOSITION: COMPARING ELECTRONIC NOSE, GAS CHROMATOGRAPHY, AND SENSORY METHODS By HAOXIAN ZHANG A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTL^L FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2003

PAGE 2

Copyright 2003 by HAOXIAN ZHANG

PAGE 3

To my dearly loved husband, Hailong, with whom enjoyable and difficult times were shared throughout this journey, and to my parents, Chuanmao and Jiyun, for their love and support throughout my life.

PAGE 4

ACKNOWLEDGMENTS I would like to express my sincere thanks and appreciation to my advisor. Dr. Murat O. Balaban, for his invaluable advice, encouragement, support and guidance throughout my graduate studies at the University of Florida and for giving me the opportunity to study the interesting subject of this research. I would also like to thank my committee members, Drs. Kenneth M. Portier, Jose C. Principe, Charles A. Sims and Arthur A. Teixeira, for their help, suggestions, and words for encouraging this research. My accomplishments could not have been achieved without their support. Special thanks go to Dr. Portier and Dr. Principe for their time and patience to teach me in the topics of multivariate statistics and artificial neural networks, respectively. Finally, I am thankful to all my friends at the University of Florida, especially those at Dr. Balaban and Dr. Sims' lab, for their helpful suggestions, discussions and fiiendship. Also my appreciation goes to the panelists who helped in the sensory studies. iv

PAGE 5

TABLE OF CONTENTS Page ACKNOWLEDGMENTS iv LIST OF TABLES viii LIST OF FIGURES x ABSTRACT xii CHAPTER 1 INTRODUCTION 1 2 LITERATURE REVIEW 5 Spices 5 Electronic Nose 6 Introduction 6 Applications in Food Area 7 Equipment 8 Data Analysis Methods 11 Gas Chromatography 13 Introduction 13 Equipment 13 Sampling Methods of Volatiles 16 Application to Spice Mixture Analysis 20 Sensory Evaluation 21 Introduction 21 Measurement of Sensory Thresholds 22 Discrimination Tests 24 Difference Threshold Psychophysical Theory 26 Difference Thresholds of Mixtures 26 Neural Networks 27 Introduction 27 Multilayer Perceptrons (MLPs) Trained by Back-Propagation (BP) 28 Time-Delay Neural Network (TDNN) 30 Softmax 33 V

PAGE 6

3 OBJECTIVE 34 Electronic Nose Analysis 34 GC Analysis 34 Sensory Analysis 34 4 MATERIALS AND METHODS 35 Materials 35 Experimental Designs and Methods 36 Mixture Experimental Design of Three Components 36 Electronic Nose 38 Gas Chromatography 41 Sensory Thresholds 47 Data Analysis 51 Electronic Nose 51 MLP 51 PCA-MLP 52 TDNN 53 Gas Chromatography 53 Single volatile method 54 Five-volatiles method 55 Sensory Thresholds 56 5 RESULT AND DISSCUSION 58 Electronic Nose 58 Raw Data 58 Data Analysis Results Using MLP 59 Data Analysis Results Using PCA-MLP 60 Data Analysis Results Using TDNN 61 Discussion 63 Gas Chromatography 67 Volatile Components of Spices 67 Single Volatile Method 72 Five-volatiles Method 72 Discussion 73 Sensory Thresholds 79 Comparison between E-nose and GC/Sensory Methods 81 vi

PAGE 7

6 SUMMARY, CONCLUSIONS AND SUGGESTIONS FOR FUTURE STUDY ....85 APPENDIX A INSTANTANEOUS DATA FOR E-NOSE ANALYSIS 88 B TIME SERIES DATA FOR E-NOSE ANALYSIS 96 C RAW DATA FOR GC ANALYSIS 98 D RAW DATA FROM SENSORY ANALYSIS 1 03 LIST OF REFERENCES 115 BIOGRAPHICAL SKETCH 124 vii

PAGE 8

LIST OF TABLES Table E^ge 2L Classification of test methods in sensory evaluation 21 4-1 . The chemical names and their corresponding CAS# and KI (for DB-5) of major volatile compounds of basil, cinnamon and garlic 36 4-2. The weight fi:actions of design points for training/cross-validation 40 4-3. The weight fractions of spices for testing 40 4-4. The changes of the percentage of extraction of the total volatiles fi-om the spice mixtures with equal proportions of basil, cinnamon and garlic versus distillation time using SDE 44 4-5. The retention times of n-paraffin hydrocarbons (C7-C20) 47 46. The mixing fractions of the six pairs of spice mixtures compared by panelists 50 51 . The experimental and predicted mass Suctions of spice mixtures predicted by the optimal performance of MLP 59 5-2. The proportion and cxmiulative proportion of total variance explained by principal component 1 through 10 61 5-3. The experimental and predicted mass fi^actions of spice mixtures predicted by the optimal performance of PCA-MLP 62 5-4. The experimental and predicted mass fi-actions of spice mixtures predicted by the optimal performance of TDNN 63 5-5. The average relative amounts of all volatiles identified in each of the three spices and their corresponding retention times 68 5-6. The five most abundant unique volatiles identified in the basil, cinnamon and garlic, and the volatiles' corresponding chemical identities 70 5-7. The average fiaction (in percentage) of the volatiles in their corresponding pure spice extracts 71 viii

PAGE 9

5-9. The K values of the most abundant garlic volatile at different weights of garlic in the corresponding spice mixture 76 5-10. The K values of the five most abundant unique basil volatiles at different weights of basil in the corresponding spice mixture 76 5-1 1 . The K values of the five most abundant unique cinnamon volatiles at different weights of cinnamon in the corresponding spice mixture 78 5-12. The K values of the five most abundant unique garlic volatiles at different weights of garlic in the corresponding spice mixture 78 5-13. Comparison between the prediction error of electronic nose methods and the sensory threshold of human subjects 82 5-14. Comparison between the prediction performance of electronic nose and gas chromatography method 84 xi

PAGE 10

LIST OF FIGURES Figure page 21 . Cross section of a Fused Silica Open Tubular Column 15 2-2. Multilayer Perceptrons (MLPs) with one hidden-layer 30 2-3 . Tapped delay line with p delay units 31 2-4. Paradigm of Time-delay Neural Network with one hidden-layer 32 4-1 . The simplex factor space with three components 37 4-2. The three basic design methods in the simplex space with three components 38 4-3. The distribution of design points in the simplex space 39 4-4. The apparatus used for simultaneous distillation extraction 43 4-5. The comparison points for sensory threshold testing 49 4-6. The single volatile method for GC data analysis 55 47. The five-volatile method for GC data analysis 56 51 . A sample of time series data obtained 58 5-2. The optimal topology of MLP 60 5-3 . The optimal topology of PC A-MLP 63 5-4. The optimal topology of TDNN 65 5-5. The distribution of the prediction points, including those from e-nose MLP, MLPPCA and TDNN analysis, in the simplex space 66 5-6. The distribution of the prediction points, including those fi-om GC single volatile and five-volatile methods, in the simplex space 74 5-7. The K values of the most abundant basil volatile at different weights of basil in the corresponding spice mixture 75 X

PAGE 11

5-8. The experimental and predicted mass fractions of spice mixtures predicted by the single volatile method 72 5-9. The experimental and predicted mass fractions of spice mixtures predicted by the five-volatile method 72 5-10. The results of the triangle tests for each of the six pairs of comparisons 79 5-11. The number of correct responses necessary for the triangular tests to establish a significant difference between the two samples under comparison when the total number of the tests is 50 80 5-12. The comparison of panelists' performance between the tests carried first in that day and those carried second 81 5-13. Comparing the accuracy and the efficiency of GC and electronic nose methods ....83 A-1 . The instantaneous data obtained from e-nose experiments for training 89 A-2. The instantaneous data obtained from e-nose experiments for testing 95 C-1 . The raw GC data from each of the three pure spice extracts first injection 99 C-2. The raw GC data from each of the three pure spice extracts second injection 100 C-3. The raw GC data from the extracts from each of the five spice mixtures first injection 101 C-4. The raw GC data from the extracts from each of the five spice mixtures second injection 102 D-1. The triangle test results from the comparisons of the mixtures IT and lT-1 104 D-2. The triangle test results from the comparisons of the mixtures 3T and 3T-1 106 D-3. The triangle test results from the comparisons of the mixtures 4T and 4T-1 108 D-4. The triangle test results from the comparisons of the mixtures 4T and 4T-2 110 D-5. The triangle test results from the comparisons of the mixtures 4T and 4T-3 1 12 D-6. The triangle test results from the comparisons of the mixtures 5T and 5T-1 114 ix

PAGE 12

Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy PREDICTING SPICE MIXTURE COMPOSITION: COMPARING ELECTRONIC NOSE, GAS CHROMATOGRAPHY, AND SENSORY METHODS By Haoxian Zhang August 2003 Chair: Murat O. Balaban Major Department: Agricultural and Biological Engineering Electronic noses (e-nose) are instruments that can quickly detect odors at low cost, and their potential applications are very diverse. Limited work has been done in investigating e-noses' quantitative ability and no work has been reported on the application of e-noses to predict mixture compositions. The objective of this project was to develop a quantitative procedure that could quickly predict the compositions of a ternary spice mixture by using an e-nose for measurement and multivariate statistics/neural networks (NN) for data analysis. The relative accuracy and efficiency of the developed e-nose methods were determined by comparing them to those resulting from gas chromatography (GC) and sensory methods. Three groimd spices (basil, cinnamon and garlic) were mixed in different compositions and presented to an e-nose. Nineteen training blends were used to build a predictive model, the performance of which was tested by five other blends. Three NN structures were used for predictive model building (multilayer perceptron (MLP), MLP xii

PAGE 13

using principal components analysis as preprocess and time-delay NN). For GC analysis, the volatile components of spice mixtures were collected by simultaneous distillationextraction and quantified by GC. Mixture compositions were predicted based on the amounts of unique volatiles of each spice. The testing blends applied in GC analysis were the same as those in e-nose experiments. For sensory analysis, triangle tests were performed by 50 panelists to estimate the difference thresholds of spice mixtures. The best NN model built from e-nose data predicted the compositions of testing spice mixtures with an error less than 0.06. A difference of 0.06 between two spice mixtures was determined through sensory analysis to be lower than human sensory thresholds. The GC method provided a more accurate but less efficient prediction. Its experimental time required for each imknown sample was 8 hours, instead of 50 minutes for the e-nose method. The procedure developed in this study can predict the compositions of a ternary spice mixture with an acceptable accuracy and significantly improved efficiency. The procedure will be valuable in quality monitoring or process control, in which efficiency is essential. xiii

PAGE 14

CHAPTER 1 INTRODUCTION Electronic noses (e-noses) are instruments developed in the last 10 years to mimic the sense of smell. Consisting of olfactory sensors and multivariate signal analysis methods, they are able to quickly detect and distinguish odors at low cost. The potential applications of electronic noses are very diverse, ranging from military applications, clinical diagnosis, and environmental monitoring, to applications in food, flavor and biotechnology industries. Currently, the biggest use of the electronic nose is in the food and fragrance related areas. Until now, electronic nose studies were focused on investigating qualitative aspects of samples, e.g., detecting or classifying odor patterns, such as "fresh" versus "spoiled." More work in exploiting their quantitative ability can extend the horizon of e-nose applications. No work was found in applying e-noses to predict mixture compositions. In the food and fragrance industries, many products are formed by mixing two or more ingredients with different olfactory attributes. A simple, fast mixture composition prediction tool for such products would be valuable for routine quality or process control, as well as product matching or re-formulation. There are additional constraints in mixture experiments: the sum of all fractions must add up to one, and the fraction of each component must be non-negative. These bring specific considerations for the optimal experimental design. One of the objectives of this study was to develop an experimental procedure that can quantify mixture compositions using an electronic nose as the measurement tool. 1

PAGE 15

The signal analysis (data analysis) methods for e-noses were designed for pattern classification or recognition. Even in the few studies dealing with quantitative attributes of samples (Alpha M.O.S., 1997; Dittmann et al, 2000), the samples with different quantitative attributes were treated as patterns during data analysis. These methods perform well for qualitative analysis, but may not work well for quantitative requirements. Therefore, it is necessary to develop appropriate quantitative data analysis methods for e-noses. The development of these methods was another objective of this study. Multilayer perceptron trained by back-propagation (MLP), a simple and popular neural network topology, is a method for quantitative data analysis. Time delay neural network (TDNN), a dynamic network with embedded local memory, has demonstrated to improve an e-nose's pattern recognition ability (Zhang et al, 2003). Therefore, TDNN was selected as another possible method, with a potential to perform better than MLP. Other data analysis methods include incorporating principal component analysis (PCA), a multivariate statistical method that can preserve information in a less dimensional format, with the neural network structures described above. Spice mixtures were selected as the model system in this study because spices have high volatile content easily detectable by an e-nose; they are stable during storage making a large experiment feasible; they are valuable for their olfactory attributes, and they are regularly used as mixtures in the industry. For odor analysis, there are two traditional competitors of e-noses. One is gas chromatography (GC), an instrumental method that can separate and then identify volatile chemical compounds. Another is sensory analysis, which uses the senses of human

PAGE 16

3 subjects to measure the organoleptic characteristics of samples. The GC methods GC can provide detailed information about the volatile chemicals composing odors, and the variability of the GC methods is small. However, the experimental procedures of GC, especially the procedure of sample preparation, are involved and time-consuming. On the other hand, human sensory panels are powerful in assessing qualitative attributes of odors, and they are the ultimate arbiters in quality evaluation and product match. However, sensory methods also have limitations, including high expense and subjectivity. The assessment of how well a newly developed e-nose procedure performs in a mixture composition prediction should be based on the comparison of the performance between the e-nose and the two traditional methods. Theoretically, e-nose methods should predict the mixture composition more efficiently than GC methods (in terms of time) with some loss of prediction accuracy (the difference between prediction and real). Human subjects, whose quantitative ability is quite limited, are not expected to rival the prediction accuracy of electronic noses. A reasonable estimation of the acceptable prediction accuracy for quality evaluation or product matching can be determined by qualitative sensory methods since human subjects are the ultimate arbiters in these areas. Under the acceptable prediction accuracy levels, the difference between the real spice mixture and the predicted one should not be perceived by human subjects. The objective of this study was to develop both the experimental and data analysis methods of an e-nose to quickly predict the mixing compositions of ternary spice mixtures. The performance of the developed e-nose method(s) was determined in terms of accuracy and efficiency by comparing them to those from GC and sensory studies

PAGE 17

4 using the same ternary spice mixtures. It was expected that the e-nose method could save significant time in data collection and result in reasonable prediction accuracy.

PAGE 18

CHAPTER 2 LITERATURE REVIEW Spices According to the American Spice Trade Association's (AST A) official definition, spices are "any dried plant product used primarily for seasoning" (1999). Many parts of plants are used as spices, such as seeds, leaves, berries, bark, kernels, arils, stems, stalks, rhizomes, roots, flowers, bulbs, fioiits and flower buds. Spices are often used in their dried form to assure year-round availability, ease of processing, longer shelf life and lower cost. The essential oil of a spice can be extracted by steam distillation. The extracted essential oil may exhibit the characteristic flavor and aroma properties of the actual spice, but may or may not have exactly the same taste. Essential oils do not generally carry the "biting" principles of the spice and, therefore, may taste almost bland (Hirasa and Takemasa, 1998; Uhl, 2000). Spices are valued for the flavor, aroma, pungency and color they impart to food, not for their nutrient content. Therefore, the quality of spices must be estimated on the basis of their sensory characteristics and intensity. Spices and their extracts should be stored in closed containers under cool and dry conditions. Reconunended storage temperature is 20°C with 50% relative humidity. Light-sensitive materials such as parsley, chives and other green herbs should be protected against direct exposure to sunlight and fluorescent light (Giese, 1994). Storage under refrigeration temperatures can slow microbial growth but may negatively affect sensory quality. 5

PAGE 19

6 Electronic Nose Introduction The concept of electronic noses (e-nose) first appeared in the literature in the early 1980s (Persaud and Dodd, 1982). A well-accepted definition of the electronic nose is "an instrument which comprises an array of electronic chemical sensors with partial specificity and an appropriate pattern recognition system capable of recognizing simple or complex odors" (page 18, Gardner and Bartlett, 1994). The commercial instruments were available at the beginning of 1990s. Since then, the electronic nose has generated a widespread interest in the food, chemical, biotech and pharmaceutical industries. Currently, the biggest market for electronic noses is in the food area. Numerous studies have shown its potential to be used as a quality control or process-monitoring tool. Generally, an electronic nose is composed of a sampling system, a sensor array, a data acquisition system, and a signal-processing algorithm. Mimicking the human nose, the operation of electronic nose begins with "sniffing": collecting and conveying the volatile components of the sample to the sensor array. Sensor "states" are altered through chemical or physical interaction between volatile components and sensors, resulting in electronic signals. These are captured by a data acquisition system, and a "cleaning" procedure is then applied to restore initial conditions in both sensors and sampling system. The electronic signals are further analyzed by pattern recognition algorithms or other data analysis techniques. Electronic noses are different from traditional gas sensors in two aspects: partially selective sensors that are all broadly tuned, and multidimensional data that should be analyzed using multivariate techniques. Electronic noses can be used to evaluate foods when the food has volatile compounds that change with the characteristics under investigation, either qualitatively or

PAGE 20

7 quantitatively. The concentrations of volatile compounds in headspace should also be high enough to be detected by electronic noses. When they are too low to be detected at room temperature, heating the sample can release more volatile components into the headspace and may help the analysis. Electronic noses differ from human noses in both their sensors' types/numbers and signal processing methods, resulting in unmatched detection thresholds and odor discrimination capabilities (Doleman and Lewis, 2001). In other words, what an electronic nose smells is not the same as what a human nose smells (Burl et al, 2001). Therefore, the electronic nose is not a primary analytical technique. An electronic nose has to be trained to fulfill its capability of odor identification. The data used for training can be collected in two ways. One is by correlating an electronic nose's responses with other primary analytical methods, such as sensory evaluation, chromatography and wet chemistry analyses. The other is by collecting an electronic nose's response toward the "known" samples. The "known" samples can either be prepared or be obtained from a supplier. Applications in Food Area Substantial research has been done to apply electronic nose technology in the food area. Electronic nose technology has been applied for fruit ripeness determination (Benady et al, 1995; Simon et al, 1996; Llobet et al, 1999; Maul et al, 1999; Brezmes et al, 2000), fermentation process monitoring (Pearce et al, 1993; Eklov et al, 1998; Bachinger et al, 1998; Pinheiro et al, 2002), spoilage detection (Schweizer-Berberich et al, 1994; Blixt and Borch, 1999; Muhl et al, 2000; Evans et al, 2000; Korel and Balaban, 2002; Park et al, 2002), instrument assessment of sensory attributes (AnnorFrempong et al, 1998; Shen et al, 2001; Korel et al, 2001; Korel and Balaban, 2002;

PAGE 21

8 Garcia-Gonzalez and Aparicio, 2002; Bleibaum et al, 2002), detection of packaging odors (Holmberg et al. 1995; Poling et al, 1997; Deventer and Mallikarjunan, 2002; Heinio and Ahvenainen, 2002), and aroma profile control (Gretsh et al, 1998; Hirschfelder et al, 2000; Stella et al, 2000). Until now, most e-nose studies were aimed at detection of different odor patterns among samples, such as "fresh" versus "spoiled." There are a few studies dealing with quantitative attributes of samples. Dittmann et al. (2000) reported using an e-nose in detecting different doses of garlic flavorings. In an application note from Alpha-MOS, an e-nose was applied to discriminate binary hops blends with 3 different mix ratios (1997). However, in these studies the samples with different quantitative attributes were freated as patterns during the data analysis. This means that the developed procedure could only be used to determine whether the sample had this quantitative attribute or not, and the actual level of the quantitative attribute could not be determined. For the purpose of quality confrol or process monitoring, it would be valuable to know not only the qualitative aspects of samples but also their quantitative attributes. Equipment Dozens of companies are now designing and selling elecfronic noses. The aspects to be considered when selecting an elecfronic nose instrument include its sampling system, sensor types/numbers, and data acquisition system. The aroma of a food is a complex mixture of volatile compounds. Often the difference between a "good" or "bad" odor associated with a food is in the relative amounts of the volatile compounds. It is impossible to have sensors specific to each chemical compound. The "broadly tuned" sensors of elecfronic noses are designed to solve this problem. Except the non-specificity requfrement, ideally each sensor's

PAGE 22

9 response should be orthogonal to the others. The less the correlation between different sensors, the more information about sample properties is provided by the sensor array's responses. To evaluate a food whose quality deteriorates fast, the sensor's rapid response and recovery time is also important. Other requirements of the electronic nose sensors include reproducibility to a given odor and insensibility to changes m temperature, hiunidity and flow rate. At present, three types of gas sensors-metal-oxide, polymer, and surface acoustic wave gas sensors—are widely used in commercial instruments (Snopok and Kruglenko, 2002). The significant advantage of metal-oxide sensors (MOS) is their low sensitivity toward humidity and low drift over time. Compared with MOS, the polymer sensors have the advantages of (1) responding to a broad range of organic vapors, (2) operating at room temperature, and (3) rapid response and recovery time. Surface acoustic wave gas sensors are of two main types: the bulk acoustic wave device (BAW), also referred to as the quartz crystal microbalance (QCM), and the surface acoustic wave device (SAW). QCM has a linear response to concentration, compared with the power law of metaloxide and Langmuir response for conducting jwlymers. The disadvantage of BAW and SAW is their high sensitivity to disturbances such as temperature and hxmiidity fluctuations (Gardner and Bartlett, 1999). There are new types of chemical sensors under development for electronic noses. One example is the so-called Smell-Seeing™ (Rakow and Suslick, 2000) sensor, which detects odors using the colorimetric response from a library of immobilized vaporsensing dyes. These sensors are insensitive to humidity and provide visual identification of odors. Yet, for the purpose of objective data analysis, the color graphs generated from

PAGE 23

10 these sensors may need complex preprocessing to generate suitable multidimensional data. It is arguable whether a single sensing device that produces an array of measurements of odors can be regarded as an electronic nose sensor array. There were two such "single sensing devices" that had been put under the umbrella of electronic nose sensor arrays: mass spectrometry based sensors (MS-sensor) and gas chromatograph based SAW sensors (GC/SAW). In the MS-sensor system, volatile components are introduced into mass spectrometer without separation and then selected fragment ions are treated as response (Dittmann et al, 2000). As a mature technology, mass spectrometry provides the benefit of reproducibility and standard calibration methods. Yet the cost of the instrument may compromise its benefit. For the GC/SAW system, volatile components are separated by a fast GC column and detected by SAW sensor; the responses of the single sensor at different times are viewed as a multidimensional response. The problem with this sensor is that the single sensor xised may not respond to some of the important volatiles. The sampling system is designed to provide a stable and reproducible sensor reading environment so that all factors capable of influencing sensor responses are kept under control (Falcitelli et al, 2002). Generally the sampling system of an electronic nose has two separate chambers: a sample chamber and a sensor chamber. The temperature and/or humidity of both chambers are controlled. The static or dynamic headspace of samples is conveyed to the sensor chamber through gas flow. Inert gas is then applied to both chambers to clean any possible odor memory of previous samples or possible contamination. Some of the hand-held electronic noses, such as the product from Cyrano

PAGE 24

11 Sciences, do not have a sample chamber. It is necessary in that case to design a sample chamber to ensure repeatable measurements. Non-absorbent and inert materials should be selected to build both chambers to prevent leftover odors from previous sample or other contaminations (Falcitelli et al, 2002). The fluid dynamics of sampling system should also be properly designed to avoid any stagnant regions and the cleaning procedure of using inert gas to remove odor memory should have flexibility. For foods with strong volatile contents, a longer cleaning time or fast cleaning gas flow rate is expected. The way to pre-process the time-dependent analog sensor signal is also an important design parameter for an electronic nose instrument. The relative and fractional change of baseline is the most common output reading of elecfronic nose sensors. Steady state or static rather than fransient or dynamic are the descriptors commonly used to defme the sensors' responses. Most commercial elecfronic noses provide their pattern recognition software package (Snopok and Kruglenko, 2002). However, independent software, such as SAS®, STATISTICA®, S-Plus®, SPSS®, MATLAB® or other Neural Network packages, are also popular in elecfronic nose data analysis. Data Analysis Methods The underlying relationship between the responses of an elecfronic nose and the characteristics of the sample is determined by two basic approaches statistical multivariate analysis or artificial neural networks. Currently, the data analysis methods for elecfronic noses are solely used to classify or recognize patterns. The commonly used statistical multivariate methods include principal component analysis (PCA), discriminant function analysis (DP A), cluster analysis (CA), partial least squares (PLS)

PAGE 25

12 and canonical correlation analysis (CCA). For electronic nose data analysis, the most commonly used neural network structure is multi-layer perceptron (MLP) trained by back-propagation. Other neural network structures that have been applied to electronic nose data include neuro-fiizzy system (Ping and Jun, 1996), self-organizing maps (SOM) (Marco et al. 1998), MLP with genetic algorithms (Kermani et al, 1999), adaptive resonance theory (ART) (Distante et al, 2000) and radial basis fimction (RBF) (Evans et a/., 2000). Most statistical multivariate methods are based on a linear approach while neural networks are non-linear. In cases in which data correlations are non-linear, neural networks may perform better than conventional multivariate methods. According to the goal of the study, the data analysis methods applied for electronic noses can be divided into two categories: exploratory methods and predictive methods. Exploratory methods are used to detect whether there are systematic patterns in the investigated data set, while predictive methods are designed to find the prediction model to describe the data structures when there is priori knowledge about their existence. Among the commonly used e-nose data analysis methods, principal component analysis, cluster analysis and unsupervised neural networks, such as self organizing maps, belong to the category of the exploratory methods. Discriminant function analysis, partial least square, canonical correlation analysis and all supervised neural networks, including multilayer perceptron, are predictive methods. As one of the most widely applied multivariate techniques, principal component analysis can also be used as a "data-compressing" tool for data pre-processing in neural networks (Principe et al, 1999). By reducing the dimensionality of information, principal

PAGE 26

13 component analysis lowers the number of inputs to neural networks, resulting in fewer parameters necessary in neural networks. Like any other predictive structure, the generalization ability of a neural network will be enhanced with fewer parameters. Hence, pre-processing of data by principal component analysis improves the generalization ability of neural networks. This method has been applied to classify different brand of coffees (Pardo et al, 2000). Gas Chromatography Introduction Gas chromatography (GC) is an instrumental method for separation and identification of volatile chemical compounds. The compounds of interest must first be removed fi-om the matrix (solid or liquid phase) and isolated fi-om any interference through the process of sampling. The extracted sample is then introduced into a heated injector, carried through a separating column by an inert gas, and detected as a series of peaks on a recorder when components leave the column. Each component of the sample reaches the detector at a different time and produces a signal at a characteristic time called a retention time. The area under a peak is related to the amount of that component present in the sample. The retention time of a compound differs fi-om colvmin to column and is subject to the settings of the GC operation parameters. Kovats retention indices (KI), which reflect the retention of a given compound relative to those of the two bracketing n-paraffin hydrocarbons, are a constant for a chemical compound and can be used to identify the compound when the standard information is available. Equipment GCs are composed of four basic units: an injector, an analytical column, a detector and a signal recorder.

PAGE 27

14 The role of the injector is to rapidly vaporize the sample and thoroughly mix the volatiles so they can be swept onto the column by the carrier gas. The most commonly used injector is the programmed temperature vaporizing (PTV) injector that can be operated in two modes: split and splitless. While the temperature of the injector is programmed to increase, each component in the sample is vaporized and moved to the column. In split injection, a defined fraction of the sample vapor enters the column, with the remainder leaving the injection through a vent. In splitless injection, the split vent flow is blocked during the injection period such that all sample vapors enter the column (Jennings etal., 1997). The analytical column is where the analytes are separated. The separation relies on the distribution of analytes between two phases: the stationary phase on the inside of the column and the mobile phase of the carrier gas. Today, the Fused Silica Open Tubular (FSOT) colunm is viewed as "state of the art." Its cross section is shown in Figure 2-1. The factors to be considered in column selection include the stationary phase, film thickness of the stationary phase, column diameter and column length. Selection of stationary phase is based on sample polarity; for complex samples, the selected column should best reflect the overall polarity of the sample. In the area of food flavor analysis, the most commonly used column is the DB-5 or equivalent (apolar column with 95% polydimethylpolysiloxane and 5% Phenyl). Typically, Kovats retention indices for the polydimethylsiloxane column are a constant within a window of +/0.1%. Increasing film thickness and diameter of the column will allow for a greater sample capacity with wider peaks and lower resolutions. It is generally advisable to use small internal diameter (0.2-

PAGE 28

15 0.32 mm) columns with thin films (0.2-0.35 [im). For a general-purpose column, 30 m is usually the most useful length (Jennings et ai, 1997). Polyimide coating Fused silica tube Chemically bonded stationary phase Figure 2-1. Cross section of a Fused Silica Open Tubular Column As American Society of Testing and Materials (ASTM) have defined that detectors are devices that identify the presence of the components as they elute from the column (1996). There are many types of detectors available, each with its particular utility. Some are universal detectors such as the mass spectrometer (MS) or the flame ionization detector (FID). Some are selective in that they respond more to certain classes of compounds. Two examples of selective detectors are the flame photometric detector (FPD) used for the analysis of organophosphate pesticides, and the electron capture detector (BCD) used mainly for the analysis of chlorinated compounds (Scott and Perry, 1998). A third, newer detector is the atomic emission detector (AED), which can be specific to particular elements such as carbon, chlorine, bromine, iodine, tin, mercury, sulfur, nitrogen, phosphorous as well as others common to pesticides and herbicides, and thus is much less subject to interferences. The signal recorder is used to record the signal detected by detectors. In most cases, the signal recorder is either a computer or a printer or a combination of both. Chemically inert gas such as nitrogen, heliimi and hydrogen can be used as carrier gas. Hydrogen, allowing much higher flows with little loss of resolution, is the best among the three for GC applications (ASTM, 2000).

PAGE 29

16 Sampling Methods of Volatiles For the investigation of volatiles, the sample preparation is the most critical and most complex step in the entire analytical process. At the present, the two most common procedures are headspace methods and extraction methods. The following extraction methods have been reported in the literature: steam distillation followed by solvent extraction, direct solvent extraction, simultaneous steam distillation/extraction (SDE), co-distillation and supercritical fluid extraction (SFE) (Parliament, 1997; Sides et al, 2000). Comparing with direct solvent extraction, steam distillation followed by solvent extraction has the advantage of separating the volatiles from the nonvolatiles. Steam distillation works best for compounds that are slightly volatile and water insoluble. Steam for distillation can be generated internally or externally. The usage of external steam (indirect steam distillation) results in an extraction with less decomposition of the sample since the sample is not heated directly. If sample decomposition remains a concern, then the steam distillation may be operated under vacuum. A direct solvent extraction is regularly performed with a Soxhlet extractor. A dried sample such as a spice or a grain can be ground finely and placed in a Soxhlet thimble and extracted by an organic solvent such as diethyl ether or methylene chloride. After a nimiber of cycles, the solvents are then combined and concentrated. Nonvolatile organic materials such as lipids and pigments will also be concentrated. The sample may be analyzed directly or after removal of the solvent. If the sample contained large amounts of lipids as in the coffee or chocolate, subsequent steam distillation may be needed to remove the nonvolatiles.

PAGE 30

17 Simultaneous steam distillation/extraction (SDE) apparatus was first described by Likens and Nickerson (1964). The apparatus can ensure continuously recycling for both the steam distillate and an immiscible organic solvent. SDE technique provides a smgle operation that can remove and highly concentrate the volatiles. In the operation, just a small volume of solvent is required, which reduces the problem of artifact buildup, as solvents are concentrated. The apparatus can be operated under reduced pressure to reduce thermal decomposition but the operation imder vacuum is quite complex (Maignial et al, 1992). A number of refinements have been made to the basic SDE apparatus and several versions are commercially available (Chaintreau, 2001). Many solvents have been employed in SDE. Hexane was an excellent solvent except for lowerboiling water-soluble compounds, where diethyl ether was considerably better (Schultz et al, 1977). Use of methylene chloride has been recommended in a modified LikensNickerson extractor (Aug-Yeung and MacLeod, 1981). Currently, most researchers appear to be using pentane-diethyl ether mixtures. A relatively new technique is co-distillation. In this technique, a solvent such as diethyl ether, pentane, or methylene chloride is dispersed in the sample and the sample is distilled rapidly at 200°C until all the solvent and a small amount of water have passed over (Misharina et al, 1994). The advantages of co-distillation are that isolates are generated without a boiled note, and it takes only 15-20 minutes for a distillation. The recovery of volatiles by SFE is comparable with that achieved by direct solvent extraction using Soxhlet apparatus (Ropkins and Taylor, 1996). SFE performed well in extracts tightly bound and encapsulated volatiles, and the solvent is easy to remove. However, the equipment of the SFE system is complex (consisting of a source fluid, a

PAGE 31

18 pump, a sample vessel, a restrictor and a trapping device) and optimization of extraction conditions such as temperature and pressure is necessary (Messer et al, 1998; Burford, 1998; Diaz-Maroto et al, 2002). Headspace sampling techniques are frequently divided into two broad categories: static headspace and dynamic headspace. In the headspace analysis, volatile analytes from a solid or liquid matrix are sampled by investigation of the atmosphere adjacent to the sample (Wampler, 1997; Rouseff and Cadwallader, 2001). Solid-phase microextraction (SPME) is a relatively new technique that can be used for headspace sampling (Harmon, 1 997). Static headspace sampling is performed as follows: the food sample is placed into a headspace vial, sealed and then allowed to stand for a period of time to establish equilibrium at that temperature. The vial may be warmed to enhance vaporization of volatile. A small sample (usually about 1ml) of the atmosphere around the sample is injected directly into the GC column. The amount of gas that may be injected into a gas chromatography is limited by the capacity of injection port, the colunm, and consideration of the increase in the pressure and flow in the injection port caused by a gas phase injection. The disadvantage of static headspace methods is that they may lack sensitivity due to low concentration of volatiles in sample headspaces (Kolb and Ettre, 1997). Dynamic headspace involves moving the analytes away from the sample headspace and concentrating in some kind of "trap." Instead of allowing the sample volatiles to come to equilibrium between the sample matrix and the surrounding headspace, the atmosphere around sample matrix is constantly swept away by a flow of carrier gas.

PAGE 32

19 taking the volatiles with it. This prevents the establishment of an equilibrium state and then allows more release of volatiles into the headspace. Therefore, dynamic headspace offers increased sensitivity compared to static headspace. "Purge and trap" is another term used m literature for dynamic headspace technique. Regularly, purge and trap refers to the technique applied to a liquid sample, while "dynamic headspace" is used when the sample material is a solid. The trapping system for dynamic headspace can be either sorbent or cryogenic trapping. Most sorbent materials are porous polymers similar to materials used to fill packed GC columns for gas analysis. Tenax® is the most widely used, general purpose sorbent for dynamic headspace techniques. Both liquid nitrogen and solid carbon dioxide can be applied for cryogenic trapping. Since the presence of water on a capillary GC column can pose a serious analytical problem, it is important to remove water that is carried away from the sample and collected in the trap before transferring the trapped organics to the chromatograph. Tenax® is hydrophobic. It is usually enough to pass a source of dry carrier gas through the trap for a minute or two to vent the water from the trap without disturbing the organics. SPME was developed by Pawliszyn's research group at the University of Waterloo in the late 1980s. It is a solventless technique that incorporates extraction, concentration, and sample introduction. The SPME devices are syringe-like with an outer septumpiercing needle. A plunger houses a fused silica fiber coated with a stationary phase. The fiber can be inserted into the sample matrix (aqueous samples) or the sample headspace. After concentration of analytes on the fiber, the syringe assembly is inserted into the injection port of a gas chromatograph where the analytes are thermally desorbed from the

PAGE 33

20 fiber and cold-trapped on the head of the capillary column (Pawliszyn, 1999; Pawliszyn, 2000). Fibers coated with nonpolar polydimethylsiloxane (similar to OV-101) and the more polar polyacrylate are commercially available. For most volatile flavor analyses, a fiber having a lOO-^un coating of polydimethylsiloxane is preferred. In general, the fibers coated with thicker films will require a somewhat longer time to achieve equilibrium but might provide higher sensitivity due to the greater mass of the analytes that can be absorbed. For analysis of high-boiling-point components, fibers with a 7-micron thickness of polydimethylsiloxane usually work best. The major advantages of SPME include fast, solvent fi-ee, easy-to-automate and sensitive to high-boiling-point components. The key to getting accurate and reproducible results with SPME is to be sure to perform sampling in exactly the same way each time. The actual position of the fiber in the headspace is also important (Cai et al, 2001 ; Marsili, 2001). One should be aware that none of these techniques of volatile sampling produce an isolate that quantitatively represents the composition of the starting material. Jennings et al. (1977) has compared various sample preparation techniques, including dynamic headspace and simultaneous distillation-exti^tion. Their conclusion was that the extracts collected by distillation-extraction most nearly agreed with the origmal sample. Application to Spice Mixture Analysis The identification and quantification of volatile components of spices have been well-investigated using GC methods and their combination with mass spectrometry. However, most of published articles were based on tiie analysis of an individual spice and there were a few published articles of identification/quantification of spices fi-om a mixture. Cheng et al. (1997) reported a qualitative and quantitative spice mixture analysis

PAGE 34

21 method. The spice mixtures were extracted by simultaneous steam distillation/extraction and then injected into GC. The recognition and quantification of spices was accomplished via numerical methods based on a database consisting of 355 different spices. The accuracy of the method was not defined since the spices in the prepared mixture and the spices m the database were not identical. Sensory Evaluation Introduction Sensory evaluation has been defined as "a scientific discipline used to evoke, measure, analyze, and interpret reactions to stimuli perceived through the senses" (page 3, ASTM, 2001). As summarized in Table 2.1, the commonly used sensory tests can be divided into three categories according to their goals and their criteria for the selection of panelists (Lawless and Heymann, 1998; Meilgaard et al, 1999; ASTM, 2001). Table 2-1. Classification of test methods in sensory evaluation Categories Question of Interest Test Methods Panelist Characteristics Discrimination Is there any Triangle Screened for sensory detectable Duo-trio acuity, oriented to test difference in "A"not "A" method, sometimes products? Paired comparison trained Rating Threshold Descriptive How do products Flavor profile Screened for sensory differ in specific Textiire profile acuity and motivation. sensory QDA® trained or highly trained characteristics? Spectrum™ Time-Intensity Free-choice Affective How well are Ranking Screened for product use. products liked or Hedonic imtrained which products are preferred?

PAGE 35

22 The sensory perception of stimulus is an active and selective process, which is influenced by both physiological and psychological factors. Those factors were well discussed by Meilgaard et al. (1999). There is also standard practice for serving protocol to avoid bias generated by these factors (ASTM, 1997a). Measurement of Sensory Thresholds Threshold is a measure of human sensitivity to a given stimulus. Four types of thresholds exist, namely absolute threshold, recognition threshold, difference threshold and terminal threshold, each described below (Meilgaard et al, 1999; ASTM, 2001): Absolute threshold: also called detection threshold, is the minimum intensity of stimulation enable of eliciting a response at a probability of 50%. Recognition threshold: the minimum intensity of stimulation at which the stimulus can be recognized and identified with a specific probability, most frequently 0.50. Difference threshold: the minimum of difference required between two stimuli that will elicit a perceived difference with a specific probability, most frequently 0.50. Terminal threshold: the mtensity of stimulation above which increase in intensity cannot be detected. In the early days of psychophysics, thresholds were commonly measured by the "method of limits" or the "method of constant stimuli." In the method of limits, stimulus intensity would be raised and then lowered sequentially to find the average point at which the observer's response changed from negative to positive or fi-om positive to negative. In the method of constant stimuli, a number of stimuli values are selected to "bracket" the assumed threshold. These stimuli are then presented many times to the subjects in a randomized order. For each of the stimuli, the percentage of correct response is calculated and the threshold is interpolated based on the predetermined criterion. The

PAGE 36

23 problem associated with the classical methods is that the perception of a response depends on the observer's criteria and willingness to guess. An objective proof of detection can be achieved by incorporating the forced-choice element (subjects are forced to make guesses whenever they are in doubt) into the classic methods (Moulton et al, 1975; Baird, 1997). A typical example of such methods is the ascending series of 3 -AFC tests for absolute or recognition threshold detection, which has been described in the ASTM standard practice E679-91 and E1432-91 . The test procedure works as follow: First, based on the estimated threshold range, the test samples whose concentrations "bracket" the threshold are prepared. These samples should have over three to four concentration and each concentration differed by a factor of 2 or 3. The test samples are then presented to the test panel that should have more than 25 panelists in an ascending 3AFC series (Brown et al, 1978; ASTM, 1991a; ASTM 1991b). It should be noted that in forced-choice methods, part of the correct responses are not elicited by perceived sample characteristics, but identified by guessing. Therefore, for threshold determination usmg forced-choice methods, the criterion that defines a threshold is actually the probability of correct response excluding the guessing rate. This probability is referred as "percent correct above chance" in the literature. Group thresholds can be derived by averaging the individual thresholds or fi-om group "percent correct above chance" at each stimulus level. In the latter approach, the tests results from forced-choice method the can be converted to percent correct at a stimulus level by a defined formula (Moulton et al, 1975; ASTM, 1991b; Antinone et al, 1994).

PAGE 37

24 Most of the forced-choice tests used in the threshold determination are based on discrimination methods. Threshold testing can be viewed as a special case of discrimination (simple difference) testing. Since some discrimination methods present a more difficult task to the participants in the test, the value of a threshold is not fixed, but is method-dependent and will rise with the difficulty of the task. In recent years, the Signal Detection Theory (SDT) has become popular among psychophysicists to determine thresholds. In SDT methods, the point of interest is not the thresholds, but "the size of the psychological difference between the two stimuli", which has the name of d'. The advantage of SDT is that the decision process of subjects becomes more explicit and can be modeled statistically (Meilgaard et al, 1999). However, SDT procedures are more time-consuming, and Frijters and his co-workers' studies have shown that there is a 1 : 1 relationship between d' and the classical threshold for forced-choice methods (Frijters, 1980a; Frijters et al, 1980b). For these reasons, both ASTM (1991a and 1991b) and ISO (1999) are still using the method of limits as their practice standard for threshold measurements. Discrimination Tests There are two groups of discrimination tests: overall discrimination tests and attribute discrimination tests. Overall discrimination tests are designed to find whether a sensory difference exists between samples, while attribute discrimination tests are designed to answer the question "how does attribute X differ between samples?" Three most commonly used discrimination tests are triangle, duo-trio and n-altemative forcedchoice procedure (n-AFC) (Meilgaard et al, 1999). In the triangle test, three samples are presented simultaneously to the panelists; two samples are fi-om the same formulation and one is different. Eiach panelist has to indicate

PAGE 38

25 which sample is the odd one. Generally, 20 to 40 panelists are necessary for a triangle test, while 50 to 100 panelists are required if the study seeks to demonstrate similarity. The panelists should be familiar with the triangle test and with the product tested (ASTM, 1997b; Meilgaard et al, 1999). Triangle tests, along with duo-trio, are overall discrimination tests. N-AFC methods, also called directional difference tests, are attribute discrimination tests. In n-altemative forced-choice procedure, the panelists are asked to choose the object with the most (or the least) of an attribute from among n objects, n-1 of which have identical independent distributions of the sensory attribute. The danger with the method is that other sensory changes may occur when one attribute is modified and these may obscure the attribute in question. The panelists should be trained to be familiar with the attribute in question (Lawless and Heymann, 1998; Meilgaard et al, 1999). Power of a discrimination test is defined as 1p, where P is the probability of committing a Type II error (not rejecting the null hypothesis when the null hypothesis is not true). Triangle and duo-trio tests are much less powerful than n-AFC tests. This is due to the difficulty of the tests being performed. In n-APC methods, subjects need only determine the maximum or minimum of an attribute, while in triangle and duo-trio subjects are assumed to compare distances between pairs of percepts (Ennis, 1990). In case there is a known sensory attribute that differs in samples, n-AFC is more efficient and powerful than the overall discrimination test like triangle and duo-trio. However, when the sensory attribute(s) differing in the sample are unknown or there is a comparison for multidimensional stimuli, an overall discrimination test may be used since it is not necessary to state the sensory attribute or percept involved in making the

PAGE 39

26 discrimination judgment for this type of tests (Ennis, 1 990; Lawless and Heymann, 1998). DifTerence Threshold Psychophysical Theory Difference thresholds are also referred to as just noticeable differences fINDs) in the literature. There is at least one psychophysical law directly associated with difference thresholds: Weber's Law states that difference thresholds increase in proportion to the background intensity, e.g., AS = kS (2.1) where AS is a difference threshold, S is the intensity of a background stimulus and A: is a Weber fraction. The value of k reflects the discriminability of closely spaced stimuli: the higher k, the lower the sensitivity. A typical Weber fraction for taste (salt) is 0. 14, while that of smell is 0.24. Weber's Law is not always true, but it is good as a baseline to compare performance and as a rule-of-thumb. Weber's Law often fails near absolute threshold. A modified version of Weber's law is as follows: AS = k(S + c) (2.2) where c is a constant, usually small that represents a baseline level of stimulus that must be surpassed (Baird, 1997). DifTerence Thresholds of Mixtures For chemical senses (taste and smell), the perception behavior of a mixture can be very complicated (Gregson, 1984; Fritjers, 1987; Gregson, 1992). No study was found in the difference thresholds detection for mixtures. A related study found was by comparing the ratings of non-trained panelists, the assessment of the model that described overall similarities between binary odor mixtures (Gregson, 1984).

PAGE 40

27 Neural Networks Introduction An artificial neiiral network, inspired by the way in which the brain works, resembles the brain in two respects: (1) Knowledge is acquired by the network through a learning process and (2) Intemeuron connection strengths, known as synaptic weights, are used to store the acquired knowledge (Haykin, 1999). A neural network has the ability to learn and therefore generalize. In other words, it can build up an unplicit model based directly on real-life data and then uses the implicit model to predict the unknown cases. There are two main elements in a neural network: synapses (connections) and neurons (nodes). The way in which the neurons of a network connect with each other is called its topology or architecture. There are three fundamental classes of network topologies: single-layer feedforward, multilayer feedforward and recurrent network. The details of network topology include the number of layers, the number of nodes in each layer and full or part connection of the network. A learning algorithm refers to the welldefined rules of how to adjust the synaptic weights and bias of neural networks according to real-life data. The design of a neural network may proceed as follows: First is the learning step an appropriate architecture is selected and real-life data are used to train the network by a suitable learning algorithm. Second is the generalization step the recognition performance of the framed network is tested with data not seen before. Network size, e.g., the number of parameters adjusted by using the learning algorithm, affects the learning and generation ability of the network. The larger the network, the better the fitting to the learning data provided that enough data are available

PAGE 41

28 to train the network. Too large a network, however, cannot generalize well because the fit is too specific to the training-set data. So, the neural network should be sufficiently large to solve the problem, but not larger (Principe et al, 1999). It is ideal to randomize the order of presentation of training examples fix)m one epoch to the next. This randomization tends to make the search in weight space stochastic (involving chance or probability) over the learning cycles (Haykin, 1999). Multilayer Perceptrons (MLPs) Trained by Back-Propagation (BP) The perceptron, consisting of a single neuron with adjustable synaptic weights and bias, is the simplest form of a neural network. It can be used for the classification of patterns that are linearly separable (i.e., patterns that lie on opposite sides of a hyperplane) (Haykin, 1999). Multilayer Perceptrons (MLPs) extend the perceptron with hidden layers. Typically, a MLP is constituted with an input layer, one or more hidden layers of neurons, and an output layer, as shown in Figure 2-2. The neurons in both hidden layers and output layers have computational ability. The input signal propagates through the network on a layer-by-layer basis and in a forward direction. The computing power of the MLP lies in its hidden neurons that act as feature detector. Muhilayer Perceptrons are regularly trained in a supervised manner with a highly popular algorithm known as "back-propagation." Basically, back-propagation consists of two passes through the different layers of the network: a forward pass of signals and a backward pass of the local error. This algorithm can be applied independent of the topology of the network and the input dimensionality. It adjusts the synaptic weights and biases in five steps: 1. Initialization. When there is no prior information available (regularly the case), randomize the synaptic weights and biases. These synaptic weights

PAGE 42

or biases should forai a uniform distribution whose mean is zero and whose variance is chosen to make the standard deviation of the induced local fields of the neurons lie at the transition between the linear and saturated parts of the sigmoid activation function. 2. Presentation of Training Examples. Present the network with an epoch (batch) of training set. The sequence of forward and backward computations, as described under 3 and 4 respectively, is performed for each example in the set. 3. Forward Computation. The signals of an example in the training set are applied to the input neurons of the network, and its effect propagates through the network layer by layer. A set of outputs is then produced as the actual response of the network. The synaptic weights / biases are all fixed during the forward computation. 4. Backward Computation. The actual response of the network is subtracted fi"om a desired (target) response to produce the error signal. The error is propagated and scaled back by the chain rule. The synaptic weights are adjusted to make the actual response of the network move closer to the desired response in a statistical sense. 5. Iteration. Iterate the forward and backward computations imder 3 and 4 by presenting new epochs of training examples to the network imtil the stopping criterion is met. MLPs are universal approximators. They can be used for both pattern classification and fiinction approximation. The difference between using an ML? for fimction

PAGE 43

30 approximation and for classification is that for function approximation the output neurons are linear, while for classification the output neurons must be nonlinear. It has been mentioned that the feature detection ability lies in the hidden layer. Provided enough neurons are available in the hidden layer, any continuous fimction can be approximated by the MLP topology. Two-hidden-layers are necessary when the fiinctions or patterns to be modeled are discontinuous. There are rare applications that need more than two hidden layers (Principe et al., 1999). Input Hidden On^ut layer layer layer m p /I Figure 2-2. Multilayer Perceptrons (MLPs) with one hidden-layer Time-Delay Neural Network (TDNN) In temporal processes such as speech signals or fluctuations in stock market prices, the measurements from the world are fiinctions of time. For a neural network response to the temporal structure of information-bearing signals, it must be given memory. Memory may be divided into "short-term" and "long-term" memory. Lx)ng-term memory is stored in the synaptic weights of the network through supervised learning. Short-term memory, on the other hand, can be incorporated into tiie structure of a neural network through the

PAGE 44

31 use of time delays, which can be implemented at the synaptic level inside the network or at the input layer of the network. Memory structures are sensitive to the sequence of information presentation. By embedding memory into the structure of a static network such as an ordinary multilayer perceptron, the output of the network becomes a function of time. This approach for building a nonlinear dynamical system provides a clear separation of responsibilities: the static network accounts for nonlinearity, and the memory accounts for temporal effects. Figure 2-3 shows a diagram of the simplest and most commonly used form of short-term memory called tapped delay line memory. The delay imit, denoted by z\ is a linear system that delays the input signal by one time unit. The tapped delay line is buih from a cascade of these delay units. Sissial ox(u) UiiiH Uuit2 Uuit p 6 x(u) --1 6 x(n-l) o .x(u-2) 6 x(n-p+l) o x(u-p) Figure 2-3. Tapped delay line with p delay units The Time-delay neural network (TDNN) is a multilayer feedforward network with embedded local memory (tapped delay line) in both input and hidden layers. As shown in Figure 2-4, all the taps of the tapped delay lines are connected to the neurons of the next layer. A TDNN was first described by Waibel et al. (1989) and was devised to capture explicitly the time symmetry in the recognition of phonemes. In this study, a TDNN with two hidden layers was used to recognize three isolated words: "bee", "dee", and "gee" and achieved an average recognition rate of 98.5%.

PAGE 45

32 Figure 2-4. Paradigm of Time-delay Neural Network with one hidden-layer It should be noted that for TDNNs, not only the inputs are time signals, but also the desired responses. The problems of designing TDNN topologies are the same as for the MLP, with the addition of choosing the size of the tapped delay line (also called the memory layer). The size of memory layer depends on the number of past samples that are needed to describe the time structure of the inputs. If the size is too small, there may not be enough data from the past and the performance of neural network will suffer. On the other hand, if the

PAGE 46

33 size is too big, there will be many more weights in the network, resulting in slower training. In the TDNN, the size of memory layer has to be physically modified. Since the tapped delay line at the input does not have any free parameters, in the TDNN static back-propagation algorithm still can be used to train the network (Principe etal, 1999). Softmax It is required that each output is between 0 and 1 and all outputs sum to 1 in order to be able to interpret a MLP's outputs as posteriori probabilities (Bihop, 1995). This can be implemented by utilizing an output neuron with softmax activation fimction: _ exp(»e/J where the denominator sums over all network outputs. The softmax fimction is similar to the sigmoid neurons, except that the outputs are scaled by the total activation at the output layer. The softmax is effectively a competitive structure that forces the sum of the outputs of the neuron to be 1 .

PAGE 47

CHAPTER 3 OBJECTIVE The overall objective of this project was to develop a quantitative method to quickly predict the mixing compositions of a ternary spice mixture by using electronic nose as the measurement tool, and neural network/multivariate statistical methods for data analysis, and to compare the testing results from the electronic nose with that from GC and sensory analyses. The specific objectives of the project were: Electronic Nose Analysis To obtain sufficient e-nose response data on various spice mixtures to perform data analysis and to develop statistical/neural network data analysis procedures. The developed procedures are capable of predicting the mixture compositions of unknown spice mixtures (testing samples) with acceptable accuracy. GC Analysis To identify and quantify the volatile compounds in spice mixtures, and to determine the accuracy and efficiency of GC methods in predicting spice mixture compositions, and to compare e-nose and GC's prediction accuracies and the time/effort necessary to obtain these accuracies. Sensory Analysis To assess difference thresholds of the spice mixtures, and to determine whether the e-nose can predict the mixture compositions with an error below the perception threshold of human subjects. 34

PAGE 48

CHAPTER 4 MATERIALS AND METHODS Materials Three ground spices: basil {Ocimum basilicum), cinnamon {Cinnamomum cassia) and garlic (Allium sativum) were selected to formulate the spice mixtures. This selection was based on the criteria: 1) the major volatile components of the three spices should be different from each other and should be easily separated by GC, and thus facilitates the mixture analysis by GC, 2) the odor of the three spices should be significantly different from each other and should be pleasant to smell, which facilitates the sensory analysis using human subjects. As shown in Table 4-1, the major volatile components of basil, cinnamon and garlic were different. Their Kovats retention indices are also different enough to allow acceptable separation by a DB-5 column (Wijesekera, 1978; Yu et al, 1989; Yu et al, 1993; Vemin et al, 1994; Yu et al, 1994; Adams, 1995; Kim et al, 1995; Marotti et al, 1996; Miller et al, 1996; Lachowicz et al, 1997; Antonelli et al, 1998; Uhl, 2000; Acree and Am, 2001). Preliminary studies with human subjects showed that the three spices smelled significantly differently and had pleasant odors whether in pure form or in the formulated mixtures. Basil was purchased from Cibolo Junction Food & Spice Company (Albuquerque, NM). Ciimamon and garlic were purchased from Frontier Natural Products (Norway, lA). Each spice was sealed in a large glass container and stored imder refiigeration (3°C) for fiirther use. To eliminate the effects of different humidity on e-nose sensors, the three spices were individually equilibrated with saturated potassium carbonate solution at room 35

PAGE 49

36 temperature (21°C to 22'^C) for 20 hours to reach the same relative humidity (43.2%) before any analysis. For GC and sensory analyses, these spices with adjusted humidity were used to ensure a fair comparison among the three methods. Table 4-1 . The chemical names and their corresponding CAS# and KI (for DB-5) of major volatile compounds of basil, cirmamon and garlic Spice Name CAS# Chemical Name KI (DB-5) Basil 78-70-6 Linalool 1100 Basil 140-67-0 Estragole N/A Basil 97-53-0 Eugenol 1356 Basil 470-82-6 1,8-cineole 1030 Cinnamon 14371-10-9 trans-Cirmamaldehyde 1266 Cirmamon 103-54-8 Ciimamyl acetate 1443 Cinnamon 140-10-3 trans-Cinnamic acid 1438 Cinnamon 100-52-7 Benzaldehyde 968 Garlic 2179-57-9 Diallyl disulfide 1320 Garlic 592-88-1 Diallyl sulfide 1085 Garlic N/A Methyl allyl disulfide 1149 Experimental Designs and Methods Mixture Experimental Design of Three Components With three components, the simplex space of the mixtures is an equilateral triangle. The fi-actions of three components can be either weight or mole based, and are usually denoted by xi, X2, and xs, respectively. The compositions of any mixture with three components 1, 2 and 3 are represented as (x/, X2, X3). As shown in Figure 4-1 part A, the vertices of the triangle represent the single-component mixtures and the internal points of the triangle represent mixtures in which none of the three components are absent. The way to identify the fi-actions of a mixture in the simplex space was illustrated in Figure 41 part B. Suppose O was an arbitrary point in the simplex space. To get the value of x/, Ime EF was drawn parallel to the triangle boundary line BC. AH is a line perpendicular to

PAGE 50

37 BC. The value of jc/ is equal to the ratio of the length of segment GH to that of AH. The value X2 and X3 can be detennined in a similar way. components vj components (1,0,0) (0-5,0.5,0) (0,1,0) 5 C component 1 component 2 component 1 component 2 Figure 4-1. The simplex factor space with three components. A) The vertices of the triangle represent the single-component mixtures; the points in the boundary line are the mixture of two components; and the central point represent equal jfractions of the three components. B) For an arbitrary point O in the simplex space, the value of is equal to the ratio of the length of segment GH to that of AH while line EF was parallel to line BC and AH is a line perpendicular to BC. There are three basic mixture design methods: simplex-lattice, simplex-centroid and axial design, as shown in the parts (A), (B) and (C) of Figure 4-2. Each dot in the figure indicates a design point. The simplex-lattice and component simplex-centroid designs are boundary designs in that most of the design points, except the overall centroid, are on the boundaries of the simplex space. Axial designs, on the other hand, are designs where most of the points are positioned inside the simplex space. Some designs could be a combination of the three methods, as shown in part (D) of Figure 4-2 (Cornell, 1990).

PAGE 51

38 Figure 4-2. The three basic design methods in the simplex space with three components: (A){3, 3} simplex-lattice; (B) three components simplex-centroid, (C) axial design, and a combination of the three methods: (D) simplex-centroid augmented with 3 interior points. Electronic Nose To obtain sufficient data for prediction model building, the training points should have a satisfactory distribution throughout the experimental region and provide an internal estimate of the error variance. On the other hand, the testing points that test the performance of the model built should be randomly distributed in the experimental region. The selected spice mixtures for training/cross-validation and for testing are illustrated in Figure 4-3. The spice mixtures used for training/crossvalidation are symbolized by dots and numbered from 1 to 19; those for testing are symbolized by small triangles and denoted as IT through 5T. The fractions (based on weight) of spices

PAGE 52

39 corresponding to each design point were listed in Table 4-2 (training) and Table 4-3 (testing). The weight fraction of basil, cinnamon and garlic were represented by xj, X2 and X3, respectively, hi the testing spice mixtures, the weight fractions xi,X2, X3 were generated using the random function in Microsoft Excel using the following equations: X, = RAND{ ) x^=RAND{ )x(1-x,) 16 13 18 '3T 14' 1 BASIL 11 7 \l9 • 4T • • 10 12 • 15 CINNAMON Figure 4-3. The distribution of design points in the simplex space, with the dots, numbered from 1 to 19 representing the spice mixtures used for training/crossvalidation and the small triangles, denoted as IT through 5T, representing those for testing.

PAGE 53

40 Table 4-2. The weight fractions of design points for training/cross-validatio n Mix# xi (Basil) X2 (Cinnamon) xs (Garlic) 1 1.0 0.0 0.0 2 0.0 1.0 0.0 3 0.0 0.0 1.0 4 0.333 0.667 0 5 0.333 0.0 0.667 6 0 0.333 0.667 7 0.333 0.333 0.333 8 0.167 0.417 0.417 9 0.417 0.167 0.417 10 0.417 0.417 0.167 11 0.667 0.167 0.167 12 0.167 0.667 0.167 13 0.167 0.167 0.667 14 0.833 0.083 0.083 15 0.083 0.833 0.083 16 0.083 0.083 0.833 17 0.667 0.333 0 18 0.667 0.0 0.333 19 0 0.667 0.333 Table 4-3. The weight fractions of spices for testing Mix# xi (Basil) X2 (Cinnamon) Xi (Garlic) IT 0.218 0.015 0.767 2T 0.114 0.610 0.277 3T 0.829 0.047 0.124 4T 0.413 0.366 0.221 5T 0.052 ^.111 0.171 Spice mixtures were formulated in the weight fractions listed in Tables 4-2 and 4-3. Each experimental sample contained 10-gram formulated spice mixture, which was stored in a 70 ml weighting bottle (I.D.x H = 40 x 80 mm. Fisher Scientific, Fair Lawn, NJ). Eight samples were prepared for each training/cross-validation point and five samples for each testing point. The prepared samples were presented to an e-nose (e-Nose 4000, Neotronics, Gainesville, GA) with 12 conducting polymer sensors (sensor types: 483, 478, 464, 463, 462, 461, 460, 459, 458, 401, 298 and 297) that had been calibrated using 75% v/v propylene glycol water solution as recommended by the manufacturer

PAGE 54

41 (100% propylene glycol purchased from Fisher Scientific, Fair Lawn, NJ). The order of sample presentation was randomized. For each measurement, the sample chamber was purged by compressed dry air (BOC GASES, Murray Hill, NJ) for 4 minutes and the sensor chamber was purged 2 minutes to eliminate any foreign odor from environment or any residue odor from previous experiments. The flow rate of the compressed air (BOC GASES, Murray Hill, NJ) used for both chambers was 800 cm^/min. The sensors' responses were recorded every second up to 4 minutes. All the electronic nose experiments were run at room temperature (23*^C to 25''C). Gas Chromatography Simultaneous distillation-extraction (SDE) was selected to collect the volatile fraction of the spice samples for the following reasons: 1) SDE was the only method foxmd in the literature to deal with spice mixtures (Cheng et al, 1997). 2) Preliminary studies have shown that static headspace, Solid-Phase Microextraction (SPME) and direct solvent extraction methods were not suitable for the spice mixtures. The static headspace method did not have enough sensitivity for all three spices. For the SPME method, the fibers were overwhelmed by the volatile compoimds in cirmamon, resulting in lack of sensitivity towards the volatile compounds from garlic and basil. Significant amount of pigments, especially these from basil, were concentrated in the extracts by the direct solvent extraction method, making the GC tests difficult. 3) Both dynamic headspace and supercritical fluid extraction (SFE) are complex in operations and equipment. Dynamic headspace methods may

PAGE 55

42 have the same problem as SPME methods, while SFE may have the same problem as the direct solvent extraction method. For the SDE method, the assumption was that the volatiles were proportionally extracted and concentrated for GC analysis (Chien 1985; Lawrence and Shu, 1993; Cheng et al, 1997). Based on the above assumption and the objective of GC analysis (to compare the accuracy of the GC to predict the spice mixing fractions with that of the enose), the GC analyses were only carried out for the pure spices and the testing mixtures shown in Table 4-3. The solvent employed for SDE analysis was the pentane-diethyl ether (1:1 mixture), which was lighter than water. Both the n-pentane and diethyl ether were HPLC grade (Fisher Scientific, Fair Lawn, NJ). The apparatus (#523010, Kimble/Kontes, Vineland, NJ) used for SDE was similar to that described by Likens and Nickerson (1964) and was illustrated in Figure 4-4. The vapor of the solvent entered into the condensation chamber (composed of three cooling jackets/condensers running ice-cold water) through arm C, while the steam and the volatiles from spices entered into the chamber through arm D. The volatiles were extracted by the solvent at the large condensation surface, and most of the volatiles were transferred to the solvent phase and returned to solvent flask through arm A. The water phase with limited volatile components had to pass through the solvent frap in arm A, where more volatiles were transferred into the solvent phase. The water phase then returned to the water/spice flask through arm B. This apparatus provided continuous fresh solvent for volatile exfraction. By carefiilly controlling the temperatures of both solvent and water/spice sides, the volatiles were transferred and concentrated in the solvent flask.

PAGE 56

43 Hot plate Stirrer Figure 4-4. The apparatus used for simultaneous distillation extraction.

PAGE 57

44 As shown in the Table 4-4, preliminary results from spice mixtures with equal proportions of basil, cinnamon and garlic showed that more than 94% and 98% of total volatiles were extracted after 1 .5 and 2 hours, respectively. Table 4-4. The changes of the percentage of extraction of the total volatiles from the spice mixtures with equal proportions of basil, cinnamon and garlic versus distillation time using SDE Extraction Time (hours) Percentage of extraction (%) 1 90.9 1.5 94.5 2 98.5 2.5 98.8 3 99.6 3.5 99.9 In this study, the SDE procedure was carried as follows: The spice sample (5 grams) was put into a 500ml round flask and mixed with 100ml distilled water; Solvent (pentane-diethyl ether 1:1, 5ml) was added to the solvent flask; Distilled water (1 7ml) and solvent (1 1ml) were added into arm A and arm B, respectively; Temperature of water bath was brought to 50°C and the temperature of water bath was then controlled at 50 + 0.5°C; After 3 minutes, heating of the water/spice flask was begun. The heating and stirring rate were carefiilly controlled to keep the flask in boiling condition (lOO^C); Started to count distillation time after the condensation of the steam/volatile showed in the condensation chamber. This typically took 15 minutes after previous step; Operated in the conditions as described above for 2 hours;

PAGE 58

45 Stopped the heat for the water/spice flask; Turned off the heat for water bath after 20 minutes. By using ice, the temperature of water bath was then lowered to O-S'^C; Waited additional 10 minutes; Immediately transferred all the extracts in the solvent flask to a calibrated vial and adjusted the volume to 5ml; Added 5\i\ decane (Sigma-Aldrich , St. Louis, MO) into the extracts as a standard; The vial was capped and stored in refrigerator until use. The extracts were introduced by a 1 jal syringe (MICROLITER® #7101, Hamilton Co., Reno, NV) into the injection port of the GC (Shimadzu GC 14A, Norcross, GA). The separation of the volatiles was carried out on a DB-5 column (length: 30m; I.D.: 0.25mm; film thickness: 0.25^m; J&W Scientific, Folsom, CA) for a total of 30 minutes. The separated volatiles were detected by a flame ionization detector (FID). The temperature profile used for column was as follows: after 3 minutes holding at 30*^C, the temperature was increased fi-om 30°C to 150''C at lO^C per minute, and then increased fi-om 150*'C to 225^C at 5°C per minute. The temperature of the injection port was set as 250^C, while that of the FID detector was set at 260*^C. The carrier gas used was hydrogen that generated by Hydrogen Generator 9200 (Packard Instrument Co., Meriden CT) and the pressure for the carrier gas was set at 0.5kg/cm^. For the FID in GC, the air (HOC GASES, Murray Hill, NJ) pressure was set at 0.5kg/cm^, while the pressure for hydrogen was set at 1 .Okg/cm'^.

PAGE 59

46 Three replicates were performed for each pure spice, and the extract from each of the three replicates was injected into GC twice (3 replicates x 2 injections for each pure spice). Two replicates were performed for each of the spice mixtures listed in Table 4-3. The extract from the spice mixtures was also injected into GC twice (2 replicates x 2 injections for each mixture). For each of the three pure spices, the retention times for five biggest unique peaks were recorded. The peak areas under these retention times (a total of 5 x 3 retention times) were recorded for all injections, including those from the pure spices and those from the spice mixtures. For each injection, the peak area of the added standard (decane) was also recorded. Volatile components in the extracts were identified by either comparing their calculated Kovats retention indices (RI) to those from literatures or by adding the known chemical standards to the extracts. If the added chemical standard increases the peak area of a volatile, that volatile can be assimied to be identical to the chemical standard. All the chemical standards were purchased either from Acros Organic (Morris Plains, NJ) or Sigma-Aldrich (St. Louis, MO). For a certain volatile compounds, the Kovats retention index were calculated by formula 4-1 : C + (rt-rtr ) lU = mx 5^ ^ (4.1) where if/ is the Kovats retention index, rtci and rtc2 are the retention times of two consecutive n-paraffin hydrocarbons that bracket the retention time (rt) of the investigated compound, with rtci less than rt and rtc2 larger than rt. The value of "C is

PAGE 60

47 determined by counting the number of carbons in the molecule of the n-paraffin whose retention time is rtci. The retention times of the n-paraffin hydrocarbons in the GC set were obtained by injecting the C7-C20 mixtures (Supelco, Bellefonte, PA) into the GC. Those retention times are listed in Table 4-5. Table 4-5. The retention times of n-paraffin hydrocarbons (C7-C20) Standards Run 1 RI (minutes) Run 2 RI (minutes) Average RI (minutes) C7 4.002 3.991 3.997 C8 6.205 6.196 6.201 C9 8.498 8.493 8.496 CIO 10.626 10.62 10.623 Cll 12.506 12.503 12.505 C12 14.247 14.244 14.246 C13 15.898 15.895 15.897 C14 17.628 17.624 17.626 C15 19.474 19.469 19.472 C16 21.426 21.422 21.424 C17 23.445 23.446 23.446 C18 25.487 25.49 25.489 C19 27.514 27.517 27.516 C20 29.499 29.504 29.502 Sensory Thresholds As indicated in the literature review, sensory thresholds can be determined by combining the method of constant stimuli with the forced-choice discrimination methods. Among the discrimination methods, n-AFC, especially 2-AFC and 3-AFC are most commonly applied methods for sensory threshold detection (Baird, 1997). The benefits of n-AFCs over the overall discrimination tests such as triangle tests and duo-trio is that they are more powerful and generally more sensitive due to theusimplicity for subjects' performance, resulting in a lower level of the estimation threshold (Ennis, 1990; Lawless and Heymann, 1998). Yet the power and the sensitivity of the n-AFCs are gained through

PAGE 61

48 the directional comparison: the attribute that differs in the two samples vmder comparison is well defined and the panelists are well trained to compare the intensity of the attribute between samples. The benefits of n-AFCs are compromised in our case since the spice mixtures under comparison were different in the intensities of three attributes: the odor of basil, the odor of cinnamon and the odor of garlic. It is also difficult, if not impossible, to carry the n-AFCs methods in this study since they require intensive panel training to familiarize them with the three attributes. Triangular tests, which have a slightly higher power than duo-trio tests, were then selected in this study to assess the difference thresholds of spice mixtures. As mentioned in Chapter 2, the perception behavior for an odor mixture can be very complicated. It is difficult to define the difference/similarity between two spice mixture with different compositions. For simplicity, here the difference of any two mixtures was defined as the absolute average fraction difference, e.g., for mixture {xji, xi2, xii) and {x2i, X22, X23}, the difference is calculated by formula 4.2. difference = \'''' -^2.1 + ^2 -^221 + ^.3 -^^^I ^^qq^ ^42) To test similarity, triangle tests require at least 50 panelists (Meilgaard et al, 1999). The number of panelists required in each discrimination test limited the total number of the triangle tests performed. In order to effectively assess the range of difference thresholds over the whole simplex space, the mixture samples under comparison should be representative over that whole space. Additionally, it was reasonable to asses the difference thresholds around the e-nose testing points since the objective of the sensory tests was to determine if the e-nose could predict the mixing fictions with an error less than human perception and the e-nose prediction errors were based on the testing points.

PAGE 62

49 Based on the criteria mentioned above, the comparison points were selected. These points are illustrated in Figure 4-5. The spice mixtures used for testing in e-nose experiment were IT through 5T, as listed in Table 4-3, and lT-1, 3T-1, 5T-1, 4T-1, 4T-2 and 4T-3 were the points being compared to their corresponding sensory testing points. The difference between each of the two points under comparison was 6%, which had been shown to be difficult to discriminate in the preliminary study using 8 panelists. The mixing fractions of all the comparison points in Figure 4-5 are listed in Table 4-6. Garlic Basil Cinnamon Figure 4-5. The comparison points for sensory threshold testing, with IT through 5T were the spice mixtures used for testing in e-nose experiment and lT-1, 3T-1, 5T-1, 4T-1, 4T-2 and 4T-3 were 0.06 absolute average fraction difference from corresponding testing points Spice mixtures were formulated in the weight fractions listed in Table 4-6. Each experimental sample contained 1-gram formulated spice mixture, which was stored in a 22ml amber glass vial with aluminum-lined screw cap (Supelco, Bellefonte, PA). For

PAGE 63

50 each of the vials, the transparent part was well covered by aluminum foil to avoid giving any clue to the panelists about the spice compositions based on their visual characters. Table 4-6. The mixing fractions of the six pairs of spice mixtures compared by panelists Testmg Points XI X2 X3 Compared with XI X2 X3 4T 4T 4T IT 3T 5T 0.413 0.413 0.413 0.218 0.829 0.052 0.366 0.366 0.366 0.015 0.047 0.777 0.221 0.221 0.221 0.767 0.124 0.171 4T-1 4T-2 4T-3 lT-1 3T-1 5T-1 0.413 0.323 0.503 0.128 0.739 0.097 0.456 0.366 0.276 0.105 0.092 0.687 0.131 0.311 0.221 0.767 0.169 0.216 For the triangle tests, panelists were instructed to open the cap just before sniffing the sample and then close the cap once finished. The samples should be sniffed in an order of from left to right. Panelists was not allowed to reopen the cap to sniff when not sure which sample was the odd one to make sure that panelists sniffed only the saturated headspace of spice mixtures. No fraining was given and panelists were informed that they were evaluating spice mixtures of basil, cinnamon and garlic, and were presented samples that differed only in compositions of these three spices. The interval between two successive triangle tests were set as 2 minutes and panelists were asked to take deep breaths during the mterval to reduce carry-over and adaptation effects. In order to avoid general fatigue of the subjects, there were only two sections of triangular tests carried per day. Panelists were healthy adults selected from campus ranging in age from 20 to 53 and had no self-reported problem in their sense of smell. A total of 50 panelists performed the triangle tests on each of three consecutive days, hi the first day, the tests were performed to compare the pair of mixture IT and lT-1 and that of 3T and 3T-1. The pair of 5T and 5T-1 and the pair of 4T and 4T-1 were compared in the followmg day. The comparison between 4T and 4T-2, and between 4T and 4T-3 were carried out in the third day.

PAGE 64

51 There are a total of six possible sequences to present the odd sample (AAB, ABA, BAA, BBA, BAB and ABB). Compusensor® (Compusense Inc., Guelph, Ontario, Canada) was used to assign the six sequences to the two samples under comparison. Compusensor® also was used to generated random numbers to represent each sample and helped in collecting and recording data. Data Analysis Electronic Nose Three neural network methods were used to analyze data obtained from the e-nose experiment: Multilayer Perceptron (MLP), ML? with principal component scores as inputs (PCA-MLP) and Time-delay Neural Networks (TDNN). These are described in detail below. MLP Twelve inputs and three outputs were used in MLP. Each input corresponded to one of the twelve e-nose sensors' response. The sensors' responses used here were the instantaneous sensor responses at 4 minutes. Each of the three outputs indicated the mass fraction of a spice (basil, cinnamon or garlic) in the mixture. The values of desired outputs were set as the experimental fractions of the corresponding spice mixture, as those listed in Tables 4-2 and 4-3. The activation function of the output layer was a softmax, which can restrict each output to a value between 0 and 1 and makes them sum up to 1 . By using softmax activation fimction in the output layer, the mtemal constraint of the mixture experiments (the mass fractions of three spices must add up to one) was implemented. The training/cross-validation data set was partitioned into a training set and a crossvalidation set: 2 replicates from each spice mixture were randomly selected to build the

PAGE 65

52 cross-validation set, and the remaining 6 replicates were used as the training set. The order of replicates in the training set was randomized before presenting to a neural network to avoid training bias. The selection of the suitable MLP topology (the number of hidden layers and the number of neurons in each hidden layer) and parameters (weights and biases) was based on the performance of that topology/parameters toward the testing data. The smaller the error of the testing results generated from a set of MLP topology/parameters, the better that topology-parameter combination. The selection of topology/parameters proceeded as follows: First, a network topology was selected. This network was trained at least 10 times with random initial conditions based on how variable the testing results were, and the best testing resdt was recorded for that topology. Then, the network topology was changed, and the same training method was applied. This iterative method was continued until performance on testing data could not be improved. By the above procedure, the optimal topology-parameter combination of MLP was determined and the corresponding testing result was recorded as the performance of the network. In this study, the testing result of a neural network topology/parameters was obtained by averaging the five replicates of e-nose's responses corresponding to each testing point and then using these average responses as the inputs of that neural network. The error of a testing result was measured by mean square error (MSE). PCA-MLP This method was similar to that of MLP except that PC A (principal component analysis) was first applied to instantaneous sensor responses to reduce dimentionality of the original data. This resulted in a reduced number of inputs into the neural network. Then, principal component scores instead of original sensor responses were used as

PAGE 66

53 inputs to the neural network, and the number of inputs was equal to the number of principal components selected. TDNN This method was different from that of MLP in the following aspects: For the inputs, the time series sensor responses from 0 to 4 minutes instead of the instantaneous sensor responses at 4 minutes were used. There were embedded local memories (tapped delay line) in both input and hidden layers, which made the NN sensitive to the sequence of information in inputs. The three outputs that indicated the mass fraction of basil, cinnamon and garlic respectively were set as the experimental fictions of the corresponding spice in the mixture, regardless of the time at which the sensor responses were collected. Since the outputs of the NN were also time series data, the predicted mixture compositions for each spice were based on the average of the corresponding outputs over time. The randomized selecting and ordering of replicates was accomplished in Microsoft Excel spreadsheets, using either the random/sort fimction or customized VBA codes. PCA was performed using SAS® (SAS Institute Inc., Gary, NC), and neural networks were implemented using NeuroSolutions® (NeuroDimension Inc., Gainesville, FL). Gas Chromatography Since the concentration of decane (the added standard) in every extract was controlled (lul/ml), the relative amount of a volatile in an extract could be determined by dividing the corresponding peak area to the peak area of decane in the same injection.

PAGE 67

54 This adjustment could also eliminate the variability associated with injection and other GC operations. The GC analysis was then based on the relative amounts of volatiles within extracts. Two approaches were used to analyze the GC data. One was based on the relative amounts of the most abundant unique volatile of each spice, while another was based on the five most abimdant volatiles of each spice. Both of these two methods assumed that the relative amount of an extracted volatile was proportional to the amount of the volatile in the spice mixture. The calculation of both methods was constrained by the fact that the fractions of three spices in a mix add up to 1 . The two methods were referred later as "single volatile method" and "five-volatiles method." Single volatile method As illustrated in the Figure 4-6, this method started with the identification of the most abundant unique volatile in each of the three spices. These volatiles were labeled by their corresponding retention time h, tc and tg for basil, cinnamon and garlic, respectively. For all of the 3 replicates x 2 injections carried for the pure basil samples, the relative volatile amounts at tb were averaged and recorded as A bpThe same procedures were performed for the pure cinnamon and garlic seimples and resulted in the value of Acp and Aop. For each of the testing spice mixtures, the relative volatile amounts at tb, tc and tg that were obtained fi"om the 2 replicates x 2 injections were averaged and recorded asAg, Ac and Ac, respectively. The mixing flections of basil, cinnamon and garlic, symbolized by Xg , X(, and Jt^j respectively, were calculated using the equations listed at the right of Figure 4-6. It should be noted that the mixing fi-actions of basil, cinnamon and garlic should equal to As/ Abp, Ac /Acp and Ac/ Agp respectively based on the assumption that

PAGE 68

55 the relative amount of a volatile was proportional to the amount of the volatile in the spice mixture. These estimated mixing fractions were divided by the common factor (the sum of the estimated fractions) to satisfy the constraint that all three fractions should add up to 1 . From Training Points (Pure spic«): Relative amount of the most (3 ' epi'cate x 2 injection) abundant volatile in Basil Average Relative amount of the most <3 f»P"*<=a*« " 2 injection) abundant volatile in Cinnamon Average Relative amount of the most (3 replicate x 2 injection) abundant volatile in Garlic Average From Testing points (Spice Mixture) : Relative amount of ( 2 replicate » 2 injection) Relative amount of ( 2 replicate x 2 injection) Relative amount of ( 2 replicate x 2 injection) volaitile att. A BP at retention time t|, A CP at retention time tc Aqp at retention time tg Average J igp + — ^ hp Ab_ '^BP 'hp hp h 'hp Ab_ h ^BP hp hp Figure 4-6. The single volatile method for GC data analysis. Xg , X,^and X^ represented the estimated mixture compositions for basil, cinnamon and garlic, respectively Five-volatiles method As shown in Figure 4-7, this method was similar to the single volatile method except that the sum of the relative amounts of five most abundant unique volatiles were used to estimate the factions instead of the relative amount of the single most abundant unique volatile from each spice.

PAGE 69

56 From Training Points (Pure spice): Sum of relertive amounts (3 replicate " 2 injection) of five most abundant volatiles in Basil Average A BP at retention time tbi.t^j ti^ t|^ and t bS Sum of relative amounts (3 replicate « 2 injection) ^ cp at retention time of five most abundant > . \, . . . . volatiles in Cinnamon Average tc , t^. t«. tc«. and tcs Sum of relative amounts (3 replicate « 2 injection) of five most abundant volatiles in Garlic Average Agp at retention time From Testing points (Spice Mixture): Sum of volatile relative amount ( 2 replicate « 2 injection) at 'bi,fh2.*b3.*w.and tb5 Average Sum of volatile relative amount < 2 replicate « 2 injection) s* tel. te2, tt3, tel. and t«5 Average Sum of volatile relative amount ( 2 replicate « 2 injection) Average J -V„ = An + — ^ ^BP 'hp hp Jc hp Ab_ h + — ^BP hp hp io. hp Ab_ h ^BP hp hp Figure 4-7. The five-volatiles method for GC data analysis. , and Xq represented the estimated mixture compositions for basil, cinnamon and garlic, respectively Sensory Thresholds The estimation of difference thresholds was based on the criterion that the probability of correct responses should be 50% correct above chance when the difference of the two samples is at the threshold. The observed probability of correct responses from the triangle tests were converted to percent correct above chance by using formula 4-3. 100 -P. (4.3) {chance) where P(corr) equal to the percentage of correct response above the chance to a stimulus, P(obs) equal to the percentage of correct response given by subjects to that stimulus and P (chance) cqual to the percentage of correct response expected on the basis of chance alone.

PAGE 70

57 For triangle tests, the percentage of correct responses observed should be 66.7% to reach "50% correct above chance" since the guessing chance of the tests is 33.3%. In other words, if the triangle tests show that at a level of difference, the observed percentage of correct responses was lower than 66.7%, that level of difference could not be perceived by human subjects at a 0.50 probability.

PAGE 71

CHAPTER 5 RESULT AND DISSCUSION Electronic Nose Raw Data Appendix A lists the instantaneous e-nose data obtained for the purpose of training and testing. The time series data were too large to be listed. The original data are available in a CD and can be obtained from Dr. Murat O. Balaban at the University of Florida. Appendix B shows the organization and the path of these raw time series data files in the CD. A sample of the time series data is illustrated in Figure 5-1. The data were obtained as a replicate of mixture #12 (basil: cirmamon: garlic = 0.167:0.667:0.167). Atotalof2401mes of data over a period of 4 nunutes per replicate per mixture. Responses from 12 Sensors Fractions of each spice within the mijcture SI S2 S3 S12 Basil Cinnamon Garlic 0 0 0 0 0.167 0.667 0.167 0 0.01 0.01 0 0.167 0.667 0.167 0 0.04 0.04 c — 1 0 0.167 0.667 0.167 0 0.1 0.09 0.02 0.167 0.667 0.167 0 0.2 0.17 0.04 0.167 0.667 0.167 0 0.33 0.29 ^ .~ 0.08 0.167 0.667 0.167 0 0.5 0.43 0.14 0.167 0.667 0.167 0 0.7 0.6 0.21 0.167 0.667 0.167 0 0.92 0.79 0.29 0.167 0.667 0.167 0 1.12 0.97 0.39 0.167 0.667 0.167 0 1.32 1.15 0.49 0.167 0.667 0.167 0.01 1.49 1.33 0.59 0.167 0.667 0.167 0.01 1.64 1.48 0.69 0.167 0.667 0.167 0.167 0.01 1.77 1.62 0.78 0.167 0.667 1.05 6.72 6.06 3.55 0.167 0.667 0.167 1.05 6.72 6.06 3.55 0.167 0.667 0.167 Figure 5-1. A sample of time series data obtained 58

PAGE 72

59 Data Analysis Results Using MLP As mentioned in Chapters 1 and 4, MLP was applied to analyze the e-nose instantaneous data. One hidden-layer was selected for the MLP since two-hidden-layers are necessary only when the functions to be modeled are discontinuous and here sensors' responses toward mixtures were expected to be continuous. The nximber of neurons in the hidden layer was first set at 3, and the smallest prediction error of this topology on the testing set was 0.0084 (MSB). When the number of neurons increased to 8, the smallest prediction error obtained fi-om this topology was 0.0059 (MSB). With the number of neurons in the hidden layer being 4 and 5, the smallest prediction errors obtained were 0.0051 and 0.0050 in MSB, respectively. Based on above results and the criteria of network size selection (the neural network should be sufficiently large to solve the problem, but not larger), the optimal topology of MLP was determined to be one hidden layer with 12 inputs/3 outputs and 4 hidden neurons (Figure 5-2). The activation function of the hidden layer was sigmoid and that of output layer was softmax. This optimal MLP predicted the mixture compositions with an error of 0.0051 (MSB). Table 5-1 lists the prediction values for each spice at each testing point. Table 5-1 . The experimental and predicted mass fractions of spice mixtures predicted by Testing Points Bxperimental Weight Fraction Predicted Weight Fraction Average Absolute Prediction error Basil Cinnamon Garlic Basil Cinnamon Garlic IT 0.218 0.015 0.767 0.226 0.099 0.675 0.061 2T 0.114 0.61 0.277 0.116 0.625 0.259 0.012 3T 0.829 0.047 0.124 0.789 0.03 0.180 0.038 4T 0.413 0.366 0.221 0.349 0.407 0.244 0.043 5T 0.052 0.777 0.171 0.049 0.926 0.024 0.100 Mean Square Error (MSE): 0.0051

PAGE 73

60 Figure 5-2. The optimal topology of MLP Data Analysis Results Using PCA-MLP Having the potential to perform better than MLP, PCA-MLP structure was also applied to analyze the instantaneous e-nose data. It is important for this method to select suitable principal component scores as the inputs to a neural network. Khattree and Naik (2000) indicated that the amount of information contained in a multivariate data set can be measured in terms of total variance. Therefore, the selection of principal components is based on the amount of variance that each principal component preserves. The proportion and cimiulative proportion of total variance explained by principal component 1 through 10 were sequentially listed in Table 5-2. It could be found that first two principal components explain more than 99% of total variance. Based on this, only the first two principal component scores were selected as neural network inputs.

PAGE 74

61 Table 5-2. The proportion and cumulative proportion of total variance explained by principa component 1 through 10 Principal Component Proportion of Total Variance Explained Cumulative Proportion of Total Variance Explained 1 0.9850 0.9850 2 0.0066 0.9916 3 0.0032 0.9948 4 0.0017 0.9964 5 0.0013 0.9978 6 0.0007 0.9985 7 0.0006 0.9991 8 0.0004 0.9995 9 0.0002 0.9997 10 0.0001 0.9998 Due to the same considerations as those in the MLP method, one hidden layer was selected. Similar to the training procediire in the MLP method, the number of neurons in the hidden layer was first set at 3 and resulted in 0.0053 (MSE) as the smallest prediction error of this topology. Then, the number of neurons was increased to 8, and resulted in 0.0039 MSE as the smallest prediction error. The smallest prediction errors obtained were 0.0041 and 0.0043 in MSE for the networks with 4 and 5 neurons in the hidden layer, respectively. Therefore, the optimal topology of PCA-MLP was determined to be 2 inputs/3 outputs and 1 hidden layer with 4 hidden neurons (Figure 5-3), and the prediction values for each spice at each testing point are listed in Table 5-3. The activation function of the hidden layer was sigmoid and that of output layer was softmax. This optimal topology predicted the mixture compositions with an error of 0.0041 (MSE). Data Analysis Results Using TDNN TDNN was applied here to analyze the e-nose time series data. One hidden layer was used here for the same reasons as in MLP and PCA-MLP. It was experimentally foimd that the TDNN's performance (measured by MSE resulted from the testing set)

PAGE 75

62 was improved by averaging every four seconds' responses of e-nose sensors. This may be because the averaging procedure removed part of the noise from the time series data. The best performance of TDNN was obtained by using the time series sensors' responses collected between 2 and 4 minutes. This indicated that the time related infonnation was more abundant in the e-nose time series data form 2 to 4 minutes than that contained in the time series data from 0 to 2 minutes. The amounts of time information in the data from 0 to 2 minutes was small enough so that it could not counter the negative effects on the TDNN performance resulting from the increasing number of parameters, which was necessary to handle data in a longer time window. Table 5-3. The experimental and predicted mass fractions of spice mixtures predicted by the optimal performance of PC A-MLP Testing Points Experimental Weight Fraction Predicted Weight Fraction Average Absolute Prediction error Basil Cinnamon Garlic Basil Cinnamon Garlic IT 0.218 0.015 0.767 0.162 0.05 0.787 0.056 2T 0.114 0.61 0.277 0.08 0.629 0.291 0.034 3T 0.829 0.047 0.124 0.841 0.035 0.124 0.012 4T 0.413 0.366 0.221 0.28 0.498 0.222 0.133 ST 0.052 0.777 0.171 0.04 0.869 0.092 0.012 Mean Square Error (MSE): 0.0041 For the one-hidden layer TDNN, except for the weights and biases that can be automatically searched by back-propagation algorithm, there were several other parameters needed to be physically adjusted: the number of neurons in hidden layer, and the depth (number of the taps) and resolution (each tap delays the signal how many imits) of each memory structure. To search for the best combination of the above parameters, just one of these parameters was adjusted in every step imtil the best performance was found. It was found that the best TDNN performance predicted the mixture compositions with an error of 0.0035 (MSE). The prediction values for each spice at each testing point

PAGE 76

63 are listed in Table 5-4. The optimal architecture of TDNN, as shown in Figure 5-4, were described as follow: 12 inputs/3 outputs and 1 hidden layer with 10 neurons; The activation function of the hidden layer was tanh. The activation fiinction of the output layer was softmax. The number of taps in the input layer was 6 with 1 tap delay. The number of taps in hidden layer was 3 with 1 tap delay. Table 5-4. The experimental and predicted mass fractions of spice mixtures predicted by Testing Points Experimental Weight Fraction Predicted Weight Fraction Average Absolute Prediction error Basil Cinnamon Garlic Basil Cinnamon Garlic IT 0.218 0.015 0.767 0.192 0.129 0.679 0.076 2T 0.114 0.61 0.277 0.092 0.585 0.323 0.031 3T 0.829 0.047 0.124 0.789 0.048 0.181 0.033 4T 0.413 0.366 0.221 0.323 0.418 0.258 0.06 ST 0.052 0.777 0.171 0.052 0.839 0.109 0.041 Mean Square Error (MSE): 0.0035 Garlic Figure 5-3. The optimal topology of PCA-MLP Discussion Based on the prediction MSE values, it could be concluded that in this study PCAMLP performed better than MLP, and TDNN performed better than PCA-MLP in predicting mixture compositions. Just like any other modeling structure, the generalization ability of a MLP model will deteriorate with increasing number of

PAGE 77

64 parameters, given the same amount of inputs / outputs information (Principe et al., 1999). PCA extracted most of the inputs' information (over 99%) in a reduced dimensional format (from 12 to 2 dimensions). Since the number of parameters necessary in a ML? structure was directly correlated with the dimensions of its inputs, the preprocessing of input information by PCA significantly decreased the number of parameters necessary in the MLP structure. That resulted in improved generalization ability and a better performance of the PCA-MLP structure over MLP. It was not surprising that TDNN provided the smallest prediction error since it captured the additional information about time related differences in sensors' responses. This was not available in the instantaneous data used in MLP and PCA-MLP. Figure 5-5 shows the distribution of the prediction points, including those from MLP, MLP-PCA and TDNN, in the simplex space. It could be noted that the predictions from all the three methods skewed toward ciimamon compared with the experimental spice mixture points. In other words, all these methods tended to predict higher cirmamon fractions within a mixture. Considering that cinnamon provided the strongest odor impact among the three spices, this skewed prediction might be due to the following two reasons: first, the e-nose sensors might be overwhelmed by the ciimamon volatiles during the "sniffing" process of e-nose operations; second, there might be ciimamon odor residues in the sampling or sensor chamber despite the e-nose cleaning procedure. To predict unknown mixture compositions, the training points for prediction model building shall cover the whole mixture space and should be evenly distributed within that space. It means that the model developed based on the training data can't be used to predict an unknown mixture whose compositions are not within the fraining mixture

PAGE 78

65 space. Evenly distributed training points will ensure a fair estimation of the parameters of the prediction model. Sensor Figure 5-4. The optimal topology of TDNN There are two main factors which influence how many training points will be necessary for prediction model building: the number of components being mixed and the complexity of the sensors' response to the mixture. The more components being mixed and the more complex the sensors' response, the more training points will be necessary. The number of samples necessary for each training point depends on the variability of the e-nose response to an identical mixture composition, the differences among the mixtxires with different compositions, and the target prediction accuracy of the mixture compositions.

PAGE 79

66 Garlic Basil Cinnamon Figure 5-5. The distribution of the prediction points, including those from e-nose MLP, MLP-PCA and TDNN analysis, in the simplex space For a ternary mixture system, the prediction of compositions is actually a twodimensional problem due to the built-in constraint of compositional data. Ideally, to solve this two dimensional problem, the e-nose response data should show the following structure: the first two principal components should explain more than 90% of total variance yet the second principal component should account for a significant proportion of total variability, say, more than 20%. Accordingly, if a mixture with four components is under investigation, the first three principal components are expected to explain more than 90% of total variance yet the third principal component accoimts for a significant proportion of total variance. The e-nose response data obtained in this study did not have these ideal properties. Among the selected first two principal components used in PCAMLP, the first principal component explained more than 98% of total variance (Table 5-

PAGE 80

67 2), resulting in the second principal component having a high noise to signal ratio. The ideal structure of the data will not only improve the prediction ability of the PCA-MLP structure, but also other prediction models applied, which include both the multivariate statistical and neural network methods. A proper sensor array design/selection will make the sensors' responses approach ideal conditions. It should be noted that neural networks have the advantage over the traditional statistical methods only when the investigated data structures are highly complex. In case the investigated data structures are simple, alternative statistical methods such as principal component regression or partial least squares may perform better than neural networks in a quantitative e-nose study. Gas Chromatography Volatile Components of Spices Table 5-5 shows the average relative amounts of all the volatiles identified in each of the three spices and their corresponding retention times. The selection of the five most abundant unique volatiles in each spice was based on the following criteria: first, in order to be identified as a unique volatile of a certain spice, the ratio of the volatile relative amount in that particular spice to those in the two other spices should be equal or larger than 15; second, the relative amount of the volatile should be as large as possible in that particular spice. The retention times of the five most abundant imique volatiles in each of the three spices were recorded and listed in the Table 5-6. In this table, the star signs under the columns heading with spice names mdicated the membership of these volatiles. The time window of each of these volatiles identified by GC-FID was its corresponding retention time plus/minus 0.05 minutes. Table 5-6 also lists the calculated retention index of all the

PAGE 81

68 volatiles and the literature values of some of the volatiles (Yu et al., 1989; Yu et al, 1993; Vemin et al, 1994; Yu et al, 1994; Adams, 1995; Kim et al, 1995; Marotti et al, 1996; Miller et al, 1996; Lachowicz et al, 1997; Antonelli et al, 1998; Acree and Am, 2001). Table 5-5. The average relative amoimts of all volatiles identified in each of the three spices and their corresponding retention times Volatile H Retention Time Average volatile relative amounts (minutes) Basil Cinnamon Garlic 1 7.537 0.002 0 0.034 2 9.795 0.006 0.227 0 3 11.295 0.041 0 0 4 11.858 0 0 0.062 5 12.178 0 0 0.406 6 12.527 0.641 0 0 7 13.285 0 0 0.232 8 14.281 0.547 0 0 9 14.672 0 0.364 0.014 10 15.518 0.043 25.174 0.010 11 16.075 0.072 0.064 0.685 12 16.981 0.292 0 0 13 17.463 0.478 0.989 0 14 18.474 0.242 0 0 15 18.586 0.021 0.332 0 16 19.732 0.003 0.331 0 17 20.178 0.039 0.759 0 18 22.562 0.184 0 0 Sum of relative volatile amounts in each of the three spices: 2.610 28.240 1.442 The five most abundant unique volatiles in the basil emerged in the minutes 1 1.295, 12.527, 14.281, 16.981 and 18.474, and were labeled as volatile 2, 5, 7, 1 1 and 12 respectively. As discussed in Chapter 4, there were two methods to determine the chemical identity of a certain volatile. One was to compare the calculated retention index with the value fi-om literature. Another was to add a known chemical standard into the sample, which resulted in an increased peak whose identity was the same as that of the

PAGE 82

69 added chemical standard. The first method was called the retention index method in this chapter, while the second was called the chemical standard method. Volatiles 2, 5 and 1 1 were identified as 1,8-cineole, linalool and eugenol by both the retention index and the chemical standard methods. Volatile 7 was identified as estangole by the chemical standard method. Volatile 12 was unidentified. The five most abundant unique volatiles in the cinnamon emerged at 14.672, 15.518, 18.586, 19.732 and 20.178 minutes, and were labeled as volatile 8, 9, 13, 14 and 15, respectively. Volatiles 8, 9 and 13 were identified as geraniol, trans-cirmamaldehyde and cinnamyl acetate by both the retention index and the chemical standard methods. Volatiles 14 and 15 were imidentified. The five most abundant unique volatiles in the garlic emerged at 7.537, 1 1.858, 12.178, 13.285 and 16.075 minutes and were labeled as volatile 1, 3, 4, 6 and 10, respectively. Volatiles 4 and 10 were identified as diallyl sufide, and diallyl disulfide by both the retention index and the chemical standard methods. Volatile 6 was identified as methyl allyl disufide by the retention index method. Volatiles 1 and 3 were unidentified. The raw data obtained fi-om both the pure spices and the spice mixtures are Usted in Appendix C.

PAGE 83

70 l-H c o I 0^ (L> X 3 o 0< o CQ c c (U > o m o "o oooO'*ocNt^'^^-^>0'-Hm ooo>— ''^(N(N(Nmf«^-*-^in>o inoooor~in^(Noou-)^T}-vorsioo ONint^fNoooor^^t^oot^oorir^ (N00^>n(N(N'O>nO0N-^int^^ •^^jsirNimTj-Tj-invo^oooooNO — |(Nr«-i'a->nvot^ooON O — (N m Q i to < S ^ a 5i c3 u 1/5 3i jil — XI o op g 00 O to IS CO a ^ ^ o ^ -a S u w O
PAGE 84

71 Table 5-7 listed the average fraction (in percentage) of the volatiles (those identified in Table 5-6) in their corresponding pure spice extracts. It could be found that linalool was the most abundant component in the basil's extracts, accounting for 24.6% of the extracted total volatiles. Trans-cinnamaldehyde and diallyl disulfide were the most abundant components in the cinnamon and garlic's extracts, respectively. Transcirmamaldehyde accounted almost 90% of the total extracted volatiles from cinnamon. Diallyl disulfide accounted for nearly 50% of the total extracted volatiles fi"om garlic. Therefore, the unique volatiles selected for the single volatile method were linalool, trans-cirmamaldehyde and diallyl disulfide for basil, cinnamon and garlic, respectively. Table 5-7. The average fi-action (in percentage) of the volatiles in their corresponding pure spice extracts. Volatile # Retention Time (minutes) % in basil extracts % in cinnamon extracts % in garlic extracts 1 7.537 0.07 0 2.36 2 11.295 1.56 0 0 3 11.858 0 0 4.28 4 12.178 0 0 28.18 5 12.527 24.59 0 0 6 13.285 0 0 16.08 7 14.281 20.97 0 0 8 14.672 0 1.31 0.98 9 15.518 1.64 89.11 0.70 10 16.075 2.75 0.23 47.42 11 16.981 11.20 0 0 12 18.474 9.25 0 0 13 18.586 0.80 1.18 0 14 19.732 0.10 1.17 0 15 20.178 1.49 2.67 0 Sum of fractions of five most abundant volatiles in the spice extracts: 67.57 95.44 98.31

PAGE 85

72 Single Volatile Method The single volatile method predicted the mixture compositions with an error of 0.000755 (MSE). The prediction values for each spice at each testing point are listed in Table 5-8. Table 5-8. The experimental and predicted mass fractions of spice mixtures predicted by the single volatile method Testing Points Experimental Weight Fraction Predicted Weight Fraction Average Absolute Prediction error Basil Cinnamon Garlic Basil Cinnamon Garlic IT 0.218 0.015 0.767 0.233 0.016 0.715 0.023 2T 0.114 0.61 0.277 0.124 0.644 0.232 0.030 3T 0.829 0.047 0.124 0.816 0.063 0.121 0.011 4T 0.413 0.366 0.221 0.441 0.386 0.172 0.032 5T 0.052 0.777 0.171 0.064 0.779 0.157 0.009 Mean Square Error (MSE): 0.000755 Five-volatiles Method The five-volatiles method predicted the mixture compositions with an error of 0.00156 (MSE). The prediction values for each spice at each testing point are listed in Table 5-9. Table 5-9. The experimental and predicted mass fractions of spice mixtures predicted by the five-volatile method Testing Points Experimental Weight Fraction Predicted Weight Fraction Average Absolute Prediction error Basil Cinnamon Garlic Basil Cinnamon Garlic IT 0.218 0.015 0.767 0.224 0.018 0.758 0.006 2T 0.114 0.61 0.277 0.113 0.682 0.204 0.049 3T 0.829 0.047 0.124 0.835 0.069 0.096 0.019 4T 0.413 0.366 0.221 0.434 0.407 0.159 0.041 ST 0.052 0.777 0.171 0.060 0.817 0.123 0.032 Mean Square Error (MSE): 0.00156

PAGE 86

73 Discussion Figure 5-6 shows the distribution of the prediction points, including those from single volatile and five-volatiles methods, in the simplex space. Similar to those from enose analysis, the predictions resulting from the GC methods tended to predict more cinnamon and less garlic in a mixture. The residues in the extraction-distillation apparatus might account for the prediction skew. The data in Table 5-5 demonstrated the existence of the residues. It was found from the table that there were minor amounts of the unique components from cinnamon and garlic in the other two spices. These residues were difficult to be avoided since the complex shape of the extraction-distillation apparatus made it very hard to be cleaned. Table 5-5 also showed that the volatile concentration in cinnamon extracts was much higher than that in basil and garlic (cinnamon:basil:garlic« 55:5:3). The abundance of cinnamon volatiles within the exfraction system also added the possibility of absorbed cinnamon volatile residues. It had been mentioned in the Chapter 4 that the two GC methods were based on the assumption that the relative amount of a volatile in extracts was proportional to the amount of the volatile in the spice mixture. To test this assumption in the single volatile method, the K value was defined here as the ratio of the relative amount of a volatile to the weight of the spice from which the volatile was extracted. If the above assumption held, the K values should be constant for basil, for cinnamon and for garlic regardless of the weight of the corresponding spice in the mixture extracted. The K values of the most abundant volatile in basil, cinnamon and garlic at different weights of the corresponding spice were calculated individually and are illusfrated in Figures 5-7, 5-8 and 5-9, respectively. It should be noted that since the total weight of the spice sample for extraction was fixed (5 grams), the frend of the K values versus the corresponding spice

PAGE 87

74 weights was the same as that of the K values versus the mixing fraction of the corresponding spice. Garlic Basil Cinnamon Figure 5-6. The distribution of the prediction points, including those from GC single volatile and five-volatile methods, in the simplex space It could be found from Figure 5-7 and Figure 5-8 that the assumption of constant K values was valid for the most abundant volatile in basil and in ciimamon except at very low weights. As discussed before, minor residues existed in the extraction-distillation apparatus. When the absolute amount of a volatile decreased, the ratio of the concentration of residue volatile to that from exfraction increased, resulting in increased K value. For the most abundant volatile in garlic, the assumption held with exception at the garlic weight around 1 gram.

PAGE 88

75 0.17 0.16 0.15 0.14 0.13 ^ 0.12 0.11 o o (0 CO o 3 CO > t 1 2 3 4 5 The weights of basil in the mixture (gram) Figure 5-7. The K values of the most abundant basil volatile at different weights of basil in the corresponding spice mixture (A S ^ ® (0 S a> c 3 n 5 4 1 1 — — 1 0 1 2 3 1 4 1 5 6 The weights of cinnamon in the mixture (gram) Figure 5-8. The K values of the most abundant cinnamon volatile at different weights of cinnamon in the corresponding spice mixture

PAGE 89

76 0.17 3 0.15 M '•B >» 0.13 0.11 o M 0.09 o 3 5 0.07 0.05 — 0 1 2 3 4 5 6 The weights of garlic in the mixture (gram) Figure 5-9. The K values of the most abundant garlic volatile at different weights of garlic in the corresponding spice mixture (0 0) o > > o « 0) 3 (0 > 0.5 0.45 c 0.4 *. 0.35 0.3 ^ 0.25 > 1 0 1 2 3 4 5 The weights of basil in the mixture (gram) 6 Figure 5-10. The K values of the five most abundant unique basil volatiles at different weights of basil in the corresponding spice mixture

PAGE 90

77 To test the same assumption of the five-volatiles method, the K value was then defined as the ratio of the relative amount of a group of volatiles, here the five most abundant unique volatiles of the corresponding spice, to the weight of that spice. The K values of basil, ciimamon and garlic were calculated and illustrated in Figures 5-10, 5-1 1 and 5-12, respectively. Similar to that of the single volatile method. Figures 5-10 and 5-1 1 show that the assumption of constant K values held for basil and for cinnamon except at very low weights. However, for garlic there was an apparent trend of increasing K values with the weight of garlic within the spice mixture (Figure 5-12). This may due to the absorption of garlic volatiles into the matrices of the other two spices. With the increasing amount of the other two spices, more garlic volatiles were absorbed, and thus less garlic volatiles were extracted by solvents. This resulted in increased K values with the garlic weight. By comparing Figure 5-12 to Figure 5-9, it could be found that the increasing trend was not apparent for the most abundant volatile in garlic (diallyl disulfide). It may be speculated that the absorption rates of the other volatiles in garlic were higher than that of diallyl disulfide. Theoretically, the five-volatiles method should perform better than the single volatile method since the calculation was multivariate based instead of univariate. However, based on the resulted prediction MSE, the single volatile method provided better prediction than the five-volatiles method. The MSE of the five-volatilse method was twice as big as that of the single volatile method. As discussed above, the assumption of a constant K value was the calculation base of both methods. For garlic, this

PAGE 91

78 assumption roughly held for the single volatile method but was not valid for the fivevolatiles method. This might explain the better performance of the single volatile method. CO o > > o « o 3 (0 > 8 ^1 0 1 2 3 4 5 6 The weights of cinnamon In the mixture (gram) Figure 5-11. The K values of the five most abundant unique cinnamon volatiles at different weights of cinnamon in the corresponding spice mixture y = 0.0243x+ 0.165 = 0.7839 1 1 2 3 4 5 6 The weights of garlic in the mixture (gram) Figure 5-12. The K values of the five most abundant unique garlic volatiles at different weights of garlic in the corresponding spice mixture

PAGE 92

79 Sensory Thresholds The raw data obtained from the triangle tests, which included the order of samples presented and the random code assigned to each sample and the panelists' responses, are listed in Appendix D. Forty-two panelists came on three consecutive days to perform the triangle tests. Five other panelists performed the tests in two days and the rest of the fourteen panelists performed the tests for one day. Table 5-10 lists the number of panelists correctly recognizing the odd sample for each of the six pairs of comparisons. The correct rates and the correct above chance rates were also calculated and listed in the table. It could be found that except for the comparison between IT and lT-1, the correct rates were at or lower than 33.3%, which is the guessing rate of a triangle test. Table 5-11 lists the number of correct responses necessary for the triangular tests to establish a significant difference between the two samples under comparison when the total number of the tests was 50 (Meilgaard et al, 1999). Based on these numbers, it could be concluded that except for IT and lT-1, the five pairs of spice mixtures were perceived as similar by the panelists. Compared Between Correct Response (Out of 50 tests) Correct Rate Correct above chance 4T 4T-1 17 34% 1% 4T 4T-2 16 32% -2% 4T 4T-3 12 24% -14% IT lT-1 30 60% 40% 3T 3T-1 12 24% -14% 5T 5T-1 14 28% -8% The difference threshold is the minimum difference required between two stimuli that will elicit a perceived difference with a specific probability. As mentioned in Chapter 2, the specific probability was regularly set at 50%. Table 5-10 shows that even for the

PAGE 93

80 comparison between IT and IT1, the correct-above-chance was lower than 50%. Therefore, it could be concluded that the 6% difference as defined in Chapter 4 was below the difference threshold and was not big enough to elicit a perceived difference with a probability of 50%. Table 5-11. The number of correct responses necessary for the triangular tests to establish a significant difference between the two samples under comparison when the total number of the tests is 50 Significance Level 10% 5% 1% 0.1% No. of Correct Responses 22 23 26 28 It should be noted that the number of correct responses for the comparison between IT and lT-1 was almost twice as big as that from the rest of the five comparisons. Figure 4-6 illustrates the distribution of the points under comparison in the simplex space. It was found that the difference between IT and other testing points (3T through 5T) was that IT was very close to the boundary of the simplex space, which indicated that the content of one of the three spices (cinnamon) in the mixture was very limited. Therefore, the task of discriminating between IT and lT-1 was similar to the task of comparing a twocomponent mixture with a three-component mixture. This was a much easier task compared with discriminating two three-component mixtures. The significant performance difference between the pair of IT and lT-1 and others could also be explained by Weber's Law. Weber's Law states that the difference thresholds increase in proportion to the background intensity, e.g., the perception is more sensitive at low background intensity. Here the background intensity could be taken as the cinnamon fraction within the mixture. When the cinnamon fraction was low enough, the difference between the spice mixtures was easy to discriminate.

PAGE 94

81 As mentioned in Chapter 4, on each of three consecutive days there were two sessions of triangle tests. Table 5-12 compares the panelists' performance between the tests carried first in that day and those carried second. The performance on the pair of IT and lT-1 was not considered in this comparison due to the unique characteristics of this comparison as discussed above. It could be found that there was no significant difference of panelists' performance (correct rate) between the tests carried first and those carried second. This indicates that the 2 minutes interval between two successive triangle tests was enough to avoid sensory adaptation or other negative effects that might compromise the discrimination ability of panelists. Table 5-12. The comparison of panelists' performance between the tests carried first in that day and those carried second Compared Between Correct Response (Out of 50 tests) Correct Rate Order of the tests carried in the corresponding day IT lT-1 30 60% 1 5T 5T-1 14 28% 1 4T 4T-2 16 32% 1 3T 3T-1 12 24% 2 4T 4T-1 17 34% 2 4T 4T-3 12 24% 2 Comparison between E-nose and GC/Sensory Methods Figure 5-13 illustrates the prediction accuracy of the three neural network methods by comparing them with the estimated sensory thresholds. It could be foimd that except for a few points, the predictions fi-om all three NN methods were within a range of 6% difference fi-om the real values. Since the 6% difference had been demonstrated through the sensory analysis to be lower than the sensory threshold of human subjects, especially in the central area of the simplex space, the prediction accuracies from all of the three enose data analysis methods were acceptable.

PAGE 95

82 Garlic >6% Difference from Desired Points (below the sensory thresholds) Desired Points Sensory Connparison Points Predicted Points from MLP using senors' reponses as inputs Predicted Pointd from MLP using PCA scores as inputs Predicted Pointed from TDNN using time-series data Basil Cinnamon Figure 5-13. Comparison between the prediction error of electronic nose methods and the sensory threshold of human subjects Table 5-13 compares the prediction accuracy and efficiency of the electronic nose and GC methods applied in this study. The efficiency was measured by the time to prepare and analyze an imknown mixture leading to the prediction of its compositions. The longer the time needed, the less the efficiency. As mentioned in Chapter 4, for the GC methods there were 2 replicates x 2 injections required to investigate the composition of a sample of spice mixture. The sample preparation time (the extraction-distillation procedure) for each replicate was 3 hours and for each injection the operation time of GC was 30 minutes. Therefore, the total time needed for preparing/analyzing a mixture for composition prediction was 3x2 + 0.5x4 = 8 hours. This estimation of time did not consider the GC cool down time, which was around 20 minutes and was necessary between injections. For e-nose methods, the time needed for preparing the sample was negligible. Since five replicates were obtained for each mixture composition prediction,

PAGE 96

83 and the operation time for a replicate was 10 minutes, the total time needed for e-nose method was a few minutes more than 50 minutes, e.g., less than one hour. It was found that although the GC methods predicted the mixture compositions more accurately, their efficiencies were much lower than that of e-nose methods. Table 5-13. Comparing the accuracy and the efficiency of GC and electronic nose methods Accuracy (Prediction MSE) Efficiency (Exp. Operation Time, hours) GC Single volatile method 0.000755 More than eight Five-volatile method 0.00156 E-nose MLP 0.0051 Less than one PCA-MLP 0.0041 TDNN 0.0035 Figure 5-14 illustrates the best prediction performances resulting from electronic nose analysis and that fi-om gas chromatography methods. It should be noted that the GC method predicted the mixture compositions more accurately when the point representing the desired mixture was close to the boimdary of the simplex space, while the e-nose method generated a more uniform error distribution. For the mixtures who represented points far away fi-om the simplex space boundary (2T and 4T in the Figure 6-2), the amounts of the prediction errors were similar for the GC and e-nose methods. In other words, when there are no dominant component(s) in a mixture, the e-nose method may predict the mixture's compositions as well as the GC methods. In addition to efficiency, the e-nose methods have two other advantages over the GC methods: the non-destructive sampling procedure making the on-line or near on-line quality/processing monitoring feasible, and the solvent-fi-ee analysis that reduces the concerns of inhalation of harmful chemicals during experiments.

PAGE 97

84 Figure 5-14. Comparison between the prediction performance of electronic nose and gas chromatography method

PAGE 98

CHAPTER 6 SUMMARY, CONCLUSIONS AND SUGGESTIONS FOR FUTURE STUDY This study investigated both the experimental and the data analysis methods in applying an electronic nose to predict the mixture composition of a ternary spice mixture. The developed experimental method provided sufficient data to perform further data analysis. Three different neural network structures: multilayer perceptron (MLP), ML? using principal components analysis as preprocess and time-delay neural network (TDNN) were applied to e-nose data analysis. TDNN achieved the best prediction on testing data. The relative accuracy and efficiency of the developed methods were determined by comparing them to those from the two traditional methods: GC and sensory analysis. In GC methods, the volatile components of the spice mixtures were extracted and then quantified by GC. The mixture compositions of the testing blends were predicted based on the amounts of the extracted volatiles from each spice. The testing blends used in GC analysis were the same as those in e-nose experiments. The sensory analysis was performed to estimate the difference thresholds of the ternary spice mixtures. Except for a few testing points, the three neural network models built from e-nose data predicted the compositions of the testing mixtures with an error less than 6%. It was concluded from the sensory analysis that the 6% difference was far below the difference threshold and not big enough to elicit a perceived difference when there was no dominant component(s) in the ternary mixture. The GC methods provided a more accurate but less 85

PAGE 99

86 efficient prediction. The experimental time required for GC methods was 8 hours, instead of 50 minutes for e-nose methods. In summary, the objectives of this study have been achieved. Both the appropriate experimental and data analysis methods were developed for an e-nose to quantitatively predict the composition of a spice mixture. The combination of these methods also provided significantly improved efficiency in composition prediction compared with the GC method yet resulted with acceptable prediction accuracy. Using the developed procedure, the mixture composition can be predicted in a near on-line speed (50 minutes for an unknown sample), which makes the procedure valuable in quality monitoring or process control. The e-nose methods can also be used as a fast screening tool for product matching or re-formulation to indicate the lack or redundancy of the ingredient(s) of interest. There were three quantitative data analysis methods developed in this study and all showed acceptable prediction accuracy. These methods could be used in other quantitative analyses with an electronic nose as the measurement tool. This study has successfully developed the procedure to predict the composition of a ternary mixture. It would be interesting to investigate the e-nose' s prediction ability toward the mixture with more components (equal or more than four). As discussed in Chapter 5, the data structures of the e-nose sensors responses are not ideal. Development and selection of suitable combinations of e-nose sensors will provide better prediction for mixture compositions. The GC data analysis methods used in this study were based on unique volatiles fi-om each spice component. It will also be interesting to investigate the data analysis method for GC in case that the mixture under study without unique major

PAGE 100

87 volatile(s). As that has been done in this study for the developed e-nose data analysis methods, sensory evaluation methods can be used to evaluate the relative accuracy of the developed GC data analysis methods. The sensory perception behavior towards mixtures is complex. Further investigation on this topic will provide valuable information for product matching or re-formulation.

PAGE 101

APPENDIX A INSTANTANEOUS DATA FOR E-NOSE ANALYSIS Tables A-1 and A-2 list the instantaneous data obtained from the e-nose experiments for the purpose of training and testing, respectively. In both tables, the numbers/letters of mixture number (Mix#) represent the same spice mixtures as those listed in Tables 4-2 and 4-3. The 12 sensors of the electronic nose were denoted as SI through SI 2. The sensors' responses were recorded imder the heading of each sensor. In the columns headed as Basil, Cinnamon and Garlic, the numbers under spice indicated the fraction of that particular spice within the mixture. 88

PAGE 102

89 nnamon Garlic o o o o o o o o o o o O O O O O o o NO NO NO NO b o o o o o o o o O o O o O O o o d d Basil m m m o o o o o o o o O o o o o o o o d d S12 2.76 2.61 2.61 2.67 2.82 2.71 2.77 3.93 3.78 3.94 3.67 3.78 3.75 3.89 3.72 2.77 2.86 2.69 2.96 2.93 2.91 2.77 2.77 3.52 3.63 r training vq m «n >n (N NO 00 NO m NO NO NO 2.52 2.37 2.47 2.37 2.36 2.35 (N 2.38 CN 00 NO CN 00 ON CJN NO m NO 2.22 m CN SIO ri 1.95 1.83 2.02 2.08 2.05 2.19 3.34 3.18 3.31 3.08 3.13 3.15 3.17 3.12 2.22 2.28 1— ( (N 2.31 2.33 m (N 2.14 2.05 2.92 3.02 lents foi 6S 00 NO vq ON oo oo 00 ON 00 2.99 2.83 3.01 2.84 2.83 2.82 2.93 2.86 1.96 2.04 1.85 2.11 2.11 2.08 1.91 1.86 2.63 2.74 obtained from e-nose experim S8 1.27 1.08 1.15 0.98 r— < 1.25 1.17 1.22 (N n CN ON 00 rn iri CN NO OO ON S7 2.22 1.97 2.04 1.77 2.15 2.26 2.05 2.27 3.71 3.46 3.74 3.59 3.51 3.56 3.79 3.58 2.33 2.46 (N CN 2.53 2.53 2.52 2.34 2.21 3.18 3.42 9S 4.57 4.24 4.37 4.56 4.72 4.67 6.95 6.58 6.98 6.67 6.58 6.57 6.77 NO NO 4.84 5.03 NO 5.12 5.16 5.16 r4.52 6.13 6.21 S5 NO
PAGE 103

90 3 arlic VO VO VO VO VO VO VO VO VO VO VO VO VO VO VO VO VO VO VO VO VO VO VO VO VO VO VO vq VO VO VO VO CO CO CO CO CO CO CO CO CO CO CO CO innamon G o o o o O o o c> o o o C> o O O d d d d d d d d d d d VO \^ r-~ VO VO VO VO VO VO VO VO VO VO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO O c> C> o o o o o o o o o o O o o d d d d d d d d d d d *<*^ r<-) fO r<^ ro m m r-> m CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO OQ O O O o O o O O o O o o d o o o o o o o o o d d d d vq vq (N "1 (N 00 »o 1—* *o as p 00 p 00 p as p VO p in in CO as CO VO CO 00 CN Os CO in CO rn CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO >0 n «n CO in in m V i CO CO VO VO VO VO i^ in VO CO VO i> «o CO 00 en 00 00 00 00 00 as 00 00 00 CO 00 00 o OV lO CO 51in (N
PAGE 104

91 O i u C/3 at PQ o c/3 ON 00 t/5 IT) CO !Z1 X m Nl N| N| NJ N| N| Nl N| NO NO NO NO NO NO o o o o o o o o O o o o o d d d d d d d d d d d d d m m m m m m m ro CO m •555NO NO NO NO NO NO NO NO o o o o o o o o o o o o o d d d d d d d d d d d d d m m m NO NO NO NO NO NO NO o O O O o o c> o o o o o d d d d d d d d d d d d d d 00 CM rn cn cn rn NO (N rn rn CN NO rn rtrn in rn NO rn rn ON CN ON CN NO m CN CN m 00 CN CN «n CO m m m m m m m m rn m rn m m m m ^ O 00 o ON o 00 o o On O ON o ON o in o NO o 00 o NO o o m o 00 O o m NO rCN (N cnI CN CN CN CN CN CN CN CN CN CN CN CN CN CN CN t — m m 00 in NO NO ^ m NO m 00 in ON in ON 00 CN 00 in 00 00 r~ 00 (N (N CN CN n ON in in in NO rn CN CN 00 rn rn ON rn ON rn rn NO in in ON 00 in 00 NO in Cn CN (N (N CnJ CN CN CN CN CN CN CN CN CN CN CN CN CN CN CN CN CN 00 t t — 1 — in I — On NO 00 CN On NO in 00 NO NO NO NO ON NO NO n NO NO On >n NO p 00 p NO ON CN p in ON CN p >o u-i in lO in in in NO iri in in in in in in in NO NO in NO in VO VO NO IT) NO NO ON NO m NO in in m NO m NO NO CN >n On m 00 in m On NO m NO 00 NO >n NO NO NO CN 00 On in m in 00 00 o m NO O o NO ON P 00 CN NO CN CN rn CN 00 CN CO \^ 1 CO rn CO m CO m m m m m m m m m m m oo T— < 00 00 NO OO On NO NO NO 00 in oo ON ri u-i in in in in in m' in in in in in n NO «n NO IT) m (N IT) r«-i
PAGE 105

92 <— <^'-<-^^^r— Ir— 11— I,— I,— I,— lrtrtrt^,_,_V0VOV0\0^^^^ oooooooc5ooc5ooc>oc)00ooc>oooo r~rr~ rr~ro »o so so so SO so so so so SO SO SO SO so so SO SO so o o o o o o o O o o o o d d d d d d d d o r— < so so so so so so so so so so so so so so so so so SO so SO vo SO SO so d d d d d d d d d d d d d d d d d d o o o o o o oooooooo so rn rn rn r-H IT) o iri >n 00 (N in OS 00 m (N rn m' rn cn rn TjlO OS (N OS OO OS OS n -<4>o m 00 00 OS 00 m OS SO OS (N 00 as OS OS SO so SO (N SO OS so (N SO SO (N n oo (N OS 00 (N m (N m (N m n m so so rn »— < ro OS fN CO (Z) tN CN (N (N (N n >n in >n in SO SO so so so SO so so >n in in m' n in in m in fN in r~ in in in ^ in o o ^ TjOS ^ ^ oo 00 so so'Oininmininmin rfOssosOOs-^^T}-^ •nsooosooosooosos OS fN — in m o^ooooooo o o 00 fN CO as fN OS fN rn m OS OS SO O OS O OS o so OS CO o OS OS fN r<^ rn cn m' fN CO CO c-i fN CO CO fN (N oo o so O so O ro so oo as fN «n >n oo
PAGE 106

93 O c o u (N o GO 0\ 00 C/3 CO C/J C/3 X 00 o m 00 o 00 o 00 o 00 O ro 00 O m 00 o 00 o 00 o CO 00 o CO 00 o CO oo o CO 00 o CO 00 o CO 00 o CO oo o CO CO 00 CO CO 00 CO CO 00 CO CO 00 CO CO 00 CO CO 00 CO CO 00 CO CO 00 o o o o o O o o o o o o o o o o o o o o o o o o o o .083 m 00 m 00 m 00 oo 00 m 00 m 00 m ro oo CO CO 00 CO CO 00 CO CO 00 CO CO 00 CO CO Art CO CO CO CO CO 00 CO 00 CO 00 CO 00 CO 00 CO 00 CO 00 o CO 00 o CO CO CO CO CO o o o o o o o o o o o o o o o o o o o o o o o d d m m 00 m 00 m oo ro 00 m 00 m 00 m fo 00 ro 00 m oo o CO 00 o CO 00 o CO oo o CO 00 o CO 00 o CO o CO o CO o CO oo o CO oo o CO 00 o CO 00 o ro 00 o CO oo o CO oo o NO NO NO o o o o o o o O o o o o o o o o o o o o o o o o d d CN 00 rjOs 00 ON 00 NO 00 CN On NO in NO NO CO NO >n 00 fN w> CN CN CO CN CN CN CN NO CN CN CO (N (N (N (N m CN CN CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO CO 00 00 >n r~; 00 00 m NO CN NO CN ON CN ^ CN NO fN CN m CO CN O CO p CN p CO o CN P 00 OS OS CO p CN CN CN CN CN CN CN CN CN CN CN CN CN CN CN CN On (N NO CO On O O 0\ OS N| CJn N On p NO NO NO 1 NO NO NO NO »n in >n 00 NO (N n «n OS NO NO NO CO NO OS OO r— t On m *o CN 5; CN 00 CN CO 1 1 CO CO CO VN n in n n «n' CN oo in r 00 OO NO t-H 00 On CO 00 fN NO >n NO NO OS in n CO in r~<3; OS CO OS NO n ^" NO no' NO no' NO no' NO NO >n >n n >n «n >n »n n 00 NO l> «o On >n 00 o in 00 CO o OO 00 >n 00 OS NO CN OS 00 "^ 00 OS o m n m NO NO ON NO CO o CO o in p P CN P p 00 p On OO 00 On On On OS 00 On 00 NO 00 00 00 o o o o o o o o O o O O O o O o d d in in »n in in >n >n m NO NO

PAGE 107

94 on Garlic o o o o o o 0.333 0.333 0.333 0.333 0.333 0.333 0.333 0.333 0.333 0.333 0.333 0.333 0.333 0.333 0.333 0.333 Cinnami 0.333 0.333 0.333 0.333 0.333 0.333 o o o o o o o o 0.667 0.667 0.667 0.667 0.667 0.667 0.667 0.667 Basil 0.667 0.667 0.667 0.667 0.667 0.667 0.667 0.667 0.667 0.667 0.667 0.667 0.667 0.667 o o o o o o o o S12 3.22 3.22 3.31 3.26 3.27 3.19 3.06 3.16 3.05 3.08 3.09 3.02 3.09 2.98 3.43 3.48 3.52 3.59 3.49 3.43 3.55 3.44 1.97 1.99 2.03 2.02 2.01 OS 00 00 so OS ro 00 NO 00 On so 00 00 >o 00 2.21 2.18 2.21 2.23 2.23 2.21 2.24 2.13 OlS 2.52 2.53 2.62 2.59 2.63 2.43 2.38 2.51 2.36 2.36 2.44 2.39 2.42 CN 2.92 2.91 2.92 2.94 2.99 2.93 OS CN 2.78 S9 2.26 2.28 2.34 2.32 2.32 o (N On 00 m 00 On so oo (N 00 OS 00 CN OS CO OS 00 00 00 S7 2.66 2.68 00 ri 2.79 2.85 2.97 2.57 2.56 2.44 2.44 2.55 2.56 2.62 so CN 3.13 3.17 3.21 CN 3.21 3.21 3.29 3.37 S6 5.38 5.41 5.57 5.56 5.65 5.67 5.17 5.27 5.05 5.02 r> GO m >o OS OS (N OS so 00 00 rrS4 2.96 2.93 3.01 3.04 3.07 3.16 2.77 2.87 2.78 2.79 2.84 2.79 2.81 2.86 3.31 3.25 3.31 3.31 ro CO en 3.26 3.33 S3 5.44 5.55 5.69 5.66 5.59 5.81 5.41 5.17 (N 5.25 5.26 5.25 5.24 5.99 5.93 6.01 so 5.96 6.02 6.04 6.01 S2 6.09 5.98 6.16 6.38 6.29 6.27 5.71 6.03 5.68 5.67 5.73 5.77 5.82 5.74 6.69 6.56 6.69 6.63 6.68 6.65 6.64 6.56 0.83 0.86 0.86 0.85 0.85 0.86 0.81 0.84 0.78 0.77 0.81 0.77 0.77 00 d 0.98 CN q CN q q CN q CN q Mix# 1^ 00 00 oo oo 00 00 oo oo OS On C3S On OS CTs

PAGE 108

95 C3 o; U 00 o C/3 e •c a X u n ON OO (N On VO ON CO ON ON (N (N (N (N (N (N 00 00 00 00 00 d d d d d NO NO O 00 o 00 OO 00 NO NO m NO NO NO NO cn NO NO m NO NO Z.Z.Z,' LLL Z.Z.Z, d d d d d o o o o o CO m ro (N >0 O (N o fN iri o o (N O d d d d d d d d d d oo rn vq NO r-i NO en rn rn on rn rn !i-ONONNor*i SSr^^EilCJlS'^'^ ,gNt^t^(N(Nr-rsi<^ romtN^r^inNOON r^u-^T};iOTr^--Hr^O-^^t~^OOON^ONOOONONTrvo«0-HS «ooooo(^ ON <0 oo NO On
PAGE 109

APPENDIX B TIME SERIES DATA FOR E-NOSE ANALYSIS As mentioned in Chapter 5, the data were too big to be Usted here; instead the organization of the raw data files in the prepared CD was presented. All training data were stored in the subdirectory: D:\dissertation\rawdata\e-nose\timeseries\training. There were a total of 19 folders in the above subdirectory, named by number 1 through 19. Each folder contains the time series data corresponding to the spice mixtures whose composition are listed in Table 4-2. There were eight replicates in each of the 19 folders. For example, file 1-6 stored the time series data fi-om the sixth replicate of the mixture 1. Similarly, all testing data were stored in the following subdirectory: D: \dissertation \rawdata \e-nose \timeseries \testing. A total of five folders exist in the "testing" subdirectory: IT, 2T, 3T, 4T and 5T. The compositions of the mixtures represented by IT through 5T are listed in Table 4-3. In each of the five folders, five replicates corresponding to the same mixture under the name of its folder are stored. For example, file lT-3 was the time series data obtained from the third replicate of the mixture IT. All the data files are in the format of Excel spreadsheets. There are a total of 15 columns in each data file. The first and third column records the temperature of the sensor chamber and the sample vessel, respectively. The second column records the 96

PAGE 110

97 humidity within the sensor chamber. Column fourth through fifteenth record the sensors' responses.

PAGE 111

APPENDIX C RAW DATA FOR GC ANALYSIS As mentioned in Chapter 4, there were 3 repHcates run for each of the three spices. Each extract obtained were injected into GC twice. Tables C-1 and C-2 list the raw GC data from the first and second injections, respectively. It should be noted that the total peak areas in both tables were the sum of all peak areas shown in the GC output excluding the peaks of the solvent and standard. Tables C-3 and C-4 list the GC data obtained from the spice mixture samples whose compositions were listed in Table 4-3. 98

PAGE 112

99 CM 1 a I o G O U o U CO G O U (N CQ I PQ 2 •2 g Si S Pi o 00 00 o o o o o o 00 ON o in o t^OVOVOOOOOOOOO^S 00 00 00 (N O O ON O VO 00 ON>nooooooo 00 ON m «n Tj^ ON n ON ON o ^ o — 1 m in (N o 1^ m (N fN CN in oo ON m in 00 m 00 00 On Cn) (N m 00 (N r^oo-^moooo ON in (N KO CnI VO ON in ^ _^ ^ ^ m VO in ON in vo ON rs m 00 o O 00 tN
PAGE 113

100 o I o U o CO CQ •2 g Oh o o in 00 00 0^ ON On oooo»ooooooooo>oi^ in ON m (N ON S o o 00 >n NO ^ 00 o ON S S °° g o in in in o o o o o On r00 ^ O O O (N — ' ON CN n 00 m O NO ° C3N 00 TlNO ON NO £ 22 o o CN NO LI; CN c*^ ^ ON 00 CN ON CN ON CN NO NO NO NO m NO O O CN O ON (N -"it m Tio o ^ o CN OS (T) in T}NO CN m CN 00 ON >n 00 On CN m — 1 ^ o § _ ^ ^ ^ CM O CN 2^ O O in rin m ON in r-; r-^ ON NO ON CN CN NO ON in CO in in o in NO CN ON _ CN <^o^2oi=;dNiS NO On «n ^ NO — CN — ^ ^ NO «n CN inoooot^in^cNooin-— im ONint^cNoooot^^t^ooNO cNoo^incNCNNOm" ~ — ^ CN CN m TjTjO ON in NO NO NO CN 00 CN 00 m NO NO Tt in ^ in ^• 00 00 ON O CN (N CN (Nro'^mNor^oooN Or— (NrO'^inNOt^ T3 oo -O a -a +-» o H

PAGE 114

101 O '04 > J3 o U E o X O X) U a 2 u -c fI u IT) fN •2 i Pi 1 0 ON ^ — ' — ' o >n VO 00 O O (N so so 0 0 SO ON (N 00 Vi 0 Os 0 (N ON (N 00 O O m O -H T}(N ° -H CN 00 r4 0 OS »— < 0 VO 00 ON 10 NO VO 35987 00 (N ON ON 00 ON VO ON SO (N ON m VO VO NO VO >o ro 00 0 fN NO (N n OS 0 0 (N ON n ON in VO n NO in 0 ON CN in 0 ON in VO >n VO 0 in 00 0 in 00 in NO CN CN VO CN in CN VO >n CN CN 00 NO NO ^ in CN CN m 00 m o O CN o o ^4NO ON VO 00 m in VO m 0 ON CN 00 CN NO ON CN ON ON 00 CN o in ON m 0 VO 16281 VO in ON in m ON ON 00 CN in ON ON 00 f)J o ^ o NO m 0 On 00 NO OS r-in 0 VO 0 m CN in 00 NO ON NO NO CO m NO ON ON CN o m cT) m as TlON 00 CN 00 CN CO CN 0 t-H •rt 00 ?^0OrNl CN CN in in ON VO 00 m 00 CN ON 00 m o NO VO O CN n 5 m >n m in in NO NO 0 NO 0 0 CN 00 r-H 00 in 0 CN m 0 m VO CN VO in 00 NO CN ON r~in rn ON in Os 00 00 in CN 00 in NO m CN 00 00 00 VO 00 00 in CN CN NO in 0 ON n VO VO 06 00 NO NO m ^ O CN CN CN — 'fNjr<^->4-«nvor^ooc3N'^'^'^'^^'^'^'^«o'2

PAGE 115

102 H =tt: H H o o 2: s S (N CN in lO m 00 in ON O fN in ^ in ^ 00 m ON in VO On o ON ON in 00 Ess 00 00 m (N IJ n ON " Pig Si ^ in m NO in in o o o in o O ON 00 in o o m 00 NO NO °° ^ NO CN NO NO in o O CN — CN CN — cNroTj-invor^-oooN o^cNmTi-invor--oo

PAGE 116

APPENDIX D RAW DATA FROM SENSORY ANALYSIS Table DThe triangle test results from the comparisons of the mixtures IT and lT-1 Triangle CjXp. ft Sample 1 Random i^ooe Sample2 Random LxOue Sample3 Random i^ooe Panelist's oeieciion Correct? 1 1 T 1 1 1 T 1 1 7ni /Ul 1 1-1 1 77 Iz/ -J J. 1 T 1 1 /Ul 1 T 1 11-1 oUo 1 T 1 1 J jj z J 1 T 1 11-1 1 "in Iz / 1 T 1 1 new /Ul 1 X 1 1 ^1 ^ J J J T J A H 1 T 1 11-1 1 T7 Iz / 1 T 1 11-1 oUo 1 T 1 1 701 /Ul -J J C J 1 T 1 11-1 ouo 1 T 1 1 /Ul 1 T 1 11-1 1 77 Iz / z 0 T T 1 1 J J J 1 T 1 11-1 ouo 1 T 1 11-1 1 77 111 1 1 7 1 T 1 1 1 T 1 1 7m /Ul 1X1 11-1 1 77 Iz/ -5 J o o 1 T 1 1 1 T 1 1 1-1 oUo 1 X 1 1 7fV1 /Ul Z o 1 T 1 11-1 Iz / 1 T 1 1 1 X 1 1 7f*1 /Ul Z in 1 T 1 11-1 1 07 Iz / 1 T 1 11-1 oUo 1 X 1 1 1 J 1 1 1 1 1 T 1 11-1 1 07 IZ / 1 T 1 1 j5j 1X1 1 1-1 OUO Z I"? iz 1 T 1 1 7ni /Ul 1 T 1 11-1 1 77 Iz / 1X1 11-1 ADA oUo 1 1 13 IT 701 IT 535 lT-1 606 J 14 IT 535 lT-1 127 IT 701 2 15 lT-1 127 IT 535 IT 701 3 16 lT-1 127 lT-1 606 IT 535 3 17 lT-1 606 IT 535 lT-1 127 1 0 18 IT 701 lT-1 606 lT-1 127 1 19 IT 535 IT 701 lT-1 606 3 20 IT 535 lT-1 127 IT 701 2 21 lT-1 127 IT 701 IT 535 2 0 22 lT-1 127 lT-1 606 IT 701 1 0 23 lT-1 127 IT 535 lT-1 606 1 0 24 IT 701 lT-1 606 lT-1 127 2 0 25 IT 535 IT 701 lT-1 127 3 1 103

PAGE 117

104 Table D-1 . Continued Triangle Exp.# Sample 1 Random Lode Sample2 Random Lode Sample3 Random Loue Panelist s Selection Correct? 26 IT ^ ^ 535 1 1 lT-1 606 1 1 701 1 A U 27 lT-1 606 IT 535 IT 701 1 1 1 28 lT-1 127 lT-1 606 IT 535 1 0 29 lT-1 127 IT 701 1 TP 1 lT-1 606 2 1 30 IT 701 lT-1 606 lT-1 127 1 1 31 IT 701 IT 535 lT-1 606 3 1 32 IT 701 1 TP \ lT-1 127 IT ^ ^ ^ 535 2 1 33 lT-1 127 IT 535 IT 701 3 0 34 lT-1 606 lT-1 127 IT 701 2 0 35 lT-1 127 I HP IT 701 lT-1 606 1 1 0 36 IT c c 535 lT-1 127 lT-1 606 2 0 37 IT 535 IT TA 1 701 lT-1 127 1 0 38 IT 701 lT-1 127 1 IT 535 1 0 39 lT-1 127 IT 535 IT 701 1 1 40 lT-1 127 lT-1 606 IT 701 2 0 41 lT-1 127 IT 535 lT-1 606 2 1 42 1 1 701 lT-1 127 lT-1 606 3 0 43 IT 535 IT 701 lT-1 606 2 0 44 IT 701 lT-1 127 IT 535 2 1 45 11-1 1 T7 127 IT TA 1 701 IT 535 1 1 46 lT-1 127 lT-1 606 IT 701 1 0 47 lT-1 606 IT 701 lT-1 127 2 1 48 IT 701 lT-1 606 lT-1 127 1 1 49 IT 535 IT 701 lT-1 606 3 1 50 IT 701 lT-1 606 IT 535 2 1

PAGE 118

105 Table D-2. The triangle test results from the comparisons of the mixtures 3T and 3T-1 1 X tailzie EXD # Sample 1 Ranfinm XVClXllXVJl 1 1 Code Sample2 Random Code Sample3 Random Code Panelist's Selection Correct? 1 3T •J 1 171 3T 343 3T-1 439 2 0 2 At 3T 171 3T-1 439 3T 343 3 0 3T-1 439 3T 343 3T 171 2 0 4 3T-1 X 1 439 3T-1 956 3T 343 1 0 5 3T-1 439 3T 343 3T-1 956 3 0 3T ^ X 171 3T-1 439 3T-1 956 2 0 7 3T -J X 171 3T 343 3T-1 956 2 0 fi O 3T ^ X 343 3T-1 439 3T 171 1 0 q 3T-1 ^ X X 439 3T 343 3T 171 2 0 10 3T-1 439 3T-1 956 3T 343 1 0 1 1 1 1 3T-1 ^ X 1 439 3T ^ X 343 3T-1 956 3 0 12 3T ^ X 171 3T-1 956 3T-1 439 1 1 13 J. 3T X 171 3T 343 3T-1 956 2 0 14 3T 171 3T-1 439 3T 343 1 0 15 1*/ 3T-1 ^ X X 439 3T 171 3T 343 1 1 16 3T-1 439 3T-1 956 3T 171 2 0 17 3T-1 956 3T 343 3T-1 439 3 0 18 3T 171 3T-1 956 3T-1 439 2 0 19 X ^ 3T ^ X 343 3T 171 3T-1 956 3 1 20 3T 171 3T-1 439 3T 343 1 0 21 3T-1 956 3T 171 3T 343 3 0 22 3T-1 439 3T-1 956 3T 171 2 0 23 3T-1 439 3T 171 3T-1 956 2 1 24 3T 171 3T-1 956 3T-1 439 1 1 25 3T 343 3T 171 3T-1 956 2 0

PAGE 119

Table D-2. Continued 1 nanglc Sample 1 xvanuuni Code Sample2 iVdilUUIll Code Sample3 IValliJlJlii Code 1 ull&llot d Selection Correct? J 1 171 1/1 7 JU 3T J 1 343 0 z / J 1 1 "JO 3T 171 1/1 0 7 JU 3T J 1171 1/1 1 7Q J 1 1 yj\j lT-1 J 1 1 439 X -J 0 J I -? 1 " 1 y^\} lT-1 -J 1 1 439 X 0 J I 1 71 1/1 lT-1 J 1 1 y^yj 1 1 0 D 1 171 1/1 IT J 1 343 7 z. 1 1 J J 3T 171 1/1 2 0 lT-1 y*j\) "^T-l '-tj y 3T J 1 343 J 1 1 J J t J" J I 171 1/1 lT-1 J 1 1 y jyj 0 J I J 1 I y^yj lT-1 J 1 " 1 41Q t J y 7 0 J 1 J 1 J 1 171 1/1 lT-1 J 1 1 1 1 0 JO IT 171 1/1 4^Q y IT 141 1 1 n \j jy J 1 ~ 1 yj\j 1 71 1/1 IT 141 1 1 1 1 AO y^yj 41Q IT D 1 171 1/1 9 z. 0 u 41 J 1 1 43Q "J y 3T J 1 141 lT-1 -J 1 1 7 J v 'I 0 42 -~ J 1 171 1/1 -J 1 1 41Q lT-1 J 1 1 0 43 "^T 1 71 1/1 J 1 141 lT-1 J 1 1 1 1 n 44 '?T 171 1/1 lT-1 J 1 1 41Q IT 141 'I 45 3T-1 439 3T 343 3T 171 1 1 46 3T-1 439 3T-1 956 3T 171 3 1 47 3T-1 956 3T 343 3T-1 439 3 0 48 3T 171 3T-1 439 3T-1 956 1 1 49 3T 171 3T 343 3T-1 439 2 0 50 3T 343 3T-1 439 3T 171 1 0

PAGE 120

107 Table D-3. The triangle test results from the comparisons of the mixtures 4T and 4T-1 1 1 loll^lC FXD # Sample 1 IVallUUlll Code Sample2 IVallUUlll Code Sample3 ixilllUUlll PQtl#*llCt*C 1 ollCliol d Slplprtinn Correct? 1 1 4T 1 4T 1 1 694 4T-1 til 111 DJ 1 0 2 4T 1 1 0\J 4T-1 t 1 ~ 1 4T t 1 fOA 1 1 n J 4T-1 til DJ 1 4T t 1 V/^t 4T 1 1 1K6 1 0\J 1 1 1 1 4 4T-1 llA 4T-1 t 1 ~ 1 331 J J 1 4T t 1 186 1 Ovl 9 0 5 4T-1 til 4T t 1 694 4T-1 t 1 ~ 1 331 J J 1 x 0 d 4T t 1 674 VJZ.t 4T-1 til 331 J J 1 4T-1 til llA J 0 7 4T t 1 1 1 ou 4T fOA 4T-1 til 111 DD 1 0 8 o 4T t 1 1 0\J 4T-1 t 1 1 J J 1 4T t 1 fOA u^t 1 1 0 Q 4T-1 "T 1 ~ 1 331 J J 1 4T t 1 186 1 Ovi 4T t 1 694 1 1 1 in 1 u 4T-1 4T-1 til J J 1 4T 1 1 1 0\J 1 1 1 1 1 X 4T-1 *+ 1 ~ 1 llA 4T t 1 1 0\J 4T-1 t 1 ~ 1 111 1 1 1 12 4T t 1 4T-1 t 1 ~ 1 llA 4T-1 t 1 ~ 1 331 J J 1 9 Z 0 4T *T 1 1 OVJ 4T t 1 694 4T-1 t 1 " 1 331 J J 1 1 1 n 14 4T t 1 UZ-t 4T-1 til 111 J J 1 4T t i 1 Rf> 9 1 1 15 4T-1 til 111 J J 1 AT 1 1 fOA 4T 1 1 1 X6 n 16 4T-1 t 1 " 1 J J 1 4T-1 til llA 4T 1 1 1 X6 z n w 17 4T-1 t 1 ~ 1 331 J J 1 4T 1 1 1 %(s 1 0\J 4T-1 1 1 ~ 1 llA 9 1 1 18 4T t 1 4T-1 1 1 1 llA 4T-1 1 1 1 J J 1 9 19 4T t X 694 4T 1 1 1 R6 1 ou 4T-1 1 1 1 llA -I 1 20 4T 624 4T-1 331 4T 186 2 1 21 4T-1 331 4T 624 4T 186 3 0 22 4T-1 331 4T-1 774 4T 624 2 0 23 4T-1 331 4T 624 4T-1 774 3 0 24 4T 186 4T-1 774 4T-1 331 3 0 25 4T 186 4T 624 4T-1 331 2 0

PAGE 121

108 Table D-3. Continued i 1 loll^JlC EXD # Sample 1 R ofiH Am IVallUUlll Code Sample2 IVOlllJUill Code Sample3 IvCUlUVXil Code Selection Correct? 7ft 4T 674 4T-1 774 4T 186 2 1 77 4T-1 J J) I 4T 674 4T t 1 1 86 7 0 4T-1 774 4T-1 4T t 1 0 4T-1 J J 1 4T 674 4T-1 til llA 0 DM 4T 1 Rft 1 oyj 4T-1 *T 1 1 4T-1 t 1 ~ 1 llA X 0 J 1 4T 1 86 1 o\j 4T 674 4T-1 t 1 " 1 331 J J 1 x 1 1 ^7 4T 674 4T-1 4T t 1 1 1 0\J 1 1 4T-1 "Ti l J J 1 4T t 1 1 86 1 0\J 4T t 1 fDA 1 1 1 1 4T-1 774 4T-1 t 1 ~ 1 331 J J 1 4T t 1 186 1 0\J 1 1 n 4T-1 774 4T •t 1 1 86 1 0\J 4T-1 t 1 ~ 1 331 J J 1 T. 0 4T 1 R6 4T-1 t 1 " 1 111 4T-1 til 774 0 37 4T 674 4T 1 86 4T-1 t 1 ~ 1 331 J J 1 1 1 0 38 4T 1X6 4T-1 llA 4T t 1 674 2 1 1 39 4T-1 J J 1 4T •t 1 1 86 4T t 1 674 1 1 1 "TV 4T-1 til 131 4T-1 •t 1 ~ 1 llA 4T t 1 1 86 1 _j 1 1 41 ~ J. 4T-1 ~ 1 1 331 4T " 1 674 4T-1 t 1 " 1 llA 0 yj 42 4T 186 4T-1 331 4T-1 t 1 ~ 1 llA 0 43 4T 674 4T t 1 1 86 1 OvJ 4T-1 1 1 ~ 1 DJ 1 9 44 4T ~ 1 1X6 4T-1 til J J 1 1 1 u^t 9 1 45 4T-1 774 4T 186 4T 624 3 0 46 4T-1 331 4T-1 774 4T 186 2 0 47 4T-1 331 4T 624 4T-1 774 1 0 48 4T 624 4T-1 774 4T-1 331 1 1 49 4T 624 4T 186 4T-1 774 3 1 50 4T 186 4T-1 331 4T 624 1 0

PAGE 122

109 Table D-4. The triangle test results from the comparisons of the mixtures 4T and 4T-2 Exp. # Sample! RanHnm X VdX lU V/ 11 X Code Sample2 Random Code Sample3 Random Code Panelist's Selection Correct? 1 4T 214 4T 372 4T-2 835 1 0 2 4T 214 ^ 1 ~ 4T-2 519 4T 372 2 1 3 4T-2 835 4T 372 4T 214 2 0 4 4T-2 835 4T-2 519 4T 214 3 1 5 4T-2 835 4T 372 4T-2 519 2 1 6 4T 372 4T-2 835 4T-2 519 3 0 7 4T 214 4T ~ 1 372 4T-2 835 1 0 g 4T 214 4T-2 519 4T 372 2 1 9 4T-2 835 4T 372 4T 214 3 0 10 4T-2 835 4T-2 519 4T 214 1 0 11 1. 1 4T-2 835 4T 214 4T-2 519 3 0 12 4T 214 4T-2 835 4T-2 519 2 0 13 4T 372 4T 214 4T-2 835 2 0 14 4T 372 4T-2 519 4T 214 1 0 15 4T-2 835 4T 372 4T 214 1 1 X 16 4T-2 519 4T-2 835 4T 214 1 X 0 17 4T-2 519 4T 372 4T-2 835 2 1 18 4T 372 4T-2 519 4T-2 I 1 ^ 835 1 1 1 1 19 4T 214 4T 4T-2 519 1 1 0 20 4T 214 4T-2 519 4T 372 3 0 21 4T-2 835 4T 372 4T 214 3 0 22 4T-2 835 4T-2 519 4T 372 3 1 23 4T-2 519 4T 372 4T-2 835 1 0 24 4T 372 4T-2 835 4T-2 519 2 0 25 4T 214 4T 372 4T-2 519 3 1

PAGE 123

110 Table D-4. Continued Trian&le Exp. # Sample 1 Random Code Sample2 Random Code Sample3 Random Code Panelist's Selection Correct? 26 4T 372 4T-2 835 4T 214 1 0 27 4T-2 519 4T 372 4T 214 3 0 28 4T-2 835 4T-2 519 4T 372 3 1 29 4T-2 519 4T 214 4T-2 835 2 1 30 4T 214 4T-2 519 4T-2 835 3 0 31 4T 372 4T 214 4T-2 835 1 0 32 4T 372 4T-2 519 4T 214 2 1 33 4T-2 519 4T 214 4T 372 3 0 34 4T-2 835 4T-2 519 4T 214 1 0 35 4T-2 519 4T 214 4T-2 835 1 0 36 4T 372 4T-2 835 4T-2 519 2 0 37 4T 372 4T 214 4T-2 835 1 0 38 4T 372 4T-2 519 4T 214 1 0 39 4T-2 519 4T 372 4T 214 2 0 40 4T-2 835 4T-2 519 4T 214 2 0 41 4T-2 519 4T 372 4T-2 835 1 0 42 4T 214 4T-2 519 4T-2 835 2 0 43 4T 372 4T 214 4T-2 519 2 0 44 4T 214 4T-2 519 4T 372 2 1 45 4T-2 519 4T 214 4T 372 1 1 46 4T-2 835 4T-2 519 4T 214 2 0 47 4T-2 519 4T 372 4T-2 835 2 1 48 4T 214 4T-2 519 4T-2 835 1 1 49 4T 214 4T 372 4T-2 835 1 0 50 4T 372 4T-2 519 4T 214 3 0

PAGE 124

Ill Table D-5. The triangle test results from the comparisons of the mixtures 4T and 4T-3 1 1 laii^i V Exp. # Sample! Random Code Sample2 Random Code Sample3 Random Code Panelist's Selection Correct? 1 1 4T 128 4T 263 4T-3 624 3 1 2 4T 128 4T-3 624 4T 263 2 1 4T-3 902 4T 263 4T 128 3 0 4 4T-3 624 4T-3 902 4T 263 3 1 5 4T-3 624 4T 128 4T-3 902 1 0 6 4T 128 4T-3 902 4T-3 624 2 0 7 4T 263 4T 128 4T-3 624 2 0 g 4T 263 4T-3 624 4T 128 1 0 9 4T-3 902 4T 128 4T 263 1 1 10 4T-3 624 4T-3 902 4T 263 2 0 11 4T-3 624 4T 128 4T-3 902 1 0 12 4T 128 4T-3 902 4T-3 624 1 1 13 4T 263 4T 128 4T-3 902 3 1 14 4T 263 4T-3 902 4T 128 1 0 15 4T-3 624 4T 263 4T 128 3 0 16 4T-3 624 4T-3 902 4T 128 2 0 17 4T-3 624 4T 128 4T-3 902 1 0 18 4T 263 4T-3 624 4T-3 902 1 1 19 4T 128 4T 263 4T-3 624 2 0 20 4T 263 4T-3 624 4T 128 1 0 21 4T-3 624 4T 128 4T 263 2 0 22 4T-3 902 4T-3 624 4T 128 1 0 23 4T-3 624 4T 263 4T-3 902 1 0 24 4T 128 4T-3 902 4T-3 624 3 0 25 4T 128 4T 263 4T-3 902 2 0

PAGE 125

Table D-5. Continued TrianfTlp Exp. # Sample 1 Code Sample2 Random Code Sample3 Random Code Panelist's Selection Correct? 26 4T 128 4T-3 624 4T 263 3 0 27 624 4T 263 4T 128 2 0 28 4T-3 902 4T-3 624 4T 263 2 0 29 4T-3 624 4T 128 4T-3 902 1 0 30 4T 128 4T-3 902 4T-3 624 3 0 31 4T 128 4T 263 4T-3 902 3 1 32 4T 128 4T-3 624 4T 263 3 0 33 4T-3 624 4T 263 4T 128 3 0 34 4T-3 902 4T-3 624 4T 128 2 0 35 4T-3 902 4T 128 4T-3 624 1 0 36 4T 263 4T-3 902 4T-3 624 1 1 37 4T 263 4T 128 4T-3 624 1 0 38 4T 128 4T-3 902 4T 263 1 0 39 4T-3 624 4T 263 4T 128 3 0 40 4T-3 624 4T-3 902 4T 263 1 X 0 41 4T-3 902 4T 128 4T-3 624 3 0 42 4T 263 4T-3 902 4T-3 624 3 0 43 4T 128 4T 263 4T-3 624 2 0 44 4T 263 4T-3 ~ X ^ 902 4T ~ X 128 X 2 1 45 4T-3 902 4T 263 4T 128 1 1 46 4T-3 624 4T-3 902 4T 263 2 0 47 4T-3 902 4T 263 4T-3 624 3 0 48 4T 128 4T-3 624 4T-3 902 3 0 49 4T 128 4T 263 4T-3 624 3 1 50 4T 128 4T-3 624 4T 263 1 0

PAGE 126

113 Table D-6. The triangle test results from the comparisons of the mixtures 5T and 5T-1 Trianele Exp. # Sample 1 Random X ^V4X J. V/ X XX Code Sample2 Random Code Sample3 Random Code Panelist's Selection Correct? 1 5T 428 5T 737 5T-1 950 2 0 2 5T 428 5T-1 950 5T 737 1 0 3 5T-1 355 5T 428 5T 737 2 0 4 5T-1 950 5T-1 355 5T 737 2 0 5 5T-1 950 5T 428 5T-1 355 2 1 6 5T 428 5T-1 355 5T-1 950 2 0 7 5T 737 5T 428 5T-1 950 1 0 8 5T 737 5T-1 355 5T 428 2 1 9 5T-1 355 5T 428 5T 737 3 0 10 5T-1 355 5T-1 950 5T 737 2 0 11 5T-1 355 5T 737 5T-1 950 2 1 12 5T 737 5T-1 355 5T-1 950 2 0 13 5T 737 5T 428 5T-1 355 3 1 14 5T 428 5T-1 355 5T 737 1 0 15 5T-1 355 5T 737 5T 428 3 0 16 5T-1 950 5T-1 355 5T 737 3 1 17 5T-1 950 5T 428 5T-1 355 2 1 18 5T 737 5T-1 355 5T-1 950 2 0 19 5T 428 5T 737 5T-1 950 2 0 20 5T 428 5T-1 950 5T 737 3 0 21 5T-1 355 5T 737 5T 428 1 1 22 5T-1 950 5T-1 355 5T 428 1 0 23 5T-1 355 5T 737 5T-1 950 1 0 24 5T 428 5T-1 950 5T-1 355 2 0 25 5T 737 5T 428 5T-1 950 1 0

PAGE 127

114 Table D-6. Continued 1 ••1 on rti o I riangie Fvn U Sample 1 iVallUUIIl Code Sample2 JValiUUlll Code Sample3 Code Panplist's Selection Correct? zo 7'?7 I J 1 ST-l J 1 1 QSO ^ J\J ST 428 1 0 97 z/ J 1 1 ISS J J J ST 498 ST 737 1 1 1 "78 1 J 1 1 "ISS J J J ST-1 QSO ST 737 1 J 1 1 oso ST 498 ST-1 355 J *J >J 9 Am 1 J 1 111 ST.1 J 1 1 3SS J J J 9 0 J 1 J I 498 ST 1X1 ST-1 QSO 9 Z. 0 J I 1 J 1 ST-1 ST 498 -J n '^'^ J 1 1 "^SS J J J ST ST 717 0 ST 1 "^SS J J J ST-1 J 1 1 y jyj ST 9 0 J J ST 1 J 1 -1 "^SS J J J ST 1 D 1 ST-1 QSO 9 1 1 ST ST-1 J 1 1 J J J ST-1 QSO 1 1 1 "^7 ST 498 ST 1X1 ST-1 ISS 1 1 n JO ST 498 ST-1 ST 717 1 1 0 ST 1 osn ST 1X1 ST 'I 0 ST.1 J 1 1 Qsn ST-1 -J 1 1 J> J J ST 717 0 *Tl ST-1 "^SS ST 1 D 1 ST-1 QSO y j\j 1 1 0 ST 498 ST-1 J J J ST-1 QSO 'X J 0 ST 498 ST 717 ST-1 QSO y jyj 1 1 0 44 ST 717 ST-1 J \ ~\. OSO yjyj ST 49 R 1 1 0 45 5T-1 355 5T 737 5T 428 2 0 46 5T-1 355 5T-1 950 5T 428 1 0 47 5T-1 950 5T 428 5T-1 355 2 1 48 5T 428 5T-1 950 5T-1 355 1 1 49 5T 428 5T 737 5T-1 355 1 0 50 5T IZl 5T-1 355 5T 428 1 0

PAGE 128

LIST OF REFERENCES Acree T and Am H (2001), FlavorNet, http://www.nysaes.comell.edu/flavomet/chem.html, 03/09/2001 . Adams RP (1995), Identification of Essential Oil Components by Gas Chromatography/Mass Spectroscopy, Allured Publishing Corporation, Carol Stream, IL, USA. Alpha-M.O.S. (1997), "Study of pure hops and blends with an electronic nose," Application note 24 of Alpha-M.O.S., Released in October 1997. American Spice Trade Association (1999), "Spice definition and glossary," http://www.astaspice.org/spice/frame_spice.htm, 01/1 1/1999. American Society for Testing and Materials (ASTM) (1991a), "Standard practice for determination of odor and taste threshold by a forced-choice method of limits," E679-91, \r Annual Book of ASTM Standards, American Society for Testing and Materials, Philadelphia. ASTM (1991b), "Standard practice for defining and calculating individual and group sensory thresholds fi-om forced-choice data sets of intermediate size," E-679-91, In Annual Book of ASTM Standards, American Society for Testing and Materials, Philadelphia. ASTM (1996), "Standard practice for Gas Chromatography Terms and Relationships," E?)55-96,\n Annual Book of ASTM Standards, American Society for Testing and Materials, Philadelphia. ASTM (1997a), "Standard practice for serving protocol for sensory evaluation of foods and beverages," El 87 1-97, In Annual Book of ASTM Standards, American Society for Testing and Materials, Philadelphia. ASTM (1997b), "Standard test method for sensory analysis triangle test," E1885-97, In Annual Book of ASTM Standards, American Society for Testing and Materials, Philadelphia. ASTM (2000), "Standard practice for installing fused silica open tubular capillary columns in gas chromatographs," El 5 10-95 (Reapproved), In Annual Book of ASTM Standards, American Society for Testing and Materials, Philadelphia. 115

PAGE 129

116 ASTM (2001), "Standard terminology relating to sensory evaluation of materials and products," E253-01, In Annual Book of ASTM Standards, American Society for Testing and Materials, Philadelphia. Annor-Frempong IE, Nute GR, Wood JD, Whittington FW and West A (1998), "The measurement of the responses to different odour intensities of 'boar taint' using a sensory panel and an electronic nose," Meat Science, 50(2), 139-151. Antinone MJ, Lawless HT, Ledford RA and Johnston M (1994), "Diacetyl as a flavor component in full fat cottage cheese," Journal of Food Science, 59, 38-42. Antonelli A, Fabbri C and Boseli E (1998), "Modification of dried basil {Ocinum basilicum) leaf oil by gamma and microwave irradiation," Food Chemistry, 63(4), 485-489. Aug-Yeung C and MacLeod A (1981), "A comparison of the efficiency of the Likens and Nickerson extractor for aqueous, liquid/aqueous, and lipid samples," J. Agric. Food Chem. 29, 502-506. Bachinger T, Liden H, Martensson P and Mandenius CF (1998), "On-line estimation of state variables in baker's yeast fermentation using an electronic nose," Seminars in Food Analysis, 3(1), 85-91. Baird JC (1997), Sensation and Judgment: Complementarity Theory of Psychophysics, Lawrence Erlbaum Associates, Mahwah, New Jersey. Benady M, Simon JE and Miles GE (1995), "Fruit ripeness determination by electronic sensing of aromatic volatiles," Trans. ASAE, 38, 251-257. Bihop C (1995), Neural Networks for Pattern Recognition, Oxford Press, Oxford. Bleibaum RN, Stone H, Tan T, Labreche S, Saint Martin E and Isz S (2002), "Comparison of sensory and consumer results with electronic nose and tongue sensors for apple juices," Food Quality and Preference, 13(6), 409-422. Blixt Y and Borch E (1999), "Using an electronic nose for determining the spoilage of vacuum-packaged beef," International Journal of Food Microbiology, 46(2), 123134. Brezmes J, Llobert E, Vilanova X, Saiz G and Correig X (2000), 'Fruit ripeness monitoring using an electronic nose," Sensors and Actuators, B: Chemical, 69, 223-229. Brown DGW, Clapperton, JF, Meilgaard MC and Moll M (1978), "Flavor thresholds of added substances," American Society of Brewing Chemists Journal, 36, 73-80.

PAGE 130

117 Burford MD (1998), "Supercritical fluid extraction coupled on-line with gas chromatography," In Ramsey ED, Analytical Supercritical Fluid Extraction Techniques, Kluwer Academic, Dordrecht, Netherlands, 43-68. Burl MC, Doleman BJ, Schaffer A, and Lewis NS (2001), "Assessing the ability to predict human percepts of odor quality from the detector responses of a conducting polymer composite-based electronic nose," Sensors and Actuators B: Chemical, 72, 149-159. Cai J, Liu B and Su Q (2001), "Comparison of simultaneous distillation extraction and solid-phase microextraction for the determination of volatile flavor components," Journal of Chromatography A, 930, 1-7. Chaintreau A (2001), "Simultaneous distillation-extraction: from birth to maturity review," Flavour and Fragrance Journal, 16, 136-148. Chemical Abstracts Service (2003), "The CAS registry," http://www.cas.org/EO/regsys.html, 05/04/2003. Cheng CK, Chen CC, Shu WY, Shih LL and Feng HH (1997), "Component analysis of mixed spices," In Risch SJ and Ho H, Spices Flavor Chemistry and Antioxidant Properties, American Chemical Society, Washington, DC, 160-174. Chien M (1985), "Analysis of complex mixtures by gas chromatography/mass spectrometry using a pattern recognition method," Anal. Chem, 57:348-352. Cornell JA (1 990), Experiments with Mixtures Designs, Models, and the Analysis of Mixture Data. John Wiley & Sons, New York. Deventer DV and Mallikarjunan P (2002), "Optimizing an electronic nose for analysis of volatiles from printing inks on assorted plastic films," Innovative Food Science and Emerging Technologies, 3(1), 91-99. Diaz-Maroto MC, Perez-Coello MS and Cabezudo MD (2002), "Supercritical carbon dioxide exfraction of volatiles from spices, comparison with simultaneous distillation-extraction," Journal of Chromatography A, 947, 23-29. Distante C, Siciliano P and Vasanelli L (2000), "Odor discrimination using adaptive resonance theory," Sensors and Actuators B: Chemical, 69, 248-252. Dittmaim B, Zimmermann B, Engelen C, Jany G and Nitz S (2000), "Use of the MSsensor to discriminate between different dosages of garlic flavoring in tomato sauce," J: Agric. Food Chem, 48, 2887-2892. Doleman BJ, Lewis NS (2001), "Comparison of odor detection thresholds and odor discriminablities of a conducting polymer composite electronic nose versus mammalian olfaction," Sensors and Actuators B: Chemical, 72, 4150.

PAGE 131

118 Eklov T, Johansson G, Winquist F and Lundstorm I (1998), "Monitoring sausage fermentation using an electronic nose," Journal of the Science of Food and Agriculture, 76, 525-532. Ennis DM (1990), "Relative power of difference testing methods in sensory evaluation," Food technology, 44(4), 114-117. Evans P, Persaud K, McNeish A, Sneath R, Hobson N, and Magan N (2000), "Evaluation of a radial basis fiinction neural network for the determination of wheat quality from electronic nose data," Sensors and Actuators B: Chemical, 69, 348-358. Falcitelli M, Benassi A, Francesco FD, Domenici C, Marano L and Pioggia G (2002), "Fluid dynamic simulation of a measurement chamber for electronic noses," Sensors and Actuators B: Chemical, 85, 166-174. Frijters JER (1980a), "Three-stimulus procedures in olfactory psychophysics: an experimental comparison of Thurstone-Ura and ThreeAlternative Forced-Choice models of signal detection theory," Perception and Psychophysics, 27(2), 176-178. Frijters JER, Kooistra A, and Vereijken PFG (1980b), "Tables oid' for the triangular method and the 3-AFC signal detection procedure," Perception and Psychophysics, 27(2), 176-178. Frijters JER (1987), "Psychophysical models for mixtures of tastants and mixtures of odorants," Annual of the New York Academy of Sciences, 5 1 0, 67-68. Garcia-Gonzalez DL and Aparicio R (2002), "Detection of defective virgin olive oils by metal-oxide sensors," European Food Research and Technology, 215(2), 1 18-123. Gardner JW, Bartlett PN (1994), "A brief history of electronic noses," Sensors and Actuators B: Chemical, 18-19, 21 1-220. Gardner JW and Bartlett PN (1999), Electronic Noses: Principles and Applications, Oxford University Press, Oxford. Giese J (1994), "Spices and seasoning blends: a taste for all seasons," Food Technology, 4, 88-98. Gregson RAM (1984), "Similarities between odor mixtures with known components," Perception and Psychophysics, Lawrence Eribaimi Associates, Hillsdale, New Jersey. Gregson RAM (1992), n-Dimensional Nonlinear Psychophysics: Theory and Case Studies, 35(1), 33-40. Gretsh C, Toury A, Estebaranz R and Liardon R (1998), "Sensitivity of metal oxide sensors towards coffee aroma," Seminars in Food Analysis, 3(1), 37-42.

PAGE 132

119 Harmon AD (1997), "Solid-Phase microextraction for the analysis of flavors," In Marsili R, Techniques for Analyzing Food Aroma, Marcel Dekker, New York, 81-1 12. Haykin S (1999), Neural Networks A Comprehensive Foundation, 2"** ed., Prentice-Hall, Upper Saddle River, New Jersey. Heinio RL and Ahvenainen R (2002), "Monitoring of taints related to printed solid boards with an electronic nose," Food Additives and Contaminations, 19(supplement), 209-220. Hirasa K and Takemasa M (1998), Spices Science and Technology. Marcel Dekker, New York. Hirschfelder, M., Forster, A., Ktihne, S., Langbehn, j., Junghanns, W., Pank, F., and Hanrieder, D. (2000), "Using multivariate statistics to predict sensory quality of marjoram from instrumental data," Sensors and Actuators, B: Chemical, 69, 404407. Holmberg M, Winquist F, Lundstrom 1, Gardner JW and Hines EL (1995), "Identification of paper quality using a hybrid electronic nose," Sensors and Actuators, B: Chemical, 26/27, 246-249. ISO (1999), "Sensory analysis methodology general guidance for measuring odour, flavour, and taste detection thresholds by a Three-Alternative Forced-Choice (3AFC) procedure," International Organization for Standardization, Draft International Standard ISO/ DIS 1 3 30 1. 2. ISO, Case Postale 56, 1 rue Varembe, CHI 211 Geneve, Switzerland. Jennings W and Filsoof M (1977), "Comparison of sample preparation techniques for gas chromatographic analysis," J. Agric. FoodChem. 25: 440-452. Jennings W, Mittlefehldt E and Sti-emple P (1997), Analytical Gas Chromatography, 2"** edition, Academic Press, New York. Kermani BG, Schiffman SS, Nagle HT (1999), "Using neural networks and genetic algorithms to enhance performance in an electronic nose," IEEE Transactions on Biomedical Engineering, 46(4), 429-439. Khattree R, Naik DN (2000), Multivariate Data Reduction and Discrimination with SAS® Software, SAS Institute Inc, Gary, North Carolina. Kim SM, Wu CM, Kubota K and Kobayashi A (1995), "Effect of soybean oil on garlic volatile compound isolated by distillation," Journal of Agricultural and Food Chemistry, 43, 499-452. Kolb B, Ettie LS (1997), Static Headspace-Gas Chromatography Theory and Practice, John Wiley & Sons, New York.

PAGE 133

120 Korel F, Luzuriaga DA and Balaban MO (2001), "Objective quality assessment of raw tilapia (Oreochromis niloticus) fillets using electronic nose and machine vision," Journal of Food Science, 66(7), 1018-1024. Korel F and Balaban MO (2002), "Microbial and sensory assessment of milk with an electronic nose," Journal of Food Science, 67(2), 758-764. Lachowicz KJ, Jones GP, Briggs DR, Bienvenu FE, Palmer MV, Mishra V and Hunter MM (1997), "Characteristics of plants and plant extracts from five varieties of basil (Ocimum basilicum L.) grown in Australia," Journal of Agricultural and Food Chemistry, 45, 2660-2665. Lawless HT and Heymarm H (1998), Sensory Evaluation of Food: Principle and Practices, Chapman & Hall, New York. Lawrence B and Shu CK (1993), "Essential oils as components of mixtures: analysis and differentiation," In Ho C and Manley CH, Flavor Measurement, Marcel Dekker, New York, 267-327. Likens ST and Nickerson GB (1964), "Detection of certain hop oil constituents in brewing products," In American Society of Brewing Chemists Proceedings, 5-13. Llobet E, Hines EL, Gardner JW and Franco S (1999), "Non-destructive banana ripeness determination using a neural network-based electronic nose," Meas. Sci. Technol. , 10, 538-548. Maignial L, Pibarot P, Bonetti G, Chaintreau A and Marion JP (1992), "Simultaneous distillation-extractiom under static vacuum: isolation of volatile compounds at room temperature," Journal of Chromatography, 606, 87-94. Marco S, Ortega A, Pardo A and Samitier J (1998), "Gas identification with tin oxide sensor array and self organizing maps: adaptive correction of sensor drifts," IEEE Trans. Instrum. Meas., 47, 316-321. Marotti M, Piccaglia R and Glovanelli E (1996), "Differences in essential oil composition of basil {Ocimum basilicum L.) Italian cultivars related to morphological characteristics," Journal of Agricultural and Food Chemistry, 44, 3926-3929. Marsili R (2001), "Solid-Phase microextraction," http://www.foodproductdesign.eom/archive/l 996/1 296QA.html, 04/09/200 1 . Maul F, Sargent SA, Balaban MO, Baldvdn EA, Huber DJ and Sims CA (1999), "Aroma volatile profiles fi-om ripe tomatoes are influenced by physiological maturity at harvest: an application for electronic nose technology," Journal of the American Society for Horticultural Science, 123(6), 1094-1101. Meilgaard M, Civille GV and Carr BT (1999), Sensory Evaluation Techniques, 3"* Edition, CRC Press, Boca Raton, Florida.

PAGE 134

121 Messer DC, Davies GR, Rosselli AC, Prange CG and Algaier JW (1998), "Supercritical fluid extraction instrumentation," In Ramsey ED, Analytical Supercritical Fluid Extraction Techniques, Kluwer Academic, Dordrecht, Netherlands, 43-68. Miller KG, Poole CF and Pawlowski TMP (1996), "Classification of the botanical origin of cinnamon by solid-phase microextraction and gas chromatography," Chromatographia, 42(1 1/12), 639-646. Misharina T, Golovnya R, and Beletsky I (1994), "Comparison of the efficiency of isolation of volatiles from foodstuffs by co-distillation and Likens-Nickerson methods," In Maarse H and van der Heij D, Trends in Flavor Research, Elsevier, New York, 117-129. Moulton DG, Turk A and Johnston JW JR (1975), Methods in Olfactory Research, Academic Press, New York. Muhl M, Demisch HU, Becker F and Kohl CD (2001), "Electronic nose for deterioration of fiying fat comparative studies for a new quick test," European Journal of Lipid Science and Technology, 102(8/9), 581-585. Pardo M, Niederjaufiier G, Bebussi G, Comini E, Faglia G, Sberveglieri G, Holmberg M, and Lundstorm I (2000), "Data preprocessing enhances the classification of different brands of Espresso coffee with an electronic nose," Sensors and Actuators B: Chemical, 69, 397-403. Park EY, Noh BS and Ko S (2002), "Prediction of shelf life for soybean curd by electronic nose and artificial neural network system," Food Science and Biotechnology, 11(3), 245-251. Parliament TH (1997), "Solvent extraction and distillation techniques," In Marsili R, Techniques for Analyzing Food Aroma, New York, Marcel Dekker, 1-26. Pawliszyn J (1999), "Quantitative aspects of SPME," In Pawliszyn J, Applications of Solid Phase Microextraction, Cambridge, Royal Society of Chemistry, 3-21. Pawliszyn J (2000), "Theory of solid-phase microextraction," J. Chromatographic ScL, 38:270-277. Pearce TC, Gardner JW, Friel S, Bartlett PN and Blair N (1993), "Electronic nose for monitoring the flavor of beers," The Analyst, 118, 371-377. Persaud K and Dodd GH, (1982), "Analysis of discrimination mechanisms of the mammalian olfactory system using a model nose," Nature, 299, 352-355. Ping W, and Jun X (1996), "A novel recognition method for electronic nose using artificial neural network and fuzzy recognition," Sensors and Actuators B: Chemical, 37, 169-174.

PAGE 135

122 Pinheiro C, Rodrigues CM, Schafer T and Crespo JG (2002), "Monitoring the aroma production during wine-must fermentation with an electronic nose," Biotechnology and Bioengineering, 77(6), 632-640. Poling J, Lucas Q and Weber K (1997), "Quality control of packaging with the electronic nose Fox-4000," Journal of Automatic Chemistry 19(4), 115-1 16. Principe JC, Euliano NR, Lefebvre WC. (1999), Neural and Adaptive Systems: Fundamentals through Simulations, John Wiley & Sons, New York. Rakow NA, Suslick KS (2000), "A colorimetric sensor array for odour visualization," Nature, 406,710-714. Ropkins K and Taylor AJ (1996), "The isolation of flavour compounds from foods by enhanced solvent extraction methods," In Taylor AJ and Mottram DS, Flavour Science: Recent Developments, Royal Society of Chemistry, Cambridge, 297-304. Rouseff R and Cadwallader K (2001), "Headspace techniques in food, fragrances and flavors: an overview," In Rouseff R and Cadwallader K, Headspace Analysis of Ffoods and Flavors: Theory and Practice, Kluwer Academic, New York. Schultz T, Flath R, Mon R, Eggling S, and Teranishi R (1977), "Isolation of volatile components from a model system," J. Agric. Food Chem. 25, 446-450. Schweizer-Berberich M, Vahinger S and Gopel W (1994), "Characterisation of food freshness with sensor arrays," Sensors and Actuators, B: Chemical, 18/19, 282-290. Scott RPW and Perry JA (1998), Introduction to Analytical Gas Chromatography, 2"^ ed.. Marcel Dekker, New York. Shen N, Moizuddin S, Wilson L, Duvick S, White P and Pollak L (2001), "Relationship of electronic nose analyses and sensory evaluation of vegetable oils during storage," Journal of the American Oil Chemists ' Society, 78(9), 937-940. Sides A, Robards K, and Helliwell S (2000), "Developments in extraction techniques and their application to analysis of volatiles in foods," Trends in analytical chemistry, 19(5), 322-329. Simon JE, Hetzroni A, Bordelon B, Miles GE and Charles DJ (1996), "Elecfronic sensing of aromatic volatiles for quality sorting of blueberries," Journal of Food Science, 61,967-970. Snopok BA and Kruglenko IV (2002), "Multisensor systems for chemical analysis: stateof-the-art in Electronic Nose technology and new trends in machine olfaction," Thin Solid Films, 418,21-41.

PAGE 136

123 Stella R, Barisci JN, Serra G, Wallace GG and De Rossi D (2000), "Characterization of olive oil by an electronic nose based on conducting polymer sensors," Sensors and Actuators, B: Chemical, 63, 1-9. Uhl SR (2000), Spices, Seasonings and Flavorings. Technomic Publishing, Lancaster, Pennsylvania. Vemin G, Vemin C, Metzger J, Pujol L and Parkanyi C (1994), "GC/MS analysis of cinnamon and cassia essential oils: a comparative study," In Charalambous G, Spices, Herbs and Edible Fungi, Elsevier Science, Berlin, 41 1-424. Waibel A, Hanazawa T, Hinton G, Shikano K, and Lang KJ (1989), "Phoneme recognition using time-delay neural networks," IEEE Transactions on Acoustics, Speech and Signal Processing, ASSP-37: 328-339. Wampler TP (1997), "Analysis of food volatiles using headspace-gas chromatographic techniques," In Marsili R, Techniques for Analyzing Food Aroma, Marcel Dekker, New York, 27-58. Wijesekera ROB (1978), "The chemistry and technology of cinnamon," CRC Crit. Rev. FoodSci. Nutr., 10, 1-30. Yu T, Wu C and Liou Y (1989), "Volatile compounds from garlic," Journal of Agricultural and Food Chemistry, 37, 725-730. Yu T, Wu C and Ho C (1993), "Volatile compounds of deep-oil fried, microwave-heated, and oven-baked garlic slices," Journal of Agricultural and Food Chemistry, 41, 800-805. Yu T, Lin L and Ho C (1994), "Volatile compounds of blanched, and baked blanched garlic slices,'" Journal of Agricultural and Food Chemistry, 42, 1342-1347. Zhang H, Balaban MO, Principe JC (In press), "Improving pattern recognition of electronic nose data with time-delay neural networks," Sensors and Actuators: B Chemical.

PAGE 137

BIOGRAPHICAL SKETCH Haoxian Zhang was bom in 1976, in Jiangxi province of the People's Republic of China. She is currently a Ph.D. candidate at the University of Florida, majoring in agricultural and biological engineering. She received a B.S in chemical engineering from the Beijing Institute of Light Industry in 1996 and a M.S. in biochemical engineering from the Graduate School of the Chinese Academy of Sciences in 1999. Her current interests include the instrumental and sensory evaluation of food quality, applying multivariate statistics/neural network for chemical information analysis and bio/food processing engineering. 124

PAGE 138

I certify that I have read this study and that in my opinion it conforms to acceptabl standards of scholarly presentation and is fully adequate, in scope and quality, as a dissertation for the degree of Doctor of Philosophy. lurat (X Balaban, Chairman Professor of Agricultural and Biological Engineering I certify that I have read this study and that in my opinion it conforms to acceptabl standards of scholarly presentation and is fully adequate, in scope and quality, as a dissertation for the degree of Doctor of Philosoph) Lenneth M. Portier Associate Professor of Statistics I certify that I have read this study and that in my opinion it conforms to acceptabl standards of scholarly presentation and is fully adequate^ji^.scQpe and quality, as a dissertation for the degree of Doctor of Philosophj ssor of Electrical and meering I certify that I have read this study and that in my opinion it conforms to acceptabl standards of scholarly presentation and is fully adequate, in scope and quality, as a dissertation for the degree of Doctor of Philosophy. Charles A. Sims Professor of Food Science and Human Nutrition I certify that I have read this study and that in my opinion it conforms to acceptabl standards of scholarly presentation and is fully adequate, in scope and quality, as a dissertation for the degree of Doctor of Philosophy. Arthur A. Tierra Professor of Agricul Engineering and Biological

PAGE 139

This dissertation was submitted to the Graduate Faculty of the College of Engineering and to the Graduate School and was accepted as partial fulfilhnent of the requirements for the degree of Doctor of Philosophy. August 2003 CAX^.^r^Ay U^^-^c^^-'-Jr-^ Pramod P. Khargonekar Dean, College of Engineering Winfred M. Phillips Dean, Graduate School


xml version 1.0 encoding UTF-8
REPORT xmlns http:www.fcla.edudlsmddaitss xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.fcla.edudlsmddaitssdaitssReport.xsd
INGEST IEID EXQYG95TH_BYG5HO INGEST_TIME 2015-04-14T18:58:07Z PACKAGE AA00030038_00001
AGREEMENT_INFO ACCOUNT UF PROJECT UFDC
FILES